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 | |
14 | * through a control-dependency in io_get_cqring (smp_store_release to | |
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> | |
6c271ce2 | 60 | #include <linux/kthread.h> |
2b188cc1 | 61 | #include <linux/blkdev.h> |
edafccee | 62 | #include <linux/bvec.h> |
2b188cc1 JA |
63 | #include <linux/net.h> |
64 | #include <net/sock.h> | |
65 | #include <net/af_unix.h> | |
6b06314c | 66 | #include <net/scm.h> |
2b188cc1 JA |
67 | #include <linux/anon_inodes.h> |
68 | #include <linux/sched/mm.h> | |
69 | #include <linux/uaccess.h> | |
70 | #include <linux/nospec.h> | |
edafccee JA |
71 | #include <linux/sizes.h> |
72 | #include <linux/hugetlb.h> | |
aa4c3967 | 73 | #include <linux/highmem.h> |
15b71abe JA |
74 | #include <linux/namei.h> |
75 | #include <linux/fsnotify.h> | |
4840e418 | 76 | #include <linux/fadvise.h> |
3e4827b0 | 77 | #include <linux/eventpoll.h> |
ff002b30 | 78 | #include <linux/fs_struct.h> |
7d67af2c | 79 | #include <linux/splice.h> |
b41e9852 | 80 | #include <linux/task_work.h> |
bcf5a063 | 81 | #include <linux/pagemap.h> |
0f212204 | 82 | #include <linux/io_uring.h> |
2b188cc1 | 83 | |
c826bd7a DD |
84 | #define CREATE_TRACE_POINTS |
85 | #include <trace/events/io_uring.h> | |
86 | ||
2b188cc1 JA |
87 | #include <uapi/linux/io_uring.h> |
88 | ||
89 | #include "internal.h" | |
561fb04a | 90 | #include "io-wq.h" |
2b188cc1 | 91 | |
5277deaa | 92 | #define IORING_MAX_ENTRIES 32768 |
33a107f0 | 93 | #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) |
65e19f54 JA |
94 | |
95 | /* | |
96 | * Shift of 9 is 512 entries, or exactly one page on 64-bit archs | |
97 | */ | |
98 | #define IORING_FILE_TABLE_SHIFT 9 | |
99 | #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT) | |
100 | #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1) | |
101 | #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE) | |
21b55dbc SG |
102 | #define IORING_MAX_RESTRICTIONS (IORING_RESTRICTION_LAST + \ |
103 | IORING_REGISTER_LAST + IORING_OP_LAST) | |
2b188cc1 JA |
104 | |
105 | struct io_uring { | |
106 | u32 head ____cacheline_aligned_in_smp; | |
107 | u32 tail ____cacheline_aligned_in_smp; | |
108 | }; | |
109 | ||
1e84b97b | 110 | /* |
75b28aff HV |
111 | * This data is shared with the application through the mmap at offsets |
112 | * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING. | |
1e84b97b SB |
113 | * |
114 | * The offsets to the member fields are published through struct | |
115 | * io_sqring_offsets when calling io_uring_setup. | |
116 | */ | |
75b28aff | 117 | struct io_rings { |
1e84b97b SB |
118 | /* |
119 | * Head and tail offsets into the ring; the offsets need to be | |
120 | * masked to get valid indices. | |
121 | * | |
75b28aff HV |
122 | * The kernel controls head of the sq ring and the tail of the cq ring, |
123 | * and the application controls tail of the sq ring and the head of the | |
124 | * cq ring. | |
1e84b97b | 125 | */ |
75b28aff | 126 | struct io_uring sq, cq; |
1e84b97b | 127 | /* |
75b28aff | 128 | * Bitmasks to apply to head and tail offsets (constant, equals |
1e84b97b SB |
129 | * ring_entries - 1) |
130 | */ | |
75b28aff HV |
131 | u32 sq_ring_mask, cq_ring_mask; |
132 | /* Ring sizes (constant, power of 2) */ | |
133 | u32 sq_ring_entries, cq_ring_entries; | |
1e84b97b SB |
134 | /* |
135 | * Number of invalid entries dropped by the kernel due to | |
136 | * invalid index stored in array | |
137 | * | |
138 | * Written by the kernel, shouldn't be modified by the | |
139 | * application (i.e. get number of "new events" by comparing to | |
140 | * cached value). | |
141 | * | |
142 | * After a new SQ head value was read by the application this | |
143 | * counter includes all submissions that were dropped reaching | |
144 | * the new SQ head (and possibly more). | |
145 | */ | |
75b28aff | 146 | u32 sq_dropped; |
1e84b97b | 147 | /* |
0d9b5b3a | 148 | * Runtime SQ flags |
1e84b97b SB |
149 | * |
150 | * Written by the kernel, shouldn't be modified by the | |
151 | * application. | |
152 | * | |
153 | * The application needs a full memory barrier before checking | |
154 | * for IORING_SQ_NEED_WAKEUP after updating the sq tail. | |
155 | */ | |
75b28aff | 156 | u32 sq_flags; |
0d9b5b3a SG |
157 | /* |
158 | * Runtime CQ flags | |
159 | * | |
160 | * Written by the application, shouldn't be modified by the | |
161 | * kernel. | |
162 | */ | |
163 | u32 cq_flags; | |
1e84b97b SB |
164 | /* |
165 | * Number of completion events lost because the queue was full; | |
166 | * this should be avoided by the application by making sure | |
0b4295b5 | 167 | * there are not more requests pending than there is space in |
1e84b97b SB |
168 | * the completion queue. |
169 | * | |
170 | * Written by the kernel, shouldn't be modified by the | |
171 | * application (i.e. get number of "new events" by comparing to | |
172 | * cached value). | |
173 | * | |
174 | * As completion events come in out of order this counter is not | |
175 | * ordered with any other data. | |
176 | */ | |
75b28aff | 177 | u32 cq_overflow; |
1e84b97b SB |
178 | /* |
179 | * Ring buffer of completion events. | |
180 | * | |
181 | * The kernel writes completion events fresh every time they are | |
182 | * produced, so the application is allowed to modify pending | |
183 | * entries. | |
184 | */ | |
75b28aff | 185 | struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp; |
2b188cc1 JA |
186 | }; |
187 | ||
edafccee JA |
188 | struct io_mapped_ubuf { |
189 | u64 ubuf; | |
190 | size_t len; | |
191 | struct bio_vec *bvec; | |
192 | unsigned int nr_bvecs; | |
193 | }; | |
194 | ||
65e19f54 JA |
195 | struct fixed_file_table { |
196 | struct file **files; | |
31b51510 JA |
197 | }; |
198 | ||
05589553 XW |
199 | struct fixed_file_ref_node { |
200 | struct percpu_ref refs; | |
201 | struct list_head node; | |
202 | struct list_head file_list; | |
203 | struct fixed_file_data *file_data; | |
4a38aed2 | 204 | struct llist_node llist; |
05589553 XW |
205 | }; |
206 | ||
05f3fb3c JA |
207 | struct fixed_file_data { |
208 | struct fixed_file_table *table; | |
209 | struct io_ring_ctx *ctx; | |
210 | ||
05589553 | 211 | struct percpu_ref *cur_refs; |
05f3fb3c | 212 | struct percpu_ref refs; |
05f3fb3c | 213 | struct completion done; |
05589553 XW |
214 | struct list_head ref_list; |
215 | spinlock_t lock; | |
05f3fb3c JA |
216 | }; |
217 | ||
5a2e745d JA |
218 | struct io_buffer { |
219 | struct list_head list; | |
220 | __u64 addr; | |
221 | __s32 len; | |
222 | __u16 bid; | |
223 | }; | |
224 | ||
21b55dbc SG |
225 | struct io_restriction { |
226 | DECLARE_BITMAP(register_op, IORING_REGISTER_LAST); | |
227 | DECLARE_BITMAP(sqe_op, IORING_OP_LAST); | |
228 | u8 sqe_flags_allowed; | |
229 | u8 sqe_flags_required; | |
7e84e1c7 | 230 | bool registered; |
21b55dbc SG |
231 | }; |
232 | ||
534ca6d6 JA |
233 | struct io_sq_data { |
234 | refcount_t refs; | |
69fb2131 JA |
235 | struct mutex lock; |
236 | ||
237 | /* ctx's that are using this sqd */ | |
238 | struct list_head ctx_list; | |
239 | struct list_head ctx_new_list; | |
240 | struct mutex ctx_lock; | |
241 | ||
534ca6d6 JA |
242 | struct task_struct *thread; |
243 | struct wait_queue_head wait; | |
244 | }; | |
245 | ||
2b188cc1 JA |
246 | struct io_ring_ctx { |
247 | struct { | |
248 | struct percpu_ref refs; | |
249 | } ____cacheline_aligned_in_smp; | |
250 | ||
251 | struct { | |
252 | unsigned int flags; | |
e1d85334 | 253 | unsigned int compat: 1; |
aad5d8da | 254 | unsigned int limit_mem: 1; |
e1d85334 RD |
255 | unsigned int cq_overflow_flushed: 1; |
256 | unsigned int drain_next: 1; | |
257 | unsigned int eventfd_async: 1; | |
21b55dbc | 258 | unsigned int restricted: 1; |
2b188cc1 | 259 | |
75b28aff HV |
260 | /* |
261 | * Ring buffer of indices into array of io_uring_sqe, which is | |
262 | * mmapped by the application using the IORING_OFF_SQES offset. | |
263 | * | |
264 | * This indirection could e.g. be used to assign fixed | |
265 | * io_uring_sqe entries to operations and only submit them to | |
266 | * the queue when needed. | |
267 | * | |
268 | * The kernel modifies neither the indices array nor the entries | |
269 | * array. | |
270 | */ | |
271 | u32 *sq_array; | |
2b188cc1 JA |
272 | unsigned cached_sq_head; |
273 | unsigned sq_entries; | |
274 | unsigned sq_mask; | |
6c271ce2 | 275 | unsigned sq_thread_idle; |
498ccd9e | 276 | unsigned cached_sq_dropped; |
206aefde | 277 | atomic_t cached_cq_overflow; |
ad3eb2c8 | 278 | unsigned long sq_check_overflow; |
de0617e4 JA |
279 | |
280 | struct list_head defer_list; | |
5262f567 | 281 | struct list_head timeout_list; |
1d7bb1d5 | 282 | struct list_head cq_overflow_list; |
fcb323cc JA |
283 | |
284 | wait_queue_head_t inflight_wait; | |
ad3eb2c8 | 285 | struct io_uring_sqe *sq_sqes; |
2b188cc1 JA |
286 | } ____cacheline_aligned_in_smp; |
287 | ||
206aefde JA |
288 | struct io_rings *rings; |
289 | ||
2b188cc1 | 290 | /* IO offload */ |
561fb04a | 291 | struct io_wq *io_wq; |
2aede0e4 JA |
292 | |
293 | /* | |
294 | * For SQPOLL usage - we hold a reference to the parent task, so we | |
295 | * have access to the ->files | |
296 | */ | |
297 | struct task_struct *sqo_task; | |
298 | ||
299 | /* Only used for accounting purposes */ | |
300 | struct mm_struct *mm_account; | |
301 | ||
534ca6d6 JA |
302 | struct io_sq_data *sq_data; /* if using sq thread polling */ |
303 | ||
90554200 | 304 | struct wait_queue_head sqo_sq_wait; |
6a779382 | 305 | struct wait_queue_entry sqo_wait_entry; |
69fb2131 | 306 | struct list_head sqd_list; |
75b28aff | 307 | |
6b06314c JA |
308 | /* |
309 | * If used, fixed file set. Writers must ensure that ->refs is dead, | |
310 | * readers must ensure that ->refs is alive as long as the file* is | |
311 | * used. Only updated through io_uring_register(2). | |
312 | */ | |
05f3fb3c | 313 | struct fixed_file_data *file_data; |
6b06314c JA |
314 | unsigned nr_user_files; |
315 | ||
edafccee JA |
316 | /* if used, fixed mapped user buffers */ |
317 | unsigned nr_user_bufs; | |
318 | struct io_mapped_ubuf *user_bufs; | |
319 | ||
2b188cc1 JA |
320 | struct user_struct *user; |
321 | ||
0b8c0ec7 | 322 | const struct cred *creds; |
181e448d | 323 | |
0f158b4c JA |
324 | struct completion ref_comp; |
325 | struct completion sq_thread_comp; | |
206aefde | 326 | |
0ddf92e8 JA |
327 | /* if all else fails... */ |
328 | struct io_kiocb *fallback_req; | |
329 | ||
206aefde JA |
330 | #if defined(CONFIG_UNIX) |
331 | struct socket *ring_sock; | |
332 | #endif | |
333 | ||
5a2e745d JA |
334 | struct idr io_buffer_idr; |
335 | ||
071698e1 JA |
336 | struct idr personality_idr; |
337 | ||
206aefde JA |
338 | struct { |
339 | unsigned cached_cq_tail; | |
340 | unsigned cq_entries; | |
341 | unsigned cq_mask; | |
342 | atomic_t cq_timeouts; | |
ad3eb2c8 | 343 | unsigned long cq_check_overflow; |
206aefde JA |
344 | struct wait_queue_head cq_wait; |
345 | struct fasync_struct *cq_fasync; | |
346 | struct eventfd_ctx *cq_ev_fd; | |
347 | } ____cacheline_aligned_in_smp; | |
2b188cc1 JA |
348 | |
349 | struct { | |
350 | struct mutex uring_lock; | |
351 | wait_queue_head_t wait; | |
352 | } ____cacheline_aligned_in_smp; | |
353 | ||
354 | struct { | |
355 | spinlock_t completion_lock; | |
e94f141b | 356 | |
def596e9 | 357 | /* |
540e32a0 | 358 | * ->iopoll_list is protected by the ctx->uring_lock for |
def596e9 JA |
359 | * io_uring instances that don't use IORING_SETUP_SQPOLL. |
360 | * For SQPOLL, only the single threaded io_sq_thread() will | |
361 | * manipulate the list, hence no extra locking is needed there. | |
362 | */ | |
540e32a0 | 363 | struct list_head iopoll_list; |
78076bb6 JA |
364 | struct hlist_head *cancel_hash; |
365 | unsigned cancel_hash_bits; | |
e94f141b | 366 | bool poll_multi_file; |
31b51510 | 367 | |
fcb323cc JA |
368 | spinlock_t inflight_lock; |
369 | struct list_head inflight_list; | |
2b188cc1 | 370 | } ____cacheline_aligned_in_smp; |
85faa7b8 | 371 | |
4a38aed2 JA |
372 | struct delayed_work file_put_work; |
373 | struct llist_head file_put_llist; | |
374 | ||
85faa7b8 | 375 | struct work_struct exit_work; |
21b55dbc | 376 | struct io_restriction restrictions; |
2b188cc1 JA |
377 | }; |
378 | ||
09bb8394 JA |
379 | /* |
380 | * First field must be the file pointer in all the | |
381 | * iocb unions! See also 'struct kiocb' in <linux/fs.h> | |
382 | */ | |
221c5eb2 JA |
383 | struct io_poll_iocb { |
384 | struct file *file; | |
0969e783 JA |
385 | union { |
386 | struct wait_queue_head *head; | |
387 | u64 addr; | |
388 | }; | |
221c5eb2 | 389 | __poll_t events; |
8c838788 | 390 | bool done; |
221c5eb2 | 391 | bool canceled; |
392edb45 | 392 | struct wait_queue_entry wait; |
221c5eb2 JA |
393 | }; |
394 | ||
b5dba59e JA |
395 | struct io_close { |
396 | struct file *file; | |
397 | struct file *put_file; | |
398 | int fd; | |
399 | }; | |
400 | ||
ad8a48ac JA |
401 | struct io_timeout_data { |
402 | struct io_kiocb *req; | |
403 | struct hrtimer timer; | |
404 | struct timespec64 ts; | |
405 | enum hrtimer_mode mode; | |
406 | }; | |
407 | ||
8ed8d3c3 JA |
408 | struct io_accept { |
409 | struct file *file; | |
410 | struct sockaddr __user *addr; | |
411 | int __user *addr_len; | |
412 | int flags; | |
09952e3e | 413 | unsigned long nofile; |
8ed8d3c3 JA |
414 | }; |
415 | ||
416 | struct io_sync { | |
417 | struct file *file; | |
418 | loff_t len; | |
419 | loff_t off; | |
420 | int flags; | |
d63d1b5e | 421 | int mode; |
8ed8d3c3 JA |
422 | }; |
423 | ||
fbf23849 JA |
424 | struct io_cancel { |
425 | struct file *file; | |
426 | u64 addr; | |
427 | }; | |
428 | ||
b29472ee JA |
429 | struct io_timeout { |
430 | struct file *file; | |
431 | u64 addr; | |
432 | int flags; | |
bfe68a22 PB |
433 | u32 off; |
434 | u32 target_seq; | |
135fcde8 | 435 | struct list_head list; |
b29472ee JA |
436 | }; |
437 | ||
9adbd45d JA |
438 | struct io_rw { |
439 | /* NOTE: kiocb has the file as the first member, so don't do it here */ | |
440 | struct kiocb kiocb; | |
441 | u64 addr; | |
442 | u64 len; | |
443 | }; | |
444 | ||
3fbb51c1 JA |
445 | struct io_connect { |
446 | struct file *file; | |
447 | struct sockaddr __user *addr; | |
448 | int addr_len; | |
449 | }; | |
450 | ||
e47293fd JA |
451 | struct io_sr_msg { |
452 | struct file *file; | |
fddaface | 453 | union { |
270a5940 | 454 | struct user_msghdr __user *umsg; |
fddaface JA |
455 | void __user *buf; |
456 | }; | |
e47293fd | 457 | int msg_flags; |
bcda7baa | 458 | int bgid; |
fddaface | 459 | size_t len; |
bcda7baa | 460 | struct io_buffer *kbuf; |
e47293fd JA |
461 | }; |
462 | ||
15b71abe JA |
463 | struct io_open { |
464 | struct file *file; | |
465 | int dfd; | |
15b71abe | 466 | struct filename *filename; |
c12cedf2 | 467 | struct open_how how; |
4022e7af | 468 | unsigned long nofile; |
15b71abe JA |
469 | }; |
470 | ||
05f3fb3c JA |
471 | struct io_files_update { |
472 | struct file *file; | |
473 | u64 arg; | |
474 | u32 nr_args; | |
475 | u32 offset; | |
476 | }; | |
477 | ||
4840e418 JA |
478 | struct io_fadvise { |
479 | struct file *file; | |
480 | u64 offset; | |
481 | u32 len; | |
482 | u32 advice; | |
483 | }; | |
484 | ||
c1ca757b JA |
485 | struct io_madvise { |
486 | struct file *file; | |
487 | u64 addr; | |
488 | u32 len; | |
489 | u32 advice; | |
490 | }; | |
491 | ||
3e4827b0 JA |
492 | struct io_epoll { |
493 | struct file *file; | |
494 | int epfd; | |
495 | int op; | |
496 | int fd; | |
497 | struct epoll_event event; | |
e47293fd JA |
498 | }; |
499 | ||
7d67af2c PB |
500 | struct io_splice { |
501 | struct file *file_out; | |
502 | struct file *file_in; | |
503 | loff_t off_out; | |
504 | loff_t off_in; | |
505 | u64 len; | |
506 | unsigned int flags; | |
507 | }; | |
508 | ||
ddf0322d JA |
509 | struct io_provide_buf { |
510 | struct file *file; | |
511 | __u64 addr; | |
512 | __s32 len; | |
513 | __u32 bgid; | |
514 | __u16 nbufs; | |
515 | __u16 bid; | |
516 | }; | |
517 | ||
1d9e1288 BM |
518 | struct io_statx { |
519 | struct file *file; | |
520 | int dfd; | |
521 | unsigned int mask; | |
522 | unsigned int flags; | |
e62753e4 | 523 | const char __user *filename; |
1d9e1288 BM |
524 | struct statx __user *buffer; |
525 | }; | |
526 | ||
3ca405eb PB |
527 | struct io_completion { |
528 | struct file *file; | |
529 | struct list_head list; | |
0f7e466b | 530 | int cflags; |
3ca405eb PB |
531 | }; |
532 | ||
f499a021 JA |
533 | struct io_async_connect { |
534 | struct sockaddr_storage address; | |
535 | }; | |
536 | ||
03b1230c JA |
537 | struct io_async_msghdr { |
538 | struct iovec fast_iov[UIO_FASTIOV]; | |
539 | struct iovec *iov; | |
540 | struct sockaddr __user *uaddr; | |
541 | struct msghdr msg; | |
b537916c | 542 | struct sockaddr_storage addr; |
03b1230c JA |
543 | }; |
544 | ||
f67676d1 JA |
545 | struct io_async_rw { |
546 | struct iovec fast_iov[UIO_FASTIOV]; | |
ff6165b2 JA |
547 | const struct iovec *free_iovec; |
548 | struct iov_iter iter; | |
227c0c96 | 549 | size_t bytes_done; |
bcf5a063 | 550 | struct wait_page_queue wpq; |
f67676d1 JA |
551 | }; |
552 | ||
1a6b74fc | 553 | struct io_async_ctx { |
f67676d1 JA |
554 | union { |
555 | struct io_async_rw rw; | |
03b1230c | 556 | struct io_async_msghdr msg; |
f499a021 | 557 | struct io_async_connect connect; |
2d28390a | 558 | struct io_timeout_data timeout; |
f67676d1 | 559 | }; |
1a6b74fc JA |
560 | }; |
561 | ||
6b47ee6e PB |
562 | enum { |
563 | REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT, | |
564 | REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT, | |
565 | REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT, | |
566 | REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT, | |
567 | REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT, | |
bcda7baa | 568 | REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT, |
6b47ee6e | 569 | |
dea3b49c | 570 | REQ_F_LINK_HEAD_BIT, |
6b47ee6e PB |
571 | REQ_F_FAIL_LINK_BIT, |
572 | REQ_F_INFLIGHT_BIT, | |
573 | REQ_F_CUR_POS_BIT, | |
574 | REQ_F_NOWAIT_BIT, | |
6b47ee6e | 575 | REQ_F_LINK_TIMEOUT_BIT, |
6b47ee6e | 576 | REQ_F_ISREG_BIT, |
6b47ee6e | 577 | REQ_F_COMP_LOCKED_BIT, |
99bc4c38 | 578 | REQ_F_NEED_CLEANUP_BIT, |
d7718a9d | 579 | REQ_F_POLLED_BIT, |
bcda7baa | 580 | REQ_F_BUFFER_SELECTED_BIT, |
5b0bbee4 | 581 | REQ_F_NO_FILE_TABLE_BIT, |
7cdaf587 | 582 | REQ_F_WORK_INITIALIZED_BIT, |
84557871 JA |
583 | |
584 | /* not a real bit, just to check we're not overflowing the space */ | |
585 | __REQ_F_LAST_BIT, | |
6b47ee6e PB |
586 | }; |
587 | ||
588 | enum { | |
589 | /* ctx owns file */ | |
590 | REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT), | |
591 | /* drain existing IO first */ | |
592 | REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT), | |
593 | /* linked sqes */ | |
594 | REQ_F_LINK = BIT(REQ_F_LINK_BIT), | |
595 | /* doesn't sever on completion < 0 */ | |
596 | REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT), | |
597 | /* IOSQE_ASYNC */ | |
598 | REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT), | |
bcda7baa JA |
599 | /* IOSQE_BUFFER_SELECT */ |
600 | REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT), | |
6b47ee6e | 601 | |
dea3b49c PB |
602 | /* head of a link */ |
603 | REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT), | |
6b47ee6e PB |
604 | /* fail rest of links */ |
605 | REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT), | |
606 | /* on inflight list */ | |
607 | REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT), | |
608 | /* read/write uses file position */ | |
609 | REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT), | |
610 | /* must not punt to workers */ | |
611 | REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT), | |
6b47ee6e PB |
612 | /* has linked timeout */ |
613 | REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT), | |
6b47ee6e PB |
614 | /* regular file */ |
615 | REQ_F_ISREG = BIT(REQ_F_ISREG_BIT), | |
6b47ee6e PB |
616 | /* completion under lock */ |
617 | REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT), | |
99bc4c38 PB |
618 | /* needs cleanup */ |
619 | REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT), | |
d7718a9d JA |
620 | /* already went through poll handler */ |
621 | REQ_F_POLLED = BIT(REQ_F_POLLED_BIT), | |
bcda7baa JA |
622 | /* buffer already selected */ |
623 | REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT), | |
5b0bbee4 JA |
624 | /* doesn't need file table for this request */ |
625 | REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT), | |
7cdaf587 XW |
626 | /* io_wq_work is initialized */ |
627 | REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT), | |
d7718a9d JA |
628 | }; |
629 | ||
630 | struct async_poll { | |
631 | struct io_poll_iocb poll; | |
807abcb0 | 632 | struct io_poll_iocb *double_poll; |
6b47ee6e PB |
633 | }; |
634 | ||
09bb8394 JA |
635 | /* |
636 | * NOTE! Each of the iocb union members has the file pointer | |
637 | * as the first entry in their struct definition. So you can | |
638 | * access the file pointer through any of the sub-structs, | |
639 | * or directly as just 'ki_filp' in this struct. | |
640 | */ | |
2b188cc1 | 641 | struct io_kiocb { |
221c5eb2 | 642 | union { |
09bb8394 | 643 | struct file *file; |
9adbd45d | 644 | struct io_rw rw; |
221c5eb2 | 645 | struct io_poll_iocb poll; |
8ed8d3c3 JA |
646 | struct io_accept accept; |
647 | struct io_sync sync; | |
fbf23849 | 648 | struct io_cancel cancel; |
b29472ee | 649 | struct io_timeout timeout; |
3fbb51c1 | 650 | struct io_connect connect; |
e47293fd | 651 | struct io_sr_msg sr_msg; |
15b71abe | 652 | struct io_open open; |
b5dba59e | 653 | struct io_close close; |
05f3fb3c | 654 | struct io_files_update files_update; |
4840e418 | 655 | struct io_fadvise fadvise; |
c1ca757b | 656 | struct io_madvise madvise; |
3e4827b0 | 657 | struct io_epoll epoll; |
7d67af2c | 658 | struct io_splice splice; |
ddf0322d | 659 | struct io_provide_buf pbuf; |
1d9e1288 | 660 | struct io_statx statx; |
3ca405eb PB |
661 | /* use only after cleaning per-op data, see io_clean_op() */ |
662 | struct io_completion compl; | |
221c5eb2 | 663 | }; |
2b188cc1 | 664 | |
1a6b74fc | 665 | struct io_async_ctx *io; |
d625c6ee | 666 | u8 opcode; |
65a6543d XW |
667 | /* polled IO has completed */ |
668 | u8 iopoll_completed; | |
2b188cc1 | 669 | |
4f4eeba8 | 670 | u16 buf_index; |
9cf7c104 | 671 | u32 result; |
4f4eeba8 | 672 | |
010e8e6b PB |
673 | struct io_ring_ctx *ctx; |
674 | unsigned int flags; | |
675 | refcount_t refs; | |
676 | struct task_struct *task; | |
677 | u64 user_data; | |
d7718a9d | 678 | |
010e8e6b | 679 | struct list_head link_list; |
fcb323cc | 680 | |
d21ffe7e PB |
681 | /* |
682 | * 1. used with ctx->iopoll_list with reads/writes | |
683 | * 2. to track reqs with ->files (see io_op_def::file_table) | |
684 | */ | |
010e8e6b PB |
685 | struct list_head inflight_entry; |
686 | ||
687 | struct percpu_ref *fixed_file_refs; | |
688 | struct callback_head task_work; | |
689 | /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */ | |
690 | struct hlist_node hash_node; | |
691 | struct async_poll *apoll; | |
692 | struct io_wq_work work; | |
2b188cc1 | 693 | }; |
05589553 | 694 | |
27dc8338 PB |
695 | struct io_defer_entry { |
696 | struct list_head list; | |
697 | struct io_kiocb *req; | |
9cf7c104 | 698 | u32 seq; |
2b188cc1 JA |
699 | }; |
700 | ||
def596e9 | 701 | #define IO_IOPOLL_BATCH 8 |
2b188cc1 | 702 | |
013538bd JA |
703 | struct io_comp_state { |
704 | unsigned int nr; | |
705 | struct list_head list; | |
706 | struct io_ring_ctx *ctx; | |
707 | }; | |
708 | ||
9a56a232 JA |
709 | struct io_submit_state { |
710 | struct blk_plug plug; | |
711 | ||
2579f913 JA |
712 | /* |
713 | * io_kiocb alloc cache | |
714 | */ | |
715 | void *reqs[IO_IOPOLL_BATCH]; | |
6c8a3134 | 716 | unsigned int free_reqs; |
2579f913 | 717 | |
013538bd JA |
718 | /* |
719 | * Batch completion logic | |
720 | */ | |
721 | struct io_comp_state comp; | |
722 | ||
9a56a232 JA |
723 | /* |
724 | * File reference cache | |
725 | */ | |
726 | struct file *file; | |
727 | unsigned int fd; | |
728 | unsigned int has_refs; | |
9a56a232 JA |
729 | unsigned int ios_left; |
730 | }; | |
731 | ||
d3656344 JA |
732 | struct io_op_def { |
733 | /* needs req->io allocated for deferral/async */ | |
734 | unsigned async_ctx : 1; | |
735 | /* needs current->mm setup, does mm access */ | |
736 | unsigned needs_mm : 1; | |
737 | /* needs req->file assigned */ | |
738 | unsigned needs_file : 1; | |
fd2206e4 JA |
739 | /* don't fail if file grab fails */ |
740 | unsigned needs_file_no_error : 1; | |
d3656344 JA |
741 | /* hash wq insertion if file is a regular file */ |
742 | unsigned hash_reg_file : 1; | |
743 | /* unbound wq insertion if file is a non-regular file */ | |
744 | unsigned unbound_nonreg_file : 1; | |
66f4af93 JA |
745 | /* opcode is not supported by this kernel */ |
746 | unsigned not_supported : 1; | |
f86cd20c JA |
747 | /* needs file table */ |
748 | unsigned file_table : 1; | |
ff002b30 JA |
749 | /* needs ->fs */ |
750 | unsigned needs_fs : 1; | |
8a72758c JA |
751 | /* set if opcode supports polled "wait" */ |
752 | unsigned pollin : 1; | |
753 | unsigned pollout : 1; | |
bcda7baa JA |
754 | /* op supports buffer selection */ |
755 | unsigned buffer_select : 1; | |
57f1a649 | 756 | unsigned needs_fsize : 1; |
d3656344 JA |
757 | }; |
758 | ||
738277ad | 759 | static const struct io_op_def io_op_defs[] __read_mostly = { |
0463b6c5 PB |
760 | [IORING_OP_NOP] = {}, |
761 | [IORING_OP_READV] = { | |
d3656344 JA |
762 | .async_ctx = 1, |
763 | .needs_mm = 1, | |
764 | .needs_file = 1, | |
765 | .unbound_nonreg_file = 1, | |
8a72758c | 766 | .pollin = 1, |
4d954c25 | 767 | .buffer_select = 1, |
d3656344 | 768 | }, |
0463b6c5 | 769 | [IORING_OP_WRITEV] = { |
d3656344 JA |
770 | .async_ctx = 1, |
771 | .needs_mm = 1, | |
772 | .needs_file = 1, | |
773 | .hash_reg_file = 1, | |
774 | .unbound_nonreg_file = 1, | |
8a72758c | 775 | .pollout = 1, |
57f1a649 | 776 | .needs_fsize = 1, |
d3656344 | 777 | }, |
0463b6c5 | 778 | [IORING_OP_FSYNC] = { |
d3656344 JA |
779 | .needs_file = 1, |
780 | }, | |
0463b6c5 | 781 | [IORING_OP_READ_FIXED] = { |
d3656344 JA |
782 | .needs_file = 1, |
783 | .unbound_nonreg_file = 1, | |
8a72758c | 784 | .pollin = 1, |
d3656344 | 785 | }, |
0463b6c5 | 786 | [IORING_OP_WRITE_FIXED] = { |
d3656344 JA |
787 | .needs_file = 1, |
788 | .hash_reg_file = 1, | |
789 | .unbound_nonreg_file = 1, | |
8a72758c | 790 | .pollout = 1, |
57f1a649 | 791 | .needs_fsize = 1, |
d3656344 | 792 | }, |
0463b6c5 | 793 | [IORING_OP_POLL_ADD] = { |
d3656344 JA |
794 | .needs_file = 1, |
795 | .unbound_nonreg_file = 1, | |
796 | }, | |
0463b6c5 PB |
797 | [IORING_OP_POLL_REMOVE] = {}, |
798 | [IORING_OP_SYNC_FILE_RANGE] = { | |
d3656344 JA |
799 | .needs_file = 1, |
800 | }, | |
0463b6c5 | 801 | [IORING_OP_SENDMSG] = { |
d3656344 JA |
802 | .async_ctx = 1, |
803 | .needs_mm = 1, | |
804 | .needs_file = 1, | |
805 | .unbound_nonreg_file = 1, | |
ff002b30 | 806 | .needs_fs = 1, |
8a72758c | 807 | .pollout = 1, |
d3656344 | 808 | }, |
0463b6c5 | 809 | [IORING_OP_RECVMSG] = { |
d3656344 JA |
810 | .async_ctx = 1, |
811 | .needs_mm = 1, | |
812 | .needs_file = 1, | |
813 | .unbound_nonreg_file = 1, | |
ff002b30 | 814 | .needs_fs = 1, |
8a72758c | 815 | .pollin = 1, |
52de1fe1 | 816 | .buffer_select = 1, |
d3656344 | 817 | }, |
0463b6c5 | 818 | [IORING_OP_TIMEOUT] = { |
d3656344 JA |
819 | .async_ctx = 1, |
820 | .needs_mm = 1, | |
821 | }, | |
0463b6c5 PB |
822 | [IORING_OP_TIMEOUT_REMOVE] = {}, |
823 | [IORING_OP_ACCEPT] = { | |
d3656344 JA |
824 | .needs_mm = 1, |
825 | .needs_file = 1, | |
826 | .unbound_nonreg_file = 1, | |
f86cd20c | 827 | .file_table = 1, |
8a72758c | 828 | .pollin = 1, |
d3656344 | 829 | }, |
0463b6c5 PB |
830 | [IORING_OP_ASYNC_CANCEL] = {}, |
831 | [IORING_OP_LINK_TIMEOUT] = { | |
d3656344 JA |
832 | .async_ctx = 1, |
833 | .needs_mm = 1, | |
834 | }, | |
0463b6c5 | 835 | [IORING_OP_CONNECT] = { |
d3656344 JA |
836 | .async_ctx = 1, |
837 | .needs_mm = 1, | |
838 | .needs_file = 1, | |
839 | .unbound_nonreg_file = 1, | |
8a72758c | 840 | .pollout = 1, |
d3656344 | 841 | }, |
0463b6c5 | 842 | [IORING_OP_FALLOCATE] = { |
d3656344 | 843 | .needs_file = 1, |
57f1a649 | 844 | .needs_fsize = 1, |
d3656344 | 845 | }, |
0463b6c5 | 846 | [IORING_OP_OPENAT] = { |
f86cd20c | 847 | .file_table = 1, |
ff002b30 | 848 | .needs_fs = 1, |
d3656344 | 849 | }, |
0463b6c5 | 850 | [IORING_OP_CLOSE] = { |
fd2206e4 JA |
851 | .needs_file = 1, |
852 | .needs_file_no_error = 1, | |
f86cd20c | 853 | .file_table = 1, |
d3656344 | 854 | }, |
0463b6c5 | 855 | [IORING_OP_FILES_UPDATE] = { |
d3656344 | 856 | .needs_mm = 1, |
f86cd20c | 857 | .file_table = 1, |
d3656344 | 858 | }, |
0463b6c5 | 859 | [IORING_OP_STATX] = { |
d3656344 | 860 | .needs_mm = 1, |
ff002b30 | 861 | .needs_fs = 1, |
5b0bbee4 | 862 | .file_table = 1, |
d3656344 | 863 | }, |
0463b6c5 | 864 | [IORING_OP_READ] = { |
3a6820f2 JA |
865 | .needs_mm = 1, |
866 | .needs_file = 1, | |
867 | .unbound_nonreg_file = 1, | |
8a72758c | 868 | .pollin = 1, |
bcda7baa | 869 | .buffer_select = 1, |
3a6820f2 | 870 | }, |
0463b6c5 | 871 | [IORING_OP_WRITE] = { |
3a6820f2 JA |
872 | .needs_mm = 1, |
873 | .needs_file = 1, | |
874 | .unbound_nonreg_file = 1, | |
8a72758c | 875 | .pollout = 1, |
57f1a649 | 876 | .needs_fsize = 1, |
3a6820f2 | 877 | }, |
0463b6c5 | 878 | [IORING_OP_FADVISE] = { |
4840e418 JA |
879 | .needs_file = 1, |
880 | }, | |
0463b6c5 | 881 | [IORING_OP_MADVISE] = { |
c1ca757b JA |
882 | .needs_mm = 1, |
883 | }, | |
0463b6c5 | 884 | [IORING_OP_SEND] = { |
fddaface JA |
885 | .needs_mm = 1, |
886 | .needs_file = 1, | |
887 | .unbound_nonreg_file = 1, | |
8a72758c | 888 | .pollout = 1, |
fddaface | 889 | }, |
0463b6c5 | 890 | [IORING_OP_RECV] = { |
fddaface JA |
891 | .needs_mm = 1, |
892 | .needs_file = 1, | |
893 | .unbound_nonreg_file = 1, | |
8a72758c | 894 | .pollin = 1, |
bcda7baa | 895 | .buffer_select = 1, |
fddaface | 896 | }, |
0463b6c5 | 897 | [IORING_OP_OPENAT2] = { |
f86cd20c | 898 | .file_table = 1, |
ff002b30 | 899 | .needs_fs = 1, |
cebdb986 | 900 | }, |
3e4827b0 JA |
901 | [IORING_OP_EPOLL_CTL] = { |
902 | .unbound_nonreg_file = 1, | |
903 | .file_table = 1, | |
904 | }, | |
7d67af2c PB |
905 | [IORING_OP_SPLICE] = { |
906 | .needs_file = 1, | |
907 | .hash_reg_file = 1, | |
908 | .unbound_nonreg_file = 1, | |
ddf0322d JA |
909 | }, |
910 | [IORING_OP_PROVIDE_BUFFERS] = {}, | |
067524e9 | 911 | [IORING_OP_REMOVE_BUFFERS] = {}, |
f2a8d5c7 PB |
912 | [IORING_OP_TEE] = { |
913 | .needs_file = 1, | |
914 | .hash_reg_file = 1, | |
915 | .unbound_nonreg_file = 1, | |
916 | }, | |
d3656344 JA |
917 | }; |
918 | ||
2e0464d4 BM |
919 | enum io_mem_account { |
920 | ACCT_LOCKED, | |
921 | ACCT_PINNED, | |
922 | }; | |
923 | ||
81b68a5c PB |
924 | static void __io_complete_rw(struct io_kiocb *req, long res, long res2, |
925 | struct io_comp_state *cs); | |
78e19bbe | 926 | static void io_cqring_fill_event(struct io_kiocb *req, long res); |
ec9c02ad | 927 | static void io_put_req(struct io_kiocb *req); |
c40f6379 | 928 | static void io_double_put_req(struct io_kiocb *req); |
978db57e | 929 | static void __io_double_put_req(struct io_kiocb *req); |
94ae5e77 | 930 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req); |
7271ef3a | 931 | static void __io_queue_linked_timeout(struct io_kiocb *req); |
94ae5e77 | 932 | static void io_queue_linked_timeout(struct io_kiocb *req); |
05f3fb3c JA |
933 | static int __io_sqe_files_update(struct io_ring_ctx *ctx, |
934 | struct io_uring_files_update *ip, | |
935 | unsigned nr_args); | |
f56040b8 | 936 | static int io_prep_work_files(struct io_kiocb *req); |
3ca405eb | 937 | static void __io_clean_op(struct io_kiocb *req); |
b41e9852 JA |
938 | static int io_file_get(struct io_submit_state *state, struct io_kiocb *req, |
939 | int fd, struct file **out_file, bool fixed); | |
940 | static void __io_queue_sqe(struct io_kiocb *req, | |
f13fad7b JA |
941 | const struct io_uring_sqe *sqe, |
942 | struct io_comp_state *cs); | |
4349f30e | 943 | static void io_file_put_work(struct work_struct *work); |
de0617e4 | 944 | |
b63534c4 JA |
945 | static ssize_t io_import_iovec(int rw, struct io_kiocb *req, |
946 | struct iovec **iovec, struct iov_iter *iter, | |
947 | bool needs_lock); | |
ff6165b2 JA |
948 | static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec, |
949 | const struct iovec *fast_iov, | |
227c0c96 | 950 | struct iov_iter *iter, bool force); |
de0617e4 | 951 | |
2b188cc1 JA |
952 | static struct kmem_cache *req_cachep; |
953 | ||
738277ad | 954 | static const struct file_operations io_uring_fops __read_mostly; |
2b188cc1 JA |
955 | |
956 | struct sock *io_uring_get_socket(struct file *file) | |
957 | { | |
958 | #if defined(CONFIG_UNIX) | |
959 | if (file->f_op == &io_uring_fops) { | |
960 | struct io_ring_ctx *ctx = file->private_data; | |
961 | ||
962 | return ctx->ring_sock->sk; | |
963 | } | |
964 | #endif | |
965 | return NULL; | |
966 | } | |
967 | EXPORT_SYMBOL(io_uring_get_socket); | |
968 | ||
3ca405eb PB |
969 | static inline void io_clean_op(struct io_kiocb *req) |
970 | { | |
bb175342 PB |
971 | if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED | |
972 | REQ_F_INFLIGHT)) | |
3ca405eb PB |
973 | __io_clean_op(req); |
974 | } | |
975 | ||
4349f30e | 976 | static void io_sq_thread_drop_mm(void) |
c40f6379 JA |
977 | { |
978 | struct mm_struct *mm = current->mm; | |
979 | ||
980 | if (mm) { | |
981 | kthread_unuse_mm(mm); | |
982 | mmput(mm); | |
983 | } | |
984 | } | |
985 | ||
986 | static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx) | |
987 | { | |
988 | if (!current->mm) { | |
cbcf7214 | 989 | if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) || |
2aede0e4 JA |
990 | !ctx->sqo_task->mm || |
991 | !mmget_not_zero(ctx->sqo_task->mm))) | |
c40f6379 | 992 | return -EFAULT; |
2aede0e4 | 993 | kthread_use_mm(ctx->sqo_task->mm); |
c40f6379 JA |
994 | } |
995 | ||
996 | return 0; | |
997 | } | |
998 | ||
999 | static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx, | |
1000 | struct io_kiocb *req) | |
1001 | { | |
1002 | if (!io_op_defs[req->opcode].needs_mm) | |
1003 | return 0; | |
1004 | return __io_sq_thread_acquire_mm(ctx); | |
1005 | } | |
1006 | ||
1007 | static inline void req_set_fail_links(struct io_kiocb *req) | |
1008 | { | |
1009 | if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK) | |
1010 | req->flags |= REQ_F_FAIL_LINK; | |
1011 | } | |
4a38aed2 | 1012 | |
7cdaf587 XW |
1013 | /* |
1014 | * Note: must call io_req_init_async() for the first time you | |
1015 | * touch any members of io_wq_work. | |
1016 | */ | |
1017 | static inline void io_req_init_async(struct io_kiocb *req) | |
1018 | { | |
1019 | if (req->flags & REQ_F_WORK_INITIALIZED) | |
1020 | return; | |
1021 | ||
1022 | memset(&req->work, 0, sizeof(req->work)); | |
1023 | req->flags |= REQ_F_WORK_INITIALIZED; | |
1024 | } | |
1025 | ||
0cdaf760 PB |
1026 | static inline bool io_async_submit(struct io_ring_ctx *ctx) |
1027 | { | |
1028 | return ctx->flags & IORING_SETUP_SQPOLL; | |
1029 | } | |
1030 | ||
2b188cc1 JA |
1031 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) |
1032 | { | |
1033 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
1034 | ||
0f158b4c | 1035 | complete(&ctx->ref_comp); |
2b188cc1 JA |
1036 | } |
1037 | ||
8eb7e2d0 PB |
1038 | static inline bool io_is_timeout_noseq(struct io_kiocb *req) |
1039 | { | |
1040 | return !req->timeout.off; | |
1041 | } | |
1042 | ||
2b188cc1 JA |
1043 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) |
1044 | { | |
1045 | struct io_ring_ctx *ctx; | |
78076bb6 | 1046 | int hash_bits; |
2b188cc1 JA |
1047 | |
1048 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
1049 | if (!ctx) | |
1050 | return NULL; | |
1051 | ||
0ddf92e8 JA |
1052 | ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL); |
1053 | if (!ctx->fallback_req) | |
1054 | goto err; | |
1055 | ||
78076bb6 JA |
1056 | /* |
1057 | * Use 5 bits less than the max cq entries, that should give us around | |
1058 | * 32 entries per hash list if totally full and uniformly spread. | |
1059 | */ | |
1060 | hash_bits = ilog2(p->cq_entries); | |
1061 | hash_bits -= 5; | |
1062 | if (hash_bits <= 0) | |
1063 | hash_bits = 1; | |
1064 | ctx->cancel_hash_bits = hash_bits; | |
1065 | ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head), | |
1066 | GFP_KERNEL); | |
1067 | if (!ctx->cancel_hash) | |
1068 | goto err; | |
1069 | __hash_init(ctx->cancel_hash, 1U << hash_bits); | |
1070 | ||
21482896 | 1071 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
206aefde JA |
1072 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) |
1073 | goto err; | |
2b188cc1 JA |
1074 | |
1075 | ctx->flags = p->flags; | |
90554200 | 1076 | init_waitqueue_head(&ctx->sqo_sq_wait); |
69fb2131 | 1077 | INIT_LIST_HEAD(&ctx->sqd_list); |
2b188cc1 | 1078 | init_waitqueue_head(&ctx->cq_wait); |
1d7bb1d5 | 1079 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
0f158b4c JA |
1080 | init_completion(&ctx->ref_comp); |
1081 | init_completion(&ctx->sq_thread_comp); | |
5a2e745d | 1082 | idr_init(&ctx->io_buffer_idr); |
071698e1 | 1083 | idr_init(&ctx->personality_idr); |
2b188cc1 JA |
1084 | mutex_init(&ctx->uring_lock); |
1085 | init_waitqueue_head(&ctx->wait); | |
1086 | spin_lock_init(&ctx->completion_lock); | |
540e32a0 | 1087 | INIT_LIST_HEAD(&ctx->iopoll_list); |
de0617e4 | 1088 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 1089 | INIT_LIST_HEAD(&ctx->timeout_list); |
fcb323cc JA |
1090 | init_waitqueue_head(&ctx->inflight_wait); |
1091 | spin_lock_init(&ctx->inflight_lock); | |
1092 | INIT_LIST_HEAD(&ctx->inflight_list); | |
4a38aed2 JA |
1093 | INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work); |
1094 | init_llist_head(&ctx->file_put_llist); | |
2b188cc1 | 1095 | return ctx; |
206aefde | 1096 | err: |
0ddf92e8 JA |
1097 | if (ctx->fallback_req) |
1098 | kmem_cache_free(req_cachep, ctx->fallback_req); | |
78076bb6 | 1099 | kfree(ctx->cancel_hash); |
206aefde JA |
1100 | kfree(ctx); |
1101 | return NULL; | |
2b188cc1 JA |
1102 | } |
1103 | ||
9cf7c104 | 1104 | static bool req_need_defer(struct io_kiocb *req, u32 seq) |
7adf4eaf | 1105 | { |
2bc9930e JA |
1106 | if (unlikely(req->flags & REQ_F_IO_DRAIN)) { |
1107 | struct io_ring_ctx *ctx = req->ctx; | |
a197f664 | 1108 | |
9cf7c104 | 1109 | return seq != ctx->cached_cq_tail |
31af27c7 | 1110 | + atomic_read(&ctx->cached_cq_overflow); |
2bc9930e | 1111 | } |
de0617e4 | 1112 | |
9d858b21 | 1113 | return false; |
de0617e4 JA |
1114 | } |
1115 | ||
de0617e4 | 1116 | static void __io_commit_cqring(struct io_ring_ctx *ctx) |
2b188cc1 | 1117 | { |
75b28aff | 1118 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 1119 | |
07910158 PB |
1120 | /* order cqe stores with ring update */ |
1121 | smp_store_release(&rings->cq.tail, ctx->cached_cq_tail); | |
2b188cc1 | 1122 | |
07910158 PB |
1123 | if (wq_has_sleeper(&ctx->cq_wait)) { |
1124 | wake_up_interruptible(&ctx->cq_wait); | |
1125 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
2b188cc1 JA |
1126 | } |
1127 | } | |
1128 | ||
51a4cc11 JA |
1129 | /* |
1130 | * Returns true if we need to defer file table putting. This can only happen | |
1131 | * from the error path with REQ_F_COMP_LOCKED set. | |
1132 | */ | |
1133 | static bool io_req_clean_work(struct io_kiocb *req) | |
18d9be1a | 1134 | { |
7cdaf587 | 1135 | if (!(req->flags & REQ_F_WORK_INITIALIZED)) |
51a4cc11 JA |
1136 | return false; |
1137 | ||
1138 | req->flags &= ~REQ_F_WORK_INITIALIZED; | |
7cdaf587 | 1139 | |
cccf0ee8 JA |
1140 | if (req->work.mm) { |
1141 | mmdrop(req->work.mm); | |
1142 | req->work.mm = NULL; | |
1143 | } | |
1144 | if (req->work.creds) { | |
1145 | put_cred(req->work.creds); | |
1146 | req->work.creds = NULL; | |
1147 | } | |
ff002b30 JA |
1148 | if (req->work.fs) { |
1149 | struct fs_struct *fs = req->work.fs; | |
1150 | ||
51a4cc11 JA |
1151 | if (req->flags & REQ_F_COMP_LOCKED) |
1152 | return true; | |
1153 | ||
ff002b30 JA |
1154 | spin_lock(&req->work.fs->lock); |
1155 | if (--fs->users) | |
1156 | fs = NULL; | |
1157 | spin_unlock(&req->work.fs->lock); | |
1158 | if (fs) | |
1159 | free_fs_struct(fs); | |
b65e0dd6 | 1160 | req->work.fs = NULL; |
ff002b30 | 1161 | } |
51a4cc11 JA |
1162 | |
1163 | return false; | |
561fb04a JA |
1164 | } |
1165 | ||
cbdcb435 | 1166 | static void io_prep_async_work(struct io_kiocb *req) |
18d9be1a | 1167 | { |
d3656344 | 1168 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
54a91f3b | 1169 | |
16d59803 PB |
1170 | io_req_init_async(req); |
1171 | ||
d3656344 | 1172 | if (req->flags & REQ_F_ISREG) { |
eefdf30f | 1173 | if (def->hash_reg_file || (req->ctx->flags & IORING_SETUP_IOPOLL)) |
8766dd51 | 1174 | io_wq_hash_work(&req->work, file_inode(req->file)); |
d3656344 JA |
1175 | } else { |
1176 | if (def->unbound_nonreg_file) | |
3529d8c2 | 1177 | req->work.flags |= IO_WQ_WORK_UNBOUND; |
54a91f3b | 1178 | } |
dca9cf8b PB |
1179 | if (!req->work.mm && def->needs_mm) { |
1180 | mmgrab(current->mm); | |
1181 | req->work.mm = current->mm; | |
1182 | } | |
1183 | if (!req->work.creds) | |
1184 | req->work.creds = get_current_cred(); | |
1185 | if (!req->work.fs && def->needs_fs) { | |
1186 | spin_lock(¤t->fs->lock); | |
1187 | if (!current->fs->in_exec) { | |
1188 | req->work.fs = current->fs; | |
1189 | req->work.fs->users++; | |
1190 | } else { | |
1191 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
1192 | } | |
1193 | spin_unlock(¤t->fs->lock); | |
1194 | } | |
57f1a649 PB |
1195 | if (def->needs_fsize) |
1196 | req->work.fsize = rlimit(RLIMIT_FSIZE); | |
1197 | else | |
1198 | req->work.fsize = RLIM_INFINITY; | |
561fb04a | 1199 | } |
cccf0ee8 | 1200 | |
cbdcb435 | 1201 | static void io_prep_async_link(struct io_kiocb *req) |
561fb04a | 1202 | { |
cbdcb435 | 1203 | struct io_kiocb *cur; |
54a91f3b | 1204 | |
cbdcb435 PB |
1205 | io_prep_async_work(req); |
1206 | if (req->flags & REQ_F_LINK_HEAD) | |
1207 | list_for_each_entry(cur, &req->link_list, link_list) | |
1208 | io_prep_async_work(cur); | |
561fb04a JA |
1209 | } |
1210 | ||
7271ef3a | 1211 | static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req) |
561fb04a | 1212 | { |
a197f664 | 1213 | struct io_ring_ctx *ctx = req->ctx; |
cbdcb435 | 1214 | struct io_kiocb *link = io_prep_linked_timeout(req); |
561fb04a | 1215 | |
8766dd51 PB |
1216 | trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req, |
1217 | &req->work, req->flags); | |
1218 | io_wq_enqueue(ctx->io_wq, &req->work); | |
7271ef3a | 1219 | return link; |
18d9be1a JA |
1220 | } |
1221 | ||
cbdcb435 PB |
1222 | static void io_queue_async_work(struct io_kiocb *req) |
1223 | { | |
7271ef3a JA |
1224 | struct io_kiocb *link; |
1225 | ||
cbdcb435 PB |
1226 | /* init ->work of the whole link before punting */ |
1227 | io_prep_async_link(req); | |
7271ef3a JA |
1228 | link = __io_queue_async_work(req); |
1229 | ||
1230 | if (link) | |
1231 | io_queue_linked_timeout(link); | |
cbdcb435 PB |
1232 | } |
1233 | ||
5262f567 JA |
1234 | static void io_kill_timeout(struct io_kiocb *req) |
1235 | { | |
1236 | int ret; | |
1237 | ||
2d28390a | 1238 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
5262f567 | 1239 | if (ret != -1) { |
01cec8c1 PB |
1240 | atomic_set(&req->ctx->cq_timeouts, |
1241 | atomic_read(&req->ctx->cq_timeouts) + 1); | |
135fcde8 | 1242 | list_del_init(&req->timeout.list); |
f0e20b89 | 1243 | req->flags |= REQ_F_COMP_LOCKED; |
78e19bbe | 1244 | io_cqring_fill_event(req, 0); |
ec9c02ad | 1245 | io_put_req(req); |
5262f567 JA |
1246 | } |
1247 | } | |
1248 | ||
f3606e3a JA |
1249 | static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk) |
1250 | { | |
1251 | struct io_ring_ctx *ctx = req->ctx; | |
1252 | ||
1253 | if (!tsk || req->task == tsk) | |
1254 | return true; | |
534ca6d6 JA |
1255 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
1256 | if (ctx->sq_data && req->task == ctx->sq_data->thread) | |
1257 | return true; | |
1258 | } | |
f3606e3a JA |
1259 | return false; |
1260 | } | |
1261 | ||
76e1b642 JA |
1262 | /* |
1263 | * Returns true if we found and killed one or more timeouts | |
1264 | */ | |
1265 | static bool io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk) | |
5262f567 JA |
1266 | { |
1267 | struct io_kiocb *req, *tmp; | |
76e1b642 | 1268 | int canceled = 0; |
5262f567 JA |
1269 | |
1270 | spin_lock_irq(&ctx->completion_lock); | |
f3606e3a | 1271 | list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) { |
76e1b642 | 1272 | if (io_task_match(req, tsk)) { |
f3606e3a | 1273 | io_kill_timeout(req); |
76e1b642 JA |
1274 | canceled++; |
1275 | } | |
f3606e3a | 1276 | } |
5262f567 | 1277 | spin_unlock_irq(&ctx->completion_lock); |
76e1b642 | 1278 | return canceled != 0; |
5262f567 JA |
1279 | } |
1280 | ||
04518945 | 1281 | static void __io_queue_deferred(struct io_ring_ctx *ctx) |
de0617e4 | 1282 | { |
04518945 | 1283 | do { |
27dc8338 PB |
1284 | struct io_defer_entry *de = list_first_entry(&ctx->defer_list, |
1285 | struct io_defer_entry, list); | |
7271ef3a | 1286 | struct io_kiocb *link; |
de0617e4 | 1287 | |
9cf7c104 | 1288 | if (req_need_defer(de->req, de->seq)) |
04518945 | 1289 | break; |
27dc8338 | 1290 | list_del_init(&de->list); |
cbdcb435 | 1291 | /* punt-init is done before queueing for defer */ |
7271ef3a JA |
1292 | link = __io_queue_async_work(de->req); |
1293 | if (link) { | |
1294 | __io_queue_linked_timeout(link); | |
1295 | /* drop submission reference */ | |
1296 | link->flags |= REQ_F_COMP_LOCKED; | |
1297 | io_put_req(link); | |
1298 | } | |
27dc8338 | 1299 | kfree(de); |
04518945 PB |
1300 | } while (!list_empty(&ctx->defer_list)); |
1301 | } | |
1302 | ||
360428f8 | 1303 | static void io_flush_timeouts(struct io_ring_ctx *ctx) |
de0617e4 | 1304 | { |
360428f8 PB |
1305 | while (!list_empty(&ctx->timeout_list)) { |
1306 | struct io_kiocb *req = list_first_entry(&ctx->timeout_list, | |
135fcde8 | 1307 | struct io_kiocb, timeout.list); |
de0617e4 | 1308 | |
8eb7e2d0 | 1309 | if (io_is_timeout_noseq(req)) |
360428f8 | 1310 | break; |
bfe68a22 PB |
1311 | if (req->timeout.target_seq != ctx->cached_cq_tail |
1312 | - atomic_read(&ctx->cq_timeouts)) | |
360428f8 | 1313 | break; |
bfe68a22 | 1314 | |
135fcde8 | 1315 | list_del_init(&req->timeout.list); |
5262f567 | 1316 | io_kill_timeout(req); |
360428f8 PB |
1317 | } |
1318 | } | |
5262f567 | 1319 | |
360428f8 PB |
1320 | static void io_commit_cqring(struct io_ring_ctx *ctx) |
1321 | { | |
1322 | io_flush_timeouts(ctx); | |
de0617e4 JA |
1323 | __io_commit_cqring(ctx); |
1324 | ||
04518945 PB |
1325 | if (unlikely(!list_empty(&ctx->defer_list))) |
1326 | __io_queue_deferred(ctx); | |
de0617e4 JA |
1327 | } |
1328 | ||
90554200 JA |
1329 | static inline bool io_sqring_full(struct io_ring_ctx *ctx) |
1330 | { | |
1331 | struct io_rings *r = ctx->rings; | |
1332 | ||
1333 | return READ_ONCE(r->sq.tail) - ctx->cached_sq_head == r->sq_ring_entries; | |
1334 | } | |
1335 | ||
2b188cc1 JA |
1336 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) |
1337 | { | |
75b28aff | 1338 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
1339 | unsigned tail; |
1340 | ||
1341 | tail = ctx->cached_cq_tail; | |
115e12e5 SB |
1342 | /* |
1343 | * writes to the cq entry need to come after reading head; the | |
1344 | * control dependency is enough as we're using WRITE_ONCE to | |
1345 | * fill the cq entry | |
1346 | */ | |
75b28aff | 1347 | if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries) |
2b188cc1 JA |
1348 | return NULL; |
1349 | ||
1350 | ctx->cached_cq_tail++; | |
75b28aff | 1351 | return &rings->cqes[tail & ctx->cq_mask]; |
2b188cc1 JA |
1352 | } |
1353 | ||
f2842ab5 JA |
1354 | static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx) |
1355 | { | |
f0b493e6 JA |
1356 | if (!ctx->cq_ev_fd) |
1357 | return false; | |
7e55a19c SG |
1358 | if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) |
1359 | return false; | |
f2842ab5 JA |
1360 | if (!ctx->eventfd_async) |
1361 | return true; | |
b41e9852 | 1362 | return io_wq_current_is_worker(); |
f2842ab5 JA |
1363 | } |
1364 | ||
b41e9852 | 1365 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
1d7bb1d5 JA |
1366 | { |
1367 | if (waitqueue_active(&ctx->wait)) | |
1368 | wake_up(&ctx->wait); | |
534ca6d6 JA |
1369 | if (ctx->sq_data && waitqueue_active(&ctx->sq_data->wait)) |
1370 | wake_up(&ctx->sq_data->wait); | |
b41e9852 | 1371 | if (io_should_trigger_evfd(ctx)) |
1d7bb1d5 JA |
1372 | eventfd_signal(ctx->cq_ev_fd, 1); |
1373 | } | |
1374 | ||
46930143 PB |
1375 | static void io_cqring_mark_overflow(struct io_ring_ctx *ctx) |
1376 | { | |
1377 | if (list_empty(&ctx->cq_overflow_list)) { | |
1378 | clear_bit(0, &ctx->sq_check_overflow); | |
1379 | clear_bit(0, &ctx->cq_check_overflow); | |
1380 | ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW; | |
1381 | } | |
1382 | } | |
1383 | ||
e6c8aa9a JA |
1384 | static inline bool io_match_files(struct io_kiocb *req, |
1385 | struct files_struct *files) | |
1386 | { | |
1387 | if (!files) | |
1388 | return true; | |
1389 | if (req->flags & REQ_F_WORK_INITIALIZED) | |
1390 | return req->work.files == files; | |
1391 | return false; | |
1392 | } | |
1393 | ||
c4a2ed72 | 1394 | /* Returns true if there are no backlogged entries after the flush */ |
e6c8aa9a JA |
1395 | static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force, |
1396 | struct task_struct *tsk, | |
1397 | struct files_struct *files) | |
1d7bb1d5 JA |
1398 | { |
1399 | struct io_rings *rings = ctx->rings; | |
e6c8aa9a | 1400 | struct io_kiocb *req, *tmp; |
1d7bb1d5 | 1401 | struct io_uring_cqe *cqe; |
1d7bb1d5 JA |
1402 | unsigned long flags; |
1403 | LIST_HEAD(list); | |
1404 | ||
1405 | if (!force) { | |
1406 | if (list_empty_careful(&ctx->cq_overflow_list)) | |
c4a2ed72 | 1407 | return true; |
1d7bb1d5 JA |
1408 | if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) == |
1409 | rings->cq_ring_entries)) | |
c4a2ed72 | 1410 | return false; |
1d7bb1d5 JA |
1411 | } |
1412 | ||
1413 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
1414 | ||
1415 | /* if force is set, the ring is going away. always drop after that */ | |
1416 | if (force) | |
69b3e546 | 1417 | ctx->cq_overflow_flushed = 1; |
1d7bb1d5 | 1418 | |
c4a2ed72 | 1419 | cqe = NULL; |
e6c8aa9a JA |
1420 | list_for_each_entry_safe(req, tmp, &ctx->cq_overflow_list, compl.list) { |
1421 | if (tsk && req->task != tsk) | |
1422 | continue; | |
1423 | if (!io_match_files(req, files)) | |
1424 | continue; | |
1425 | ||
1d7bb1d5 JA |
1426 | cqe = io_get_cqring(ctx); |
1427 | if (!cqe && !force) | |
1428 | break; | |
1429 | ||
40d8ddd4 | 1430 | list_move(&req->compl.list, &list); |
1d7bb1d5 JA |
1431 | if (cqe) { |
1432 | WRITE_ONCE(cqe->user_data, req->user_data); | |
1433 | WRITE_ONCE(cqe->res, req->result); | |
0f7e466b | 1434 | WRITE_ONCE(cqe->flags, req->compl.cflags); |
1d7bb1d5 JA |
1435 | } else { |
1436 | WRITE_ONCE(ctx->rings->cq_overflow, | |
1437 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
1438 | } | |
1439 | } | |
1440 | ||
1441 | io_commit_cqring(ctx); | |
46930143 PB |
1442 | io_cqring_mark_overflow(ctx); |
1443 | ||
1d7bb1d5 JA |
1444 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
1445 | io_cqring_ev_posted(ctx); | |
1446 | ||
1447 | while (!list_empty(&list)) { | |
40d8ddd4 PB |
1448 | req = list_first_entry(&list, struct io_kiocb, compl.list); |
1449 | list_del(&req->compl.list); | |
ec9c02ad | 1450 | io_put_req(req); |
1d7bb1d5 | 1451 | } |
c4a2ed72 JA |
1452 | |
1453 | return cqe != NULL; | |
1d7bb1d5 JA |
1454 | } |
1455 | ||
bcda7baa | 1456 | static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags) |
2b188cc1 | 1457 | { |
78e19bbe | 1458 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
1459 | struct io_uring_cqe *cqe; |
1460 | ||
78e19bbe | 1461 | trace_io_uring_complete(ctx, req->user_data, res); |
51c3ff62 | 1462 | |
2b188cc1 JA |
1463 | /* |
1464 | * If we can't get a cq entry, userspace overflowed the | |
1465 | * submission (by quite a lot). Increment the overflow count in | |
1466 | * the ring. | |
1467 | */ | |
1468 | cqe = io_get_cqring(ctx); | |
1d7bb1d5 | 1469 | if (likely(cqe)) { |
78e19bbe | 1470 | WRITE_ONCE(cqe->user_data, req->user_data); |
2b188cc1 | 1471 | WRITE_ONCE(cqe->res, res); |
bcda7baa | 1472 | WRITE_ONCE(cqe->flags, cflags); |
0f212204 JA |
1473 | } else if (ctx->cq_overflow_flushed || req->task->io_uring->in_idle) { |
1474 | /* | |
1475 | * If we're in ring overflow flush mode, or in task cancel mode, | |
1476 | * then we cannot store the request for later flushing, we need | |
1477 | * to drop it on the floor. | |
1478 | */ | |
498ccd9e JA |
1479 | WRITE_ONCE(ctx->rings->cq_overflow, |
1480 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
1d7bb1d5 | 1481 | } else { |
ad3eb2c8 JA |
1482 | if (list_empty(&ctx->cq_overflow_list)) { |
1483 | set_bit(0, &ctx->sq_check_overflow); | |
1484 | set_bit(0, &ctx->cq_check_overflow); | |
6d5f9049 | 1485 | ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW; |
ad3eb2c8 | 1486 | } |
40d8ddd4 | 1487 | io_clean_op(req); |
1d7bb1d5 | 1488 | req->result = res; |
0f7e466b | 1489 | req->compl.cflags = cflags; |
40d8ddd4 PB |
1490 | refcount_inc(&req->refs); |
1491 | list_add_tail(&req->compl.list, &ctx->cq_overflow_list); | |
2b188cc1 JA |
1492 | } |
1493 | } | |
1494 | ||
bcda7baa JA |
1495 | static void io_cqring_fill_event(struct io_kiocb *req, long res) |
1496 | { | |
1497 | __io_cqring_fill_event(req, res, 0); | |
1498 | } | |
1499 | ||
e1e16097 | 1500 | static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags) |
2b188cc1 | 1501 | { |
78e19bbe | 1502 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
1503 | unsigned long flags; |
1504 | ||
1505 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
bcda7baa | 1506 | __io_cqring_fill_event(req, res, cflags); |
2b188cc1 JA |
1507 | io_commit_cqring(ctx); |
1508 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1509 | ||
8c838788 | 1510 | io_cqring_ev_posted(ctx); |
2b188cc1 JA |
1511 | } |
1512 | ||
229a7b63 | 1513 | static void io_submit_flush_completions(struct io_comp_state *cs) |
bcda7baa | 1514 | { |
229a7b63 JA |
1515 | struct io_ring_ctx *ctx = cs->ctx; |
1516 | ||
1517 | spin_lock_irq(&ctx->completion_lock); | |
1518 | while (!list_empty(&cs->list)) { | |
1519 | struct io_kiocb *req; | |
1520 | ||
3ca405eb PB |
1521 | req = list_first_entry(&cs->list, struct io_kiocb, compl.list); |
1522 | list_del(&req->compl.list); | |
0f7e466b | 1523 | __io_cqring_fill_event(req, req->result, req->compl.cflags); |
229a7b63 JA |
1524 | if (!(req->flags & REQ_F_LINK_HEAD)) { |
1525 | req->flags |= REQ_F_COMP_LOCKED; | |
1526 | io_put_req(req); | |
1527 | } else { | |
1528 | spin_unlock_irq(&ctx->completion_lock); | |
1529 | io_put_req(req); | |
1530 | spin_lock_irq(&ctx->completion_lock); | |
1531 | } | |
1532 | } | |
1533 | io_commit_cqring(ctx); | |
1534 | spin_unlock_irq(&ctx->completion_lock); | |
1535 | ||
1536 | io_cqring_ev_posted(ctx); | |
1537 | cs->nr = 0; | |
1538 | } | |
1539 | ||
1540 | static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags, | |
1541 | struct io_comp_state *cs) | |
1542 | { | |
1543 | if (!cs) { | |
1544 | io_cqring_add_event(req, res, cflags); | |
1545 | io_put_req(req); | |
1546 | } else { | |
3ca405eb | 1547 | io_clean_op(req); |
229a7b63 | 1548 | req->result = res; |
0f7e466b | 1549 | req->compl.cflags = cflags; |
3ca405eb | 1550 | list_add_tail(&req->compl.list, &cs->list); |
229a7b63 JA |
1551 | if (++cs->nr >= 32) |
1552 | io_submit_flush_completions(cs); | |
1553 | } | |
e1e16097 JA |
1554 | } |
1555 | ||
1556 | static void io_req_complete(struct io_kiocb *req, long res) | |
bcda7baa | 1557 | { |
229a7b63 | 1558 | __io_req_complete(req, res, 0, NULL); |
bcda7baa JA |
1559 | } |
1560 | ||
0ddf92e8 JA |
1561 | static inline bool io_is_fallback_req(struct io_kiocb *req) |
1562 | { | |
1563 | return req == (struct io_kiocb *) | |
1564 | ((unsigned long) req->ctx->fallback_req & ~1UL); | |
1565 | } | |
1566 | ||
1567 | static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx) | |
1568 | { | |
1569 | struct io_kiocb *req; | |
1570 | ||
1571 | req = ctx->fallback_req; | |
dd461af6 | 1572 | if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req)) |
0ddf92e8 JA |
1573 | return req; |
1574 | ||
1575 | return NULL; | |
1576 | } | |
1577 | ||
0553b8bd PB |
1578 | static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx, |
1579 | struct io_submit_state *state) | |
2b188cc1 | 1580 | { |
fd6fab2c | 1581 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
2b188cc1 JA |
1582 | struct io_kiocb *req; |
1583 | ||
f6b6c7d6 | 1584 | if (!state->free_reqs) { |
2579f913 JA |
1585 | size_t sz; |
1586 | int ret; | |
1587 | ||
1588 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
fd6fab2c JA |
1589 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs); |
1590 | ||
1591 | /* | |
1592 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
1593 | * retry single alloc to be on the safe side. | |
1594 | */ | |
1595 | if (unlikely(ret <= 0)) { | |
1596 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
1597 | if (!state->reqs[0]) | |
0ddf92e8 | 1598 | goto fallback; |
fd6fab2c JA |
1599 | ret = 1; |
1600 | } | |
2579f913 | 1601 | state->free_reqs = ret - 1; |
6c8a3134 | 1602 | req = state->reqs[ret - 1]; |
2579f913 | 1603 | } else { |
2579f913 | 1604 | state->free_reqs--; |
6c8a3134 | 1605 | req = state->reqs[state->free_reqs]; |
2b188cc1 JA |
1606 | } |
1607 | ||
2579f913 | 1608 | return req; |
0ddf92e8 | 1609 | fallback: |
0553b8bd | 1610 | return io_get_fallback_req(ctx); |
2b188cc1 JA |
1611 | } |
1612 | ||
8da11c19 PB |
1613 | static inline void io_put_file(struct io_kiocb *req, struct file *file, |
1614 | bool fixed) | |
1615 | { | |
1616 | if (fixed) | |
05589553 | 1617 | percpu_ref_put(req->fixed_file_refs); |
8da11c19 PB |
1618 | else |
1619 | fput(file); | |
1620 | } | |
1621 | ||
51a4cc11 | 1622 | static bool io_dismantle_req(struct io_kiocb *req) |
2b188cc1 | 1623 | { |
3ca405eb | 1624 | io_clean_op(req); |
929a3af9 | 1625 | |
5acbbc8e JA |
1626 | if (req->io) |
1627 | kfree(req->io); | |
8da11c19 PB |
1628 | if (req->file) |
1629 | io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE)); | |
fcb323cc | 1630 | |
51a4cc11 | 1631 | return io_req_clean_work(req); |
e65ef56d JA |
1632 | } |
1633 | ||
51a4cc11 | 1634 | static void __io_free_req_finish(struct io_kiocb *req) |
c6ca97b3 | 1635 | { |
0f212204 | 1636 | struct io_uring_task *tctx = req->task->io_uring; |
51a4cc11 | 1637 | struct io_ring_ctx *ctx = req->ctx; |
c6ca97b3 | 1638 | |
0f212204 JA |
1639 | atomic_long_inc(&tctx->req_complete); |
1640 | if (tctx->in_idle) | |
1641 | wake_up(&tctx->wait); | |
e3bc8e9d JA |
1642 | put_task_struct(req->task); |
1643 | ||
b1e50e54 PB |
1644 | if (likely(!io_is_fallback_req(req))) |
1645 | kmem_cache_free(req_cachep, req); | |
1646 | else | |
ecfc5177 PB |
1647 | clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req); |
1648 | percpu_ref_put(&ctx->refs); | |
e65ef56d JA |
1649 | } |
1650 | ||
51a4cc11 JA |
1651 | static void io_req_task_file_table_put(struct callback_head *cb) |
1652 | { | |
1653 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); | |
1654 | struct fs_struct *fs = req->work.fs; | |
1655 | ||
1656 | spin_lock(&req->work.fs->lock); | |
1657 | if (--fs->users) | |
1658 | fs = NULL; | |
1659 | spin_unlock(&req->work.fs->lock); | |
1660 | if (fs) | |
1661 | free_fs_struct(fs); | |
1662 | req->work.fs = NULL; | |
1663 | __io_free_req_finish(req); | |
1664 | } | |
1665 | ||
1666 | static void __io_free_req(struct io_kiocb *req) | |
1667 | { | |
1668 | if (!io_dismantle_req(req)) { | |
1669 | __io_free_req_finish(req); | |
1670 | } else { | |
1671 | int ret; | |
1672 | ||
1673 | init_task_work(&req->task_work, io_req_task_file_table_put); | |
1674 | ret = task_work_add(req->task, &req->task_work, TWA_RESUME); | |
1675 | if (unlikely(ret)) { | |
1676 | struct task_struct *tsk; | |
1677 | ||
1678 | tsk = io_wq_get_task(req->ctx->io_wq); | |
1679 | task_work_add(tsk, &req->task_work, 0); | |
1680 | } | |
1681 | } | |
1682 | } | |
1683 | ||
a197f664 | 1684 | static bool io_link_cancel_timeout(struct io_kiocb *req) |
2665abfd | 1685 | { |
a197f664 | 1686 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd JA |
1687 | int ret; |
1688 | ||
2d28390a | 1689 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
2665abfd | 1690 | if (ret != -1) { |
78e19bbe | 1691 | io_cqring_fill_event(req, -ECANCELED); |
2665abfd | 1692 | io_commit_cqring(ctx); |
dea3b49c | 1693 | req->flags &= ~REQ_F_LINK_HEAD; |
ec9c02ad | 1694 | io_put_req(req); |
2665abfd JA |
1695 | return true; |
1696 | } | |
1697 | ||
1698 | return false; | |
e65ef56d JA |
1699 | } |
1700 | ||
ab0b6451 | 1701 | static bool __io_kill_linked_timeout(struct io_kiocb *req) |
9e645e11 | 1702 | { |
7c86ffee | 1703 | struct io_kiocb *link; |
ab0b6451 | 1704 | bool wake_ev; |
7c86ffee PB |
1705 | |
1706 | if (list_empty(&req->link_list)) | |
ab0b6451 | 1707 | return false; |
7c86ffee PB |
1708 | link = list_first_entry(&req->link_list, struct io_kiocb, link_list); |
1709 | if (link->opcode != IORING_OP_LINK_TIMEOUT) | |
ab0b6451 | 1710 | return false; |
7c86ffee PB |
1711 | |
1712 | list_del_init(&link->link_list); | |
9b7adba9 | 1713 | link->flags |= REQ_F_COMP_LOCKED; |
7c86ffee PB |
1714 | wake_ev = io_link_cancel_timeout(link); |
1715 | req->flags &= ~REQ_F_LINK_TIMEOUT; | |
ab0b6451 JA |
1716 | return wake_ev; |
1717 | } | |
1718 | ||
1719 | static void io_kill_linked_timeout(struct io_kiocb *req) | |
9e645e11 | 1720 | { |
2665abfd | 1721 | struct io_ring_ctx *ctx = req->ctx; |
ab0b6451 | 1722 | bool wake_ev; |
9e645e11 | 1723 | |
ab0b6451 JA |
1724 | if (!(req->flags & REQ_F_COMP_LOCKED)) { |
1725 | unsigned long flags; | |
1726 | ||
1727 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
1728 | wake_ev = __io_kill_linked_timeout(req); | |
7c86ffee | 1729 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
ab0b6451 JA |
1730 | } else { |
1731 | wake_ev = __io_kill_linked_timeout(req); | |
1732 | } | |
1733 | ||
7c86ffee PB |
1734 | if (wake_ev) |
1735 | io_cqring_ev_posted(ctx); | |
1736 | } | |
1737 | ||
9b5f7bd9 | 1738 | static struct io_kiocb *io_req_link_next(struct io_kiocb *req) |
7c86ffee PB |
1739 | { |
1740 | struct io_kiocb *nxt; | |
4d7dd462 | 1741 | |
9e645e11 JA |
1742 | /* |
1743 | * The list should never be empty when we are called here. But could | |
1744 | * potentially happen if the chain is messed up, check to be on the | |
1745 | * safe side. | |
1746 | */ | |
7c86ffee | 1747 | if (unlikely(list_empty(&req->link_list))) |
9b5f7bd9 | 1748 | return NULL; |
2665abfd | 1749 | |
7c86ffee PB |
1750 | nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list); |
1751 | list_del_init(&req->link_list); | |
1752 | if (!list_empty(&nxt->link_list)) | |
1753 | nxt->flags |= REQ_F_LINK_HEAD; | |
9b5f7bd9 | 1754 | return nxt; |
9e645e11 JA |
1755 | } |
1756 | ||
1757 | /* | |
dea3b49c | 1758 | * Called if REQ_F_LINK_HEAD is set, and we fail the head request |
9e645e11 | 1759 | */ |
7c86ffee | 1760 | static void __io_fail_links(struct io_kiocb *req) |
9e645e11 | 1761 | { |
2665abfd | 1762 | struct io_ring_ctx *ctx = req->ctx; |
9e645e11 JA |
1763 | |
1764 | while (!list_empty(&req->link_list)) { | |
4493233e PB |
1765 | struct io_kiocb *link = list_first_entry(&req->link_list, |
1766 | struct io_kiocb, link_list); | |
9e645e11 | 1767 | |
4493233e | 1768 | list_del_init(&link->link_list); |
c826bd7a | 1769 | trace_io_uring_fail_link(req, link); |
2665abfd | 1770 | |
7c86ffee | 1771 | io_cqring_fill_event(link, -ECANCELED); |
9b7adba9 | 1772 | link->flags |= REQ_F_COMP_LOCKED; |
7c86ffee | 1773 | __io_double_put_req(link); |
5d960724 | 1774 | req->flags &= ~REQ_F_LINK_TIMEOUT; |
9e645e11 | 1775 | } |
2665abfd JA |
1776 | |
1777 | io_commit_cqring(ctx); | |
2665abfd | 1778 | io_cqring_ev_posted(ctx); |
9e645e11 JA |
1779 | } |
1780 | ||
7c86ffee | 1781 | static void io_fail_links(struct io_kiocb *req) |
9e645e11 | 1782 | { |
7c86ffee | 1783 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd | 1784 | |
7c86ffee | 1785 | if (!(req->flags & REQ_F_COMP_LOCKED)) { |
2665abfd JA |
1786 | unsigned long flags; |
1787 | ||
2665abfd | 1788 | spin_lock_irqsave(&ctx->completion_lock, flags); |
7c86ffee | 1789 | __io_fail_links(req); |
2665abfd JA |
1790 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
1791 | } else { | |
7c86ffee | 1792 | __io_fail_links(req); |
9e645e11 JA |
1793 | } |
1794 | ||
2665abfd | 1795 | io_cqring_ev_posted(ctx); |
9e645e11 JA |
1796 | } |
1797 | ||
3fa5e0f3 | 1798 | static struct io_kiocb *__io_req_find_next(struct io_kiocb *req) |
c69f8dbe | 1799 | { |
9b0d911a | 1800 | req->flags &= ~REQ_F_LINK_HEAD; |
7c86ffee PB |
1801 | if (req->flags & REQ_F_LINK_TIMEOUT) |
1802 | io_kill_linked_timeout(req); | |
944e58bf | 1803 | |
9e645e11 JA |
1804 | /* |
1805 | * If LINK is set, we have dependent requests in this chain. If we | |
1806 | * didn't fail this request, queue the first one up, moving any other | |
1807 | * dependencies to the next request. In case of failure, fail the rest | |
1808 | * of the chain. | |
1809 | */ | |
9b5f7bd9 PB |
1810 | if (likely(!(req->flags & REQ_F_FAIL_LINK))) |
1811 | return io_req_link_next(req); | |
1812 | io_fail_links(req); | |
1813 | return NULL; | |
4d7dd462 | 1814 | } |
9e645e11 | 1815 | |
3fa5e0f3 PB |
1816 | static struct io_kiocb *io_req_find_next(struct io_kiocb *req) |
1817 | { | |
1818 | if (likely(!(req->flags & REQ_F_LINK_HEAD))) | |
1819 | return NULL; | |
1820 | return __io_req_find_next(req); | |
1821 | } | |
1822 | ||
fd7d6de2 JA |
1823 | static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb, |
1824 | bool twa_signal_ok) | |
c2c4c83c JA |
1825 | { |
1826 | struct task_struct *tsk = req->task; | |
1827 | struct io_ring_ctx *ctx = req->ctx; | |
0ba9c9ed | 1828 | int ret, notify; |
c2c4c83c | 1829 | |
6200b0ae JA |
1830 | if (tsk->flags & PF_EXITING) |
1831 | return -ESRCH; | |
1832 | ||
c2c4c83c | 1833 | /* |
0ba9c9ed JA |
1834 | * SQPOLL kernel thread doesn't need notification, just a wakeup. For |
1835 | * all other cases, use TWA_SIGNAL unconditionally to ensure we're | |
1836 | * processing task_work. There's no reliable way to tell if TWA_RESUME | |
1837 | * will do the job. | |
c2c4c83c | 1838 | */ |
0ba9c9ed | 1839 | notify = 0; |
fd7d6de2 | 1840 | if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok) |
c2c4c83c JA |
1841 | notify = TWA_SIGNAL; |
1842 | ||
1843 | ret = task_work_add(tsk, cb, notify); | |
1844 | if (!ret) | |
1845 | wake_up_process(tsk); | |
0ba9c9ed | 1846 | |
c2c4c83c JA |
1847 | return ret; |
1848 | } | |
1849 | ||
c40f6379 JA |
1850 | static void __io_req_task_cancel(struct io_kiocb *req, int error) |
1851 | { | |
1852 | struct io_ring_ctx *ctx = req->ctx; | |
1853 | ||
1854 | spin_lock_irq(&ctx->completion_lock); | |
1855 | io_cqring_fill_event(req, error); | |
1856 | io_commit_cqring(ctx); | |
1857 | spin_unlock_irq(&ctx->completion_lock); | |
1858 | ||
1859 | io_cqring_ev_posted(ctx); | |
1860 | req_set_fail_links(req); | |
1861 | io_double_put_req(req); | |
1862 | } | |
1863 | ||
1864 | static void io_req_task_cancel(struct callback_head *cb) | |
1865 | { | |
1866 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); | |
87ceb6a6 | 1867 | struct io_ring_ctx *ctx = req->ctx; |
c40f6379 JA |
1868 | |
1869 | __io_req_task_cancel(req, -ECANCELED); | |
87ceb6a6 | 1870 | percpu_ref_put(&ctx->refs); |
c40f6379 JA |
1871 | } |
1872 | ||
1873 | static void __io_req_task_submit(struct io_kiocb *req) | |
1874 | { | |
1875 | struct io_ring_ctx *ctx = req->ctx; | |
1876 | ||
c40f6379 JA |
1877 | if (!__io_sq_thread_acquire_mm(ctx)) { |
1878 | mutex_lock(&ctx->uring_lock); | |
1879 | __io_queue_sqe(req, NULL, NULL); | |
1880 | mutex_unlock(&ctx->uring_lock); | |
1881 | } else { | |
1882 | __io_req_task_cancel(req, -EFAULT); | |
1883 | } | |
1884 | } | |
1885 | ||
1886 | static void io_req_task_submit(struct callback_head *cb) | |
1887 | { | |
1888 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); | |
6d816e08 | 1889 | struct io_ring_ctx *ctx = req->ctx; |
c40f6379 JA |
1890 | |
1891 | __io_req_task_submit(req); | |
6d816e08 | 1892 | percpu_ref_put(&ctx->refs); |
c40f6379 JA |
1893 | } |
1894 | ||
1895 | static void io_req_task_queue(struct io_kiocb *req) | |
1896 | { | |
c40f6379 JA |
1897 | int ret; |
1898 | ||
1899 | init_task_work(&req->task_work, io_req_task_submit); | |
6d816e08 | 1900 | percpu_ref_get(&req->ctx->refs); |
c40f6379 | 1901 | |
fd7d6de2 | 1902 | ret = io_req_task_work_add(req, &req->task_work, true); |
c40f6379 | 1903 | if (unlikely(ret)) { |
c2c4c83c JA |
1904 | struct task_struct *tsk; |
1905 | ||
c40f6379 JA |
1906 | init_task_work(&req->task_work, io_req_task_cancel); |
1907 | tsk = io_wq_get_task(req->ctx->io_wq); | |
c2c4c83c JA |
1908 | task_work_add(tsk, &req->task_work, 0); |
1909 | wake_up_process(tsk); | |
c40f6379 | 1910 | } |
c40f6379 JA |
1911 | } |
1912 | ||
c3524383 | 1913 | static void io_queue_next(struct io_kiocb *req) |
c69f8dbe | 1914 | { |
9b5f7bd9 | 1915 | struct io_kiocb *nxt = io_req_find_next(req); |
944e58bf PB |
1916 | |
1917 | if (nxt) | |
906a8c3f | 1918 | io_req_task_queue(nxt); |
c69f8dbe JL |
1919 | } |
1920 | ||
c3524383 | 1921 | static void io_free_req(struct io_kiocb *req) |
7a743e22 | 1922 | { |
c3524383 PB |
1923 | io_queue_next(req); |
1924 | __io_free_req(req); | |
1925 | } | |
8766dd51 | 1926 | |
2d6500d4 PB |
1927 | struct req_batch { |
1928 | void *reqs[IO_IOPOLL_BATCH]; | |
1929 | int to_free; | |
7a743e22 | 1930 | |
5af1d13e PB |
1931 | struct task_struct *task; |
1932 | int task_refs; | |
2d6500d4 PB |
1933 | }; |
1934 | ||
5af1d13e PB |
1935 | static inline void io_init_req_batch(struct req_batch *rb) |
1936 | { | |
1937 | rb->to_free = 0; | |
1938 | rb->task_refs = 0; | |
1939 | rb->task = NULL; | |
1940 | } | |
1941 | ||
2d6500d4 PB |
1942 | static void __io_req_free_batch_flush(struct io_ring_ctx *ctx, |
1943 | struct req_batch *rb) | |
1944 | { | |
1945 | kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs); | |
1946 | percpu_ref_put_many(&ctx->refs, rb->to_free); | |
1947 | rb->to_free = 0; | |
1948 | } | |
1949 | ||
1950 | static void io_req_free_batch_finish(struct io_ring_ctx *ctx, | |
1951 | struct req_batch *rb) | |
1952 | { | |
1953 | if (rb->to_free) | |
1954 | __io_req_free_batch_flush(ctx, rb); | |
5af1d13e | 1955 | if (rb->task) { |
0f212204 | 1956 | atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete); |
5af1d13e PB |
1957 | put_task_struct_many(rb->task, rb->task_refs); |
1958 | rb->task = NULL; | |
1959 | } | |
2d6500d4 PB |
1960 | } |
1961 | ||
1962 | static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req) | |
1963 | { | |
1964 | if (unlikely(io_is_fallback_req(req))) { | |
1965 | io_free_req(req); | |
1966 | return; | |
1967 | } | |
1968 | if (req->flags & REQ_F_LINK_HEAD) | |
1969 | io_queue_next(req); | |
1970 | ||
e3bc8e9d | 1971 | if (req->task != rb->task) { |
0f212204 JA |
1972 | if (rb->task) { |
1973 | atomic_long_add(rb->task_refs, &rb->task->io_uring->req_complete); | |
e3bc8e9d | 1974 | put_task_struct_many(rb->task, rb->task_refs); |
0f212204 | 1975 | } |
e3bc8e9d JA |
1976 | rb->task = req->task; |
1977 | rb->task_refs = 0; | |
5af1d13e | 1978 | } |
e3bc8e9d | 1979 | rb->task_refs++; |
5af1d13e | 1980 | |
51a4cc11 | 1981 | WARN_ON_ONCE(io_dismantle_req(req)); |
2d6500d4 PB |
1982 | rb->reqs[rb->to_free++] = req; |
1983 | if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs))) | |
1984 | __io_req_free_batch_flush(req->ctx, rb); | |
7a743e22 PB |
1985 | } |
1986 | ||
ba816ad6 JA |
1987 | /* |
1988 | * Drop reference to request, return next in chain (if there is one) if this | |
1989 | * was the last reference to this request. | |
1990 | */ | |
9b5f7bd9 | 1991 | static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req) |
e65ef56d | 1992 | { |
9b5f7bd9 PB |
1993 | struct io_kiocb *nxt = NULL; |
1994 | ||
2a44f467 | 1995 | if (refcount_dec_and_test(&req->refs)) { |
9b5f7bd9 | 1996 | nxt = io_req_find_next(req); |
4d7dd462 | 1997 | __io_free_req(req); |
2a44f467 | 1998 | } |
9b5f7bd9 | 1999 | return nxt; |
2b188cc1 JA |
2000 | } |
2001 | ||
e65ef56d JA |
2002 | static void io_put_req(struct io_kiocb *req) |
2003 | { | |
2004 | if (refcount_dec_and_test(&req->refs)) | |
2005 | io_free_req(req); | |
2b188cc1 JA |
2006 | } |
2007 | ||
f4db7182 | 2008 | static struct io_wq_work *io_steal_work(struct io_kiocb *req) |
7a743e22 | 2009 | { |
6df1db6b | 2010 | struct io_kiocb *nxt; |
f4db7182 | 2011 | |
7a743e22 | 2012 | /* |
f4db7182 PB |
2013 | * A ref is owned by io-wq in which context we're. So, if that's the |
2014 | * last one, it's safe to steal next work. False negatives are Ok, | |
2015 | * it just will be re-punted async in io_put_work() | |
7a743e22 | 2016 | */ |
f4db7182 PB |
2017 | if (refcount_read(&req->refs) != 1) |
2018 | return NULL; | |
7a743e22 | 2019 | |
9b5f7bd9 | 2020 | nxt = io_req_find_next(req); |
6df1db6b | 2021 | return nxt ? &nxt->work : NULL; |
7a743e22 PB |
2022 | } |
2023 | ||
978db57e JA |
2024 | /* |
2025 | * Must only be used if we don't need to care about links, usually from | |
2026 | * within the completion handling itself. | |
2027 | */ | |
2028 | static void __io_double_put_req(struct io_kiocb *req) | |
78e19bbe JA |
2029 | { |
2030 | /* drop both submit and complete references */ | |
2031 | if (refcount_sub_and_test(2, &req->refs)) | |
2032 | __io_free_req(req); | |
2033 | } | |
2034 | ||
978db57e JA |
2035 | static void io_double_put_req(struct io_kiocb *req) |
2036 | { | |
2037 | /* drop both submit and complete references */ | |
2038 | if (refcount_sub_and_test(2, &req->refs)) | |
2039 | io_free_req(req); | |
2040 | } | |
2041 | ||
1d7bb1d5 | 2042 | static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush) |
a3a0e43f | 2043 | { |
84f97dc2 JA |
2044 | struct io_rings *rings = ctx->rings; |
2045 | ||
ad3eb2c8 JA |
2046 | if (test_bit(0, &ctx->cq_check_overflow)) { |
2047 | /* | |
2048 | * noflush == true is from the waitqueue handler, just ensure | |
2049 | * we wake up the task, and the next invocation will flush the | |
2050 | * entries. We cannot safely to it from here. | |
2051 | */ | |
2052 | if (noflush && !list_empty(&ctx->cq_overflow_list)) | |
2053 | return -1U; | |
1d7bb1d5 | 2054 | |
e6c8aa9a | 2055 | io_cqring_overflow_flush(ctx, false, NULL, NULL); |
ad3eb2c8 | 2056 | } |
1d7bb1d5 | 2057 | |
a3a0e43f JA |
2058 | /* See comment at the top of this file */ |
2059 | smp_rmb(); | |
ad3eb2c8 | 2060 | return ctx->cached_cq_tail - READ_ONCE(rings->cq.head); |
a3a0e43f JA |
2061 | } |
2062 | ||
fb5ccc98 PB |
2063 | static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) |
2064 | { | |
2065 | struct io_rings *rings = ctx->rings; | |
2066 | ||
2067 | /* make sure SQ entry isn't read before tail */ | |
2068 | return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; | |
2069 | } | |
2070 | ||
8ff069bf | 2071 | static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf) |
e94f141b | 2072 | { |
8ff069bf | 2073 | unsigned int cflags; |
e94f141b | 2074 | |
bcda7baa JA |
2075 | cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT; |
2076 | cflags |= IORING_CQE_F_BUFFER; | |
0e1b6fe3 | 2077 | req->flags &= ~REQ_F_BUFFER_SELECTED; |
bcda7baa JA |
2078 | kfree(kbuf); |
2079 | return cflags; | |
e94f141b JA |
2080 | } |
2081 | ||
8ff069bf | 2082 | static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req) |
bcda7baa | 2083 | { |
4d954c25 | 2084 | struct io_buffer *kbuf; |
bcda7baa | 2085 | |
4d954c25 | 2086 | kbuf = (struct io_buffer *) (unsigned long) req->rw.addr; |
8ff069bf PB |
2087 | return io_put_kbuf(req, kbuf); |
2088 | } | |
2089 | ||
4c6e277c JA |
2090 | static inline bool io_run_task_work(void) |
2091 | { | |
6200b0ae JA |
2092 | /* |
2093 | * Not safe to run on exiting task, and the task_work handling will | |
2094 | * not add work to such a task. | |
2095 | */ | |
2096 | if (unlikely(current->flags & PF_EXITING)) | |
2097 | return false; | |
4c6e277c JA |
2098 | if (current->task_works) { |
2099 | __set_current_state(TASK_RUNNING); | |
2100 | task_work_run(); | |
2101 | return true; | |
2102 | } | |
2103 | ||
2104 | return false; | |
bcda7baa JA |
2105 | } |
2106 | ||
bbde017a XW |
2107 | static void io_iopoll_queue(struct list_head *again) |
2108 | { | |
2109 | struct io_kiocb *req; | |
2110 | ||
2111 | do { | |
d21ffe7e PB |
2112 | req = list_first_entry(again, struct io_kiocb, inflight_entry); |
2113 | list_del(&req->inflight_entry); | |
81b68a5c | 2114 | __io_complete_rw(req, -EAGAIN, 0, NULL); |
bbde017a XW |
2115 | } while (!list_empty(again)); |
2116 | } | |
2117 | ||
def596e9 JA |
2118 | /* |
2119 | * Find and free completed poll iocbs | |
2120 | */ | |
2121 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
2122 | struct list_head *done) | |
2123 | { | |
8237e045 | 2124 | struct req_batch rb; |
def596e9 | 2125 | struct io_kiocb *req; |
bbde017a XW |
2126 | LIST_HEAD(again); |
2127 | ||
2128 | /* order with ->result store in io_complete_rw_iopoll() */ | |
2129 | smp_rmb(); | |
def596e9 | 2130 | |
5af1d13e | 2131 | io_init_req_batch(&rb); |
def596e9 | 2132 | while (!list_empty(done)) { |
bcda7baa JA |
2133 | int cflags = 0; |
2134 | ||
d21ffe7e | 2135 | req = list_first_entry(done, struct io_kiocb, inflight_entry); |
bbde017a | 2136 | if (READ_ONCE(req->result) == -EAGAIN) { |
56450c20 | 2137 | req->result = 0; |
bbde017a | 2138 | req->iopoll_completed = 0; |
d21ffe7e | 2139 | list_move_tail(&req->inflight_entry, &again); |
bbde017a XW |
2140 | continue; |
2141 | } | |
d21ffe7e | 2142 | list_del(&req->inflight_entry); |
def596e9 | 2143 | |
bcda7baa | 2144 | if (req->flags & REQ_F_BUFFER_SELECTED) |
8ff069bf | 2145 | cflags = io_put_rw_kbuf(req); |
bcda7baa JA |
2146 | |
2147 | __io_cqring_fill_event(req, req->result, cflags); | |
def596e9 JA |
2148 | (*nr_events)++; |
2149 | ||
c3524383 | 2150 | if (refcount_dec_and_test(&req->refs)) |
2d6500d4 | 2151 | io_req_free_batch(&rb, req); |
def596e9 | 2152 | } |
def596e9 | 2153 | |
09bb8394 | 2154 | io_commit_cqring(ctx); |
32b2244a XW |
2155 | if (ctx->flags & IORING_SETUP_SQPOLL) |
2156 | io_cqring_ev_posted(ctx); | |
2d6500d4 | 2157 | io_req_free_batch_finish(ctx, &rb); |
581f9810 | 2158 | |
bbde017a XW |
2159 | if (!list_empty(&again)) |
2160 | io_iopoll_queue(&again); | |
581f9810 BM |
2161 | } |
2162 | ||
def596e9 JA |
2163 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, |
2164 | long min) | |
2165 | { | |
2166 | struct io_kiocb *req, *tmp; | |
2167 | LIST_HEAD(done); | |
2168 | bool spin; | |
2169 | int ret; | |
2170 | ||
2171 | /* | |
2172 | * Only spin for completions if we don't have multiple devices hanging | |
2173 | * off our complete list, and we're under the requested amount. | |
2174 | */ | |
2175 | spin = !ctx->poll_multi_file && *nr_events < min; | |
2176 | ||
2177 | ret = 0; | |
d21ffe7e | 2178 | list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) { |
9adbd45d | 2179 | struct kiocb *kiocb = &req->rw.kiocb; |
def596e9 JA |
2180 | |
2181 | /* | |
581f9810 BM |
2182 | * Move completed and retryable entries to our local lists. |
2183 | * If we find a request that requires polling, break out | |
2184 | * and complete those lists first, if we have entries there. | |
def596e9 | 2185 | */ |
65a6543d | 2186 | if (READ_ONCE(req->iopoll_completed)) { |
d21ffe7e | 2187 | list_move_tail(&req->inflight_entry, &done); |
def596e9 JA |
2188 | continue; |
2189 | } | |
2190 | if (!list_empty(&done)) | |
2191 | break; | |
2192 | ||
2193 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
2194 | if (ret < 0) | |
2195 | break; | |
2196 | ||
3aadc23e PB |
2197 | /* iopoll may have completed current req */ |
2198 | if (READ_ONCE(req->iopoll_completed)) | |
d21ffe7e | 2199 | list_move_tail(&req->inflight_entry, &done); |
3aadc23e | 2200 | |
def596e9 JA |
2201 | if (ret && spin) |
2202 | spin = false; | |
2203 | ret = 0; | |
2204 | } | |
2205 | ||
2206 | if (!list_empty(&done)) | |
2207 | io_iopoll_complete(ctx, nr_events, &done); | |
2208 | ||
2209 | return ret; | |
2210 | } | |
2211 | ||
2212 | /* | |
d195a66e | 2213 | * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a |
def596e9 JA |
2214 | * non-spinning poll check - we'll still enter the driver poll loop, but only |
2215 | * as a non-spinning completion check. | |
2216 | */ | |
2217 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
2218 | long min) | |
2219 | { | |
540e32a0 | 2220 | while (!list_empty(&ctx->iopoll_list) && !need_resched()) { |
def596e9 JA |
2221 | int ret; |
2222 | ||
2223 | ret = io_do_iopoll(ctx, nr_events, min); | |
2224 | if (ret < 0) | |
2225 | return ret; | |
eba0a4dd | 2226 | if (*nr_events >= min) |
def596e9 JA |
2227 | return 0; |
2228 | } | |
2229 | ||
2230 | return 1; | |
2231 | } | |
2232 | ||
2233 | /* | |
2234 | * We can't just wait for polled events to come to us, we have to actively | |
2235 | * find and complete them. | |
2236 | */ | |
b2edc0a7 | 2237 | static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) |
def596e9 JA |
2238 | { |
2239 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
2240 | return; | |
2241 | ||
2242 | mutex_lock(&ctx->uring_lock); | |
540e32a0 | 2243 | while (!list_empty(&ctx->iopoll_list)) { |
def596e9 JA |
2244 | unsigned int nr_events = 0; |
2245 | ||
b2edc0a7 | 2246 | io_do_iopoll(ctx, &nr_events, 0); |
08f5439f | 2247 | |
b2edc0a7 PB |
2248 | /* let it sleep and repeat later if can't complete a request */ |
2249 | if (nr_events == 0) | |
2250 | break; | |
08f5439f JA |
2251 | /* |
2252 | * Ensure we allow local-to-the-cpu processing to take place, | |
2253 | * in this case we need to ensure that we reap all events. | |
3fcee5a6 | 2254 | * Also let task_work, etc. to progress by releasing the mutex |
08f5439f | 2255 | */ |
3fcee5a6 PB |
2256 | if (need_resched()) { |
2257 | mutex_unlock(&ctx->uring_lock); | |
2258 | cond_resched(); | |
2259 | mutex_lock(&ctx->uring_lock); | |
2260 | } | |
def596e9 JA |
2261 | } |
2262 | mutex_unlock(&ctx->uring_lock); | |
2263 | } | |
2264 | ||
7668b92a | 2265 | static int io_iopoll_check(struct io_ring_ctx *ctx, long min) |
def596e9 | 2266 | { |
7668b92a | 2267 | unsigned int nr_events = 0; |
2b2ed975 | 2268 | int iters = 0, ret = 0; |
500f9fba | 2269 | |
c7849be9 XW |
2270 | /* |
2271 | * We disallow the app entering submit/complete with polling, but we | |
2272 | * still need to lock the ring to prevent racing with polled issue | |
2273 | * that got punted to a workqueue. | |
2274 | */ | |
2275 | mutex_lock(&ctx->uring_lock); | |
def596e9 | 2276 | do { |
a3a0e43f JA |
2277 | /* |
2278 | * Don't enter poll loop if we already have events pending. | |
2279 | * If we do, we can potentially be spinning for commands that | |
2280 | * already triggered a CQE (eg in error). | |
2281 | */ | |
1d7bb1d5 | 2282 | if (io_cqring_events(ctx, false)) |
a3a0e43f JA |
2283 | break; |
2284 | ||
500f9fba JA |
2285 | /* |
2286 | * If a submit got punted to a workqueue, we can have the | |
2287 | * application entering polling for a command before it gets | |
2288 | * issued. That app will hold the uring_lock for the duration | |
2289 | * of the poll right here, so we need to take a breather every | |
2290 | * now and then to ensure that the issue has a chance to add | |
2291 | * the poll to the issued list. Otherwise we can spin here | |
2292 | * forever, while the workqueue is stuck trying to acquire the | |
2293 | * very same mutex. | |
2294 | */ | |
2295 | if (!(++iters & 7)) { | |
2296 | mutex_unlock(&ctx->uring_lock); | |
4c6e277c | 2297 | io_run_task_work(); |
500f9fba JA |
2298 | mutex_lock(&ctx->uring_lock); |
2299 | } | |
2300 | ||
7668b92a | 2301 | ret = io_iopoll_getevents(ctx, &nr_events, min); |
def596e9 JA |
2302 | if (ret <= 0) |
2303 | break; | |
2304 | ret = 0; | |
7668b92a | 2305 | } while (min && !nr_events && !need_resched()); |
def596e9 | 2306 | |
500f9fba | 2307 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
2308 | return ret; |
2309 | } | |
2310 | ||
491381ce | 2311 | static void kiocb_end_write(struct io_kiocb *req) |
2b188cc1 | 2312 | { |
491381ce JA |
2313 | /* |
2314 | * Tell lockdep we inherited freeze protection from submission | |
2315 | * thread. | |
2316 | */ | |
2317 | if (req->flags & REQ_F_ISREG) { | |
2318 | struct inode *inode = file_inode(req->file); | |
2b188cc1 | 2319 | |
491381ce | 2320 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); |
2b188cc1 | 2321 | } |
491381ce | 2322 | file_end_write(req->file); |
2b188cc1 JA |
2323 | } |
2324 | ||
a1d7c393 JA |
2325 | static void io_complete_rw_common(struct kiocb *kiocb, long res, |
2326 | struct io_comp_state *cs) | |
2b188cc1 | 2327 | { |
9adbd45d | 2328 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
bcda7baa | 2329 | int cflags = 0; |
2b188cc1 | 2330 | |
491381ce JA |
2331 | if (kiocb->ki_flags & IOCB_WRITE) |
2332 | kiocb_end_write(req); | |
2b188cc1 | 2333 | |
4e88d6e7 JA |
2334 | if (res != req->result) |
2335 | req_set_fail_links(req); | |
bcda7baa | 2336 | if (req->flags & REQ_F_BUFFER_SELECTED) |
8ff069bf | 2337 | cflags = io_put_rw_kbuf(req); |
a1d7c393 | 2338 | __io_req_complete(req, res, cflags, cs); |
ba816ad6 JA |
2339 | } |
2340 | ||
b63534c4 JA |
2341 | #ifdef CONFIG_BLOCK |
2342 | static bool io_resubmit_prep(struct io_kiocb *req, int error) | |
2343 | { | |
2344 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
2345 | ssize_t ret = -ECANCELED; | |
2346 | struct iov_iter iter; | |
2347 | int rw; | |
2348 | ||
2349 | if (error) { | |
2350 | ret = error; | |
2351 | goto end_req; | |
2352 | } | |
2353 | ||
2354 | switch (req->opcode) { | |
2355 | case IORING_OP_READV: | |
2356 | case IORING_OP_READ_FIXED: | |
2357 | case IORING_OP_READ: | |
2358 | rw = READ; | |
2359 | break; | |
2360 | case IORING_OP_WRITEV: | |
2361 | case IORING_OP_WRITE_FIXED: | |
2362 | case IORING_OP_WRITE: | |
2363 | rw = WRITE; | |
2364 | break; | |
2365 | default: | |
2366 | printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n", | |
2367 | req->opcode); | |
2368 | goto end_req; | |
2369 | } | |
2370 | ||
8f3d7496 JA |
2371 | if (!req->io) { |
2372 | ret = io_import_iovec(rw, req, &iovec, &iter, false); | |
2373 | if (ret < 0) | |
2374 | goto end_req; | |
2375 | ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false); | |
2376 | if (!ret) | |
2377 | return true; | |
2378 | kfree(iovec); | |
2379 | } else { | |
b63534c4 | 2380 | return true; |
8f3d7496 | 2381 | } |
b63534c4 | 2382 | end_req: |
b63534c4 | 2383 | req_set_fail_links(req); |
e1e16097 | 2384 | io_req_complete(req, ret); |
b63534c4 JA |
2385 | return false; |
2386 | } | |
b63534c4 JA |
2387 | #endif |
2388 | ||
2389 | static bool io_rw_reissue(struct io_kiocb *req, long res) | |
2390 | { | |
2391 | #ifdef CONFIG_BLOCK | |
355afaeb | 2392 | umode_t mode = file_inode(req->file)->i_mode; |
b63534c4 JA |
2393 | int ret; |
2394 | ||
355afaeb JA |
2395 | if (!S_ISBLK(mode) && !S_ISREG(mode)) |
2396 | return false; | |
b63534c4 JA |
2397 | if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker()) |
2398 | return false; | |
2399 | ||
fdee946d | 2400 | ret = io_sq_thread_acquire_mm(req->ctx, req); |
6d816e08 | 2401 | |
fdee946d JA |
2402 | if (io_resubmit_prep(req, ret)) { |
2403 | refcount_inc(&req->refs); | |
2404 | io_queue_async_work(req); | |
b63534c4 | 2405 | return true; |
fdee946d JA |
2406 | } |
2407 | ||
b63534c4 JA |
2408 | #endif |
2409 | return false; | |
2410 | } | |
2411 | ||
a1d7c393 JA |
2412 | static void __io_complete_rw(struct io_kiocb *req, long res, long res2, |
2413 | struct io_comp_state *cs) | |
2414 | { | |
2415 | if (!io_rw_reissue(req, res)) | |
2416 | io_complete_rw_common(&req->rw.kiocb, res, cs); | |
ba816ad6 JA |
2417 | } |
2418 | ||
2419 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
2420 | { | |
9adbd45d | 2421 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
ba816ad6 | 2422 | |
a1d7c393 | 2423 | __io_complete_rw(req, res, res2, NULL); |
2b188cc1 JA |
2424 | } |
2425 | ||
def596e9 JA |
2426 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
2427 | { | |
9adbd45d | 2428 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
def596e9 | 2429 | |
491381ce JA |
2430 | if (kiocb->ki_flags & IOCB_WRITE) |
2431 | kiocb_end_write(req); | |
def596e9 | 2432 | |
2d7d6792 | 2433 | if (res != -EAGAIN && res != req->result) |
4e88d6e7 | 2434 | req_set_fail_links(req); |
bbde017a XW |
2435 | |
2436 | WRITE_ONCE(req->result, res); | |
2437 | /* order with io_poll_complete() checking ->result */ | |
cd664b0e PB |
2438 | smp_wmb(); |
2439 | WRITE_ONCE(req->iopoll_completed, 1); | |
def596e9 JA |
2440 | } |
2441 | ||
2442 | /* | |
2443 | * After the iocb has been issued, it's safe to be found on the poll list. | |
2444 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
2445 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
2446 | * accessing the kiocb cookie. | |
2447 | */ | |
2448 | static void io_iopoll_req_issued(struct io_kiocb *req) | |
2449 | { | |
2450 | struct io_ring_ctx *ctx = req->ctx; | |
2451 | ||
2452 | /* | |
2453 | * Track whether we have multiple files in our lists. This will impact | |
2454 | * how we do polling eventually, not spinning if we're on potentially | |
2455 | * different devices. | |
2456 | */ | |
540e32a0 | 2457 | if (list_empty(&ctx->iopoll_list)) { |
def596e9 JA |
2458 | ctx->poll_multi_file = false; |
2459 | } else if (!ctx->poll_multi_file) { | |
2460 | struct io_kiocb *list_req; | |
2461 | ||
540e32a0 | 2462 | list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb, |
d21ffe7e | 2463 | inflight_entry); |
9adbd45d | 2464 | if (list_req->file != req->file) |
def596e9 JA |
2465 | ctx->poll_multi_file = true; |
2466 | } | |
2467 | ||
2468 | /* | |
2469 | * For fast devices, IO may have already completed. If it has, add | |
2470 | * it to the front so we find it first. | |
2471 | */ | |
65a6543d | 2472 | if (READ_ONCE(req->iopoll_completed)) |
d21ffe7e | 2473 | list_add(&req->inflight_entry, &ctx->iopoll_list); |
def596e9 | 2474 | else |
d21ffe7e | 2475 | list_add_tail(&req->inflight_entry, &ctx->iopoll_list); |
bdcd3eab | 2476 | |
534ca6d6 JA |
2477 | if ((ctx->flags & IORING_SETUP_SQPOLL) && |
2478 | wq_has_sleeper(&ctx->sq_data->wait)) | |
2479 | wake_up(&ctx->sq_data->wait); | |
def596e9 JA |
2480 | } |
2481 | ||
9f13c35b | 2482 | static void __io_state_file_put(struct io_submit_state *state) |
9a56a232 | 2483 | { |
06ef3608 PB |
2484 | if (state->has_refs) |
2485 | fput_many(state->file, state->has_refs); | |
9f13c35b PB |
2486 | state->file = NULL; |
2487 | } | |
2488 | ||
2489 | static inline void io_state_file_put(struct io_submit_state *state) | |
2490 | { | |
2491 | if (state->file) | |
2492 | __io_state_file_put(state); | |
9a56a232 JA |
2493 | } |
2494 | ||
2495 | /* | |
2496 | * Get as many references to a file as we have IOs left in this submission, | |
2497 | * assuming most submissions are for one file, or at least that each file | |
2498 | * has more than one submission. | |
2499 | */ | |
8da11c19 | 2500 | static struct file *__io_file_get(struct io_submit_state *state, int fd) |
9a56a232 JA |
2501 | { |
2502 | if (!state) | |
2503 | return fget(fd); | |
2504 | ||
2505 | if (state->file) { | |
2506 | if (state->fd == fd) { | |
06ef3608 | 2507 | state->has_refs--; |
9a56a232 JA |
2508 | state->ios_left--; |
2509 | return state->file; | |
2510 | } | |
9f13c35b | 2511 | __io_state_file_put(state); |
9a56a232 JA |
2512 | } |
2513 | state->file = fget_many(fd, state->ios_left); | |
2514 | if (!state->file) | |
2515 | return NULL; | |
2516 | ||
2517 | state->fd = fd; | |
9a56a232 | 2518 | state->ios_left--; |
06ef3608 | 2519 | state->has_refs = state->ios_left; |
9a56a232 JA |
2520 | return state->file; |
2521 | } | |
2522 | ||
4503b767 JA |
2523 | static bool io_bdev_nowait(struct block_device *bdev) |
2524 | { | |
2525 | #ifdef CONFIG_BLOCK | |
2526 | return !bdev || queue_is_mq(bdev_get_queue(bdev)); | |
2527 | #else | |
2528 | return true; | |
2529 | #endif | |
2530 | } | |
2531 | ||
2b188cc1 JA |
2532 | /* |
2533 | * If we tracked the file through the SCM inflight mechanism, we could support | |
2534 | * any file. For now, just ensure that anything potentially problematic is done | |
2535 | * inline. | |
2536 | */ | |
af197f50 | 2537 | static bool io_file_supports_async(struct file *file, int rw) |
2b188cc1 JA |
2538 | { |
2539 | umode_t mode = file_inode(file)->i_mode; | |
2540 | ||
4503b767 JA |
2541 | if (S_ISBLK(mode)) { |
2542 | if (io_bdev_nowait(file->f_inode->i_bdev)) | |
2543 | return true; | |
2544 | return false; | |
2545 | } | |
2546 | if (S_ISCHR(mode) || S_ISSOCK(mode)) | |
2b188cc1 | 2547 | return true; |
4503b767 JA |
2548 | if (S_ISREG(mode)) { |
2549 | if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) && | |
2550 | file->f_op != &io_uring_fops) | |
2551 | return true; | |
2552 | return false; | |
2553 | } | |
2b188cc1 | 2554 | |
c5b85625 JA |
2555 | /* any ->read/write should understand O_NONBLOCK */ |
2556 | if (file->f_flags & O_NONBLOCK) | |
2557 | return true; | |
2558 | ||
af197f50 JA |
2559 | if (!(file->f_mode & FMODE_NOWAIT)) |
2560 | return false; | |
2561 | ||
2562 | if (rw == READ) | |
2563 | return file->f_op->read_iter != NULL; | |
2564 | ||
2565 | return file->f_op->write_iter != NULL; | |
2b188cc1 JA |
2566 | } |
2567 | ||
3529d8c2 JA |
2568 | static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2569 | bool force_nonblock) | |
2b188cc1 | 2570 | { |
def596e9 | 2571 | struct io_ring_ctx *ctx = req->ctx; |
9adbd45d | 2572 | struct kiocb *kiocb = &req->rw.kiocb; |
09bb8394 JA |
2573 | unsigned ioprio; |
2574 | int ret; | |
2b188cc1 | 2575 | |
491381ce JA |
2576 | if (S_ISREG(file_inode(req->file)->i_mode)) |
2577 | req->flags |= REQ_F_ISREG; | |
2578 | ||
2b188cc1 | 2579 | kiocb->ki_pos = READ_ONCE(sqe->off); |
ba04291e JA |
2580 | if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) { |
2581 | req->flags |= REQ_F_CUR_POS; | |
2582 | kiocb->ki_pos = req->file->f_pos; | |
2583 | } | |
2b188cc1 | 2584 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); |
3e577dcd PB |
2585 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); |
2586 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
2587 | if (unlikely(ret)) | |
2588 | return ret; | |
2b188cc1 JA |
2589 | |
2590 | ioprio = READ_ONCE(sqe->ioprio); | |
2591 | if (ioprio) { | |
2592 | ret = ioprio_check_cap(ioprio); | |
2593 | if (ret) | |
09bb8394 | 2594 | return ret; |
2b188cc1 JA |
2595 | |
2596 | kiocb->ki_ioprio = ioprio; | |
2597 | } else | |
2598 | kiocb->ki_ioprio = get_current_ioprio(); | |
2599 | ||
8449eeda | 2600 | /* don't allow async punt if RWF_NOWAIT was requested */ |
c5b85625 | 2601 | if (kiocb->ki_flags & IOCB_NOWAIT) |
8449eeda SB |
2602 | req->flags |= REQ_F_NOWAIT; |
2603 | ||
2604 | if (force_nonblock) | |
2b188cc1 | 2605 | kiocb->ki_flags |= IOCB_NOWAIT; |
8449eeda | 2606 | |
def596e9 | 2607 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
2608 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
2609 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 2610 | return -EOPNOTSUPP; |
2b188cc1 | 2611 | |
def596e9 JA |
2612 | kiocb->ki_flags |= IOCB_HIPRI; |
2613 | kiocb->ki_complete = io_complete_rw_iopoll; | |
65a6543d | 2614 | req->iopoll_completed = 0; |
def596e9 | 2615 | } else { |
09bb8394 JA |
2616 | if (kiocb->ki_flags & IOCB_HIPRI) |
2617 | return -EINVAL; | |
def596e9 JA |
2618 | kiocb->ki_complete = io_complete_rw; |
2619 | } | |
9adbd45d | 2620 | |
3529d8c2 JA |
2621 | req->rw.addr = READ_ONCE(sqe->addr); |
2622 | req->rw.len = READ_ONCE(sqe->len); | |
4f4eeba8 | 2623 | req->buf_index = READ_ONCE(sqe->buf_index); |
2b188cc1 | 2624 | return 0; |
2b188cc1 JA |
2625 | } |
2626 | ||
2627 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
2628 | { | |
2629 | switch (ret) { | |
2630 | case -EIOCBQUEUED: | |
2631 | break; | |
2632 | case -ERESTARTSYS: | |
2633 | case -ERESTARTNOINTR: | |
2634 | case -ERESTARTNOHAND: | |
2635 | case -ERESTART_RESTARTBLOCK: | |
2636 | /* | |
2637 | * We can't just restart the syscall, since previously | |
2638 | * submitted sqes may already be in progress. Just fail this | |
2639 | * IO with EINTR. | |
2640 | */ | |
2641 | ret = -EINTR; | |
df561f66 | 2642 | fallthrough; |
2b188cc1 JA |
2643 | default: |
2644 | kiocb->ki_complete(kiocb, ret, 0); | |
2645 | } | |
2646 | } | |
2647 | ||
a1d7c393 JA |
2648 | static void kiocb_done(struct kiocb *kiocb, ssize_t ret, |
2649 | struct io_comp_state *cs) | |
ba816ad6 | 2650 | { |
ba04291e JA |
2651 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
2652 | ||
227c0c96 JA |
2653 | /* add previously done IO, if any */ |
2654 | if (req->io && req->io->rw.bytes_done > 0) { | |
2655 | if (ret < 0) | |
2656 | ret = req->io->rw.bytes_done; | |
2657 | else | |
2658 | ret += req->io->rw.bytes_done; | |
2659 | } | |
2660 | ||
ba04291e JA |
2661 | if (req->flags & REQ_F_CUR_POS) |
2662 | req->file->f_pos = kiocb->ki_pos; | |
bcaec089 | 2663 | if (ret >= 0 && kiocb->ki_complete == io_complete_rw) |
a1d7c393 | 2664 | __io_complete_rw(req, ret, 0, cs); |
ba816ad6 JA |
2665 | else |
2666 | io_rw_done(kiocb, ret); | |
2667 | } | |
2668 | ||
9adbd45d | 2669 | static ssize_t io_import_fixed(struct io_kiocb *req, int rw, |
7d009165 | 2670 | struct iov_iter *iter) |
edafccee | 2671 | { |
9adbd45d JA |
2672 | struct io_ring_ctx *ctx = req->ctx; |
2673 | size_t len = req->rw.len; | |
edafccee | 2674 | struct io_mapped_ubuf *imu; |
4f4eeba8 | 2675 | u16 index, buf_index; |
edafccee JA |
2676 | size_t offset; |
2677 | u64 buf_addr; | |
2678 | ||
2679 | /* attempt to use fixed buffers without having provided iovecs */ | |
2680 | if (unlikely(!ctx->user_bufs)) | |
2681 | return -EFAULT; | |
2682 | ||
4f4eeba8 | 2683 | buf_index = req->buf_index; |
edafccee JA |
2684 | if (unlikely(buf_index >= ctx->nr_user_bufs)) |
2685 | return -EFAULT; | |
2686 | ||
2687 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
2688 | imu = &ctx->user_bufs[index]; | |
9adbd45d | 2689 | buf_addr = req->rw.addr; |
edafccee JA |
2690 | |
2691 | /* overflow */ | |
2692 | if (buf_addr + len < buf_addr) | |
2693 | return -EFAULT; | |
2694 | /* not inside the mapped region */ | |
2695 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
2696 | return -EFAULT; | |
2697 | ||
2698 | /* | |
2699 | * May not be a start of buffer, set size appropriately | |
2700 | * and advance us to the beginning. | |
2701 | */ | |
2702 | offset = buf_addr - imu->ubuf; | |
2703 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
2704 | |
2705 | if (offset) { | |
2706 | /* | |
2707 | * Don't use iov_iter_advance() here, as it's really slow for | |
2708 | * using the latter parts of a big fixed buffer - it iterates | |
2709 | * over each segment manually. We can cheat a bit here, because | |
2710 | * we know that: | |
2711 | * | |
2712 | * 1) it's a BVEC iter, we set it up | |
2713 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
2714 | * first and last bvec | |
2715 | * | |
2716 | * So just find our index, and adjust the iterator afterwards. | |
2717 | * If the offset is within the first bvec (or the whole first | |
2718 | * bvec, just use iov_iter_advance(). This makes it easier | |
2719 | * since we can just skip the first segment, which may not | |
2720 | * be PAGE_SIZE aligned. | |
2721 | */ | |
2722 | const struct bio_vec *bvec = imu->bvec; | |
2723 | ||
2724 | if (offset <= bvec->bv_len) { | |
2725 | iov_iter_advance(iter, offset); | |
2726 | } else { | |
2727 | unsigned long seg_skip; | |
2728 | ||
2729 | /* skip first vec */ | |
2730 | offset -= bvec->bv_len; | |
2731 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
2732 | ||
2733 | iter->bvec = bvec + seg_skip; | |
2734 | iter->nr_segs -= seg_skip; | |
99c79f66 | 2735 | iter->count -= bvec->bv_len + offset; |
bd11b3a3 | 2736 | iter->iov_offset = offset & ~PAGE_MASK; |
bd11b3a3 JA |
2737 | } |
2738 | } | |
2739 | ||
5e559561 | 2740 | return len; |
edafccee JA |
2741 | } |
2742 | ||
bcda7baa JA |
2743 | static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock) |
2744 | { | |
2745 | if (needs_lock) | |
2746 | mutex_unlock(&ctx->uring_lock); | |
2747 | } | |
2748 | ||
2749 | static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock) | |
2750 | { | |
2751 | /* | |
2752 | * "Normal" inline submissions always hold the uring_lock, since we | |
2753 | * grab it from the system call. Same is true for the SQPOLL offload. | |
2754 | * The only exception is when we've detached the request and issue it | |
2755 | * from an async worker thread, grab the lock for that case. | |
2756 | */ | |
2757 | if (needs_lock) | |
2758 | mutex_lock(&ctx->uring_lock); | |
2759 | } | |
2760 | ||
2761 | static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len, | |
2762 | int bgid, struct io_buffer *kbuf, | |
2763 | bool needs_lock) | |
2764 | { | |
2765 | struct io_buffer *head; | |
2766 | ||
2767 | if (req->flags & REQ_F_BUFFER_SELECTED) | |
2768 | return kbuf; | |
2769 | ||
2770 | io_ring_submit_lock(req->ctx, needs_lock); | |
2771 | ||
2772 | lockdep_assert_held(&req->ctx->uring_lock); | |
2773 | ||
2774 | head = idr_find(&req->ctx->io_buffer_idr, bgid); | |
2775 | if (head) { | |
2776 | if (!list_empty(&head->list)) { | |
2777 | kbuf = list_last_entry(&head->list, struct io_buffer, | |
2778 | list); | |
2779 | list_del(&kbuf->list); | |
2780 | } else { | |
2781 | kbuf = head; | |
2782 | idr_remove(&req->ctx->io_buffer_idr, bgid); | |
2783 | } | |
2784 | if (*len > kbuf->len) | |
2785 | *len = kbuf->len; | |
2786 | } else { | |
2787 | kbuf = ERR_PTR(-ENOBUFS); | |
2788 | } | |
2789 | ||
2790 | io_ring_submit_unlock(req->ctx, needs_lock); | |
2791 | ||
2792 | return kbuf; | |
2793 | } | |
2794 | ||
4d954c25 JA |
2795 | static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len, |
2796 | bool needs_lock) | |
2797 | { | |
2798 | struct io_buffer *kbuf; | |
4f4eeba8 | 2799 | u16 bgid; |
4d954c25 JA |
2800 | |
2801 | kbuf = (struct io_buffer *) (unsigned long) req->rw.addr; | |
4f4eeba8 | 2802 | bgid = req->buf_index; |
4d954c25 JA |
2803 | kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock); |
2804 | if (IS_ERR(kbuf)) | |
2805 | return kbuf; | |
2806 | req->rw.addr = (u64) (unsigned long) kbuf; | |
2807 | req->flags |= REQ_F_BUFFER_SELECTED; | |
2808 | return u64_to_user_ptr(kbuf->addr); | |
2809 | } | |
2810 | ||
2811 | #ifdef CONFIG_COMPAT | |
2812 | static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov, | |
2813 | bool needs_lock) | |
2814 | { | |
2815 | struct compat_iovec __user *uiov; | |
2816 | compat_ssize_t clen; | |
2817 | void __user *buf; | |
2818 | ssize_t len; | |
2819 | ||
2820 | uiov = u64_to_user_ptr(req->rw.addr); | |
2821 | if (!access_ok(uiov, sizeof(*uiov))) | |
2822 | return -EFAULT; | |
2823 | if (__get_user(clen, &uiov->iov_len)) | |
2824 | return -EFAULT; | |
2825 | if (clen < 0) | |
2826 | return -EINVAL; | |
2827 | ||
2828 | len = clen; | |
2829 | buf = io_rw_buffer_select(req, &len, needs_lock); | |
2830 | if (IS_ERR(buf)) | |
2831 | return PTR_ERR(buf); | |
2832 | iov[0].iov_base = buf; | |
2833 | iov[0].iov_len = (compat_size_t) len; | |
2834 | return 0; | |
2835 | } | |
2836 | #endif | |
2837 | ||
2838 | static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov, | |
2839 | bool needs_lock) | |
2840 | { | |
2841 | struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr); | |
2842 | void __user *buf; | |
2843 | ssize_t len; | |
2844 | ||
2845 | if (copy_from_user(iov, uiov, sizeof(*uiov))) | |
2846 | return -EFAULT; | |
2847 | ||
2848 | len = iov[0].iov_len; | |
2849 | if (len < 0) | |
2850 | return -EINVAL; | |
2851 | buf = io_rw_buffer_select(req, &len, needs_lock); | |
2852 | if (IS_ERR(buf)) | |
2853 | return PTR_ERR(buf); | |
2854 | iov[0].iov_base = buf; | |
2855 | iov[0].iov_len = len; | |
2856 | return 0; | |
2857 | } | |
2858 | ||
2859 | static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov, | |
2860 | bool needs_lock) | |
2861 | { | |
dddb3e26 JA |
2862 | if (req->flags & REQ_F_BUFFER_SELECTED) { |
2863 | struct io_buffer *kbuf; | |
2864 | ||
2865 | kbuf = (struct io_buffer *) (unsigned long) req->rw.addr; | |
2866 | iov[0].iov_base = u64_to_user_ptr(kbuf->addr); | |
2867 | iov[0].iov_len = kbuf->len; | |
4d954c25 | 2868 | return 0; |
dddb3e26 | 2869 | } |
4d954c25 JA |
2870 | if (!req->rw.len) |
2871 | return 0; | |
2872 | else if (req->rw.len > 1) | |
2873 | return -EINVAL; | |
2874 | ||
2875 | #ifdef CONFIG_COMPAT | |
2876 | if (req->ctx->compat) | |
2877 | return io_compat_import(req, iov, needs_lock); | |
2878 | #endif | |
2879 | ||
2880 | return __io_iov_buffer_select(req, iov, needs_lock); | |
2881 | } | |
2882 | ||
8452fd0c JA |
2883 | static ssize_t __io_import_iovec(int rw, struct io_kiocb *req, |
2884 | struct iovec **iovec, struct iov_iter *iter, | |
2885 | bool needs_lock) | |
2b188cc1 | 2886 | { |
9adbd45d JA |
2887 | void __user *buf = u64_to_user_ptr(req->rw.addr); |
2888 | size_t sqe_len = req->rw.len; | |
4d954c25 | 2889 | ssize_t ret; |
edafccee JA |
2890 | u8 opcode; |
2891 | ||
d625c6ee | 2892 | opcode = req->opcode; |
7d009165 | 2893 | if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) { |
edafccee | 2894 | *iovec = NULL; |
9adbd45d | 2895 | return io_import_fixed(req, rw, iter); |
edafccee | 2896 | } |
2b188cc1 | 2897 | |
bcda7baa | 2898 | /* buffer index only valid with fixed read/write, or buffer select */ |
4f4eeba8 | 2899 | if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT)) |
9adbd45d JA |
2900 | return -EINVAL; |
2901 | ||
3a6820f2 | 2902 | if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) { |
bcda7baa | 2903 | if (req->flags & REQ_F_BUFFER_SELECT) { |
4d954c25 | 2904 | buf = io_rw_buffer_select(req, &sqe_len, needs_lock); |
867a23ea | 2905 | if (IS_ERR(buf)) |
4d954c25 | 2906 | return PTR_ERR(buf); |
3f9d6441 | 2907 | req->rw.len = sqe_len; |
bcda7baa JA |
2908 | } |
2909 | ||
3a6820f2 JA |
2910 | ret = import_single_range(rw, buf, sqe_len, *iovec, iter); |
2911 | *iovec = NULL; | |
3a901598 | 2912 | return ret < 0 ? ret : sqe_len; |
3a6820f2 JA |
2913 | } |
2914 | ||
4d954c25 JA |
2915 | if (req->flags & REQ_F_BUFFER_SELECT) { |
2916 | ret = io_iov_buffer_select(req, *iovec, needs_lock); | |
3f9d6441 JA |
2917 | if (!ret) { |
2918 | ret = (*iovec)->iov_len; | |
2919 | iov_iter_init(iter, rw, *iovec, 1, ret); | |
2920 | } | |
4d954c25 JA |
2921 | *iovec = NULL; |
2922 | return ret; | |
2923 | } | |
2924 | ||
2b188cc1 | 2925 | #ifdef CONFIG_COMPAT |
cf6fd4bd | 2926 | if (req->ctx->compat) |
2b188cc1 JA |
2927 | return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV, |
2928 | iovec, iter); | |
2929 | #endif | |
2930 | ||
2931 | return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter); | |
2932 | } | |
2933 | ||
8452fd0c JA |
2934 | static ssize_t io_import_iovec(int rw, struct io_kiocb *req, |
2935 | struct iovec **iovec, struct iov_iter *iter, | |
2936 | bool needs_lock) | |
2937 | { | |
2938 | if (!req->io) | |
2939 | return __io_import_iovec(rw, req, iovec, iter, needs_lock); | |
2940 | *iovec = NULL; | |
2941 | return iov_iter_count(&req->io->rw.iter); | |
2942 | } | |
2943 | ||
0fef9483 JA |
2944 | static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb) |
2945 | { | |
2946 | return kiocb->ki_filp->f_mode & FMODE_STREAM ? NULL : &kiocb->ki_pos; | |
2947 | } | |
2948 | ||
31b51510 | 2949 | /* |
32960613 JA |
2950 | * For files that don't have ->read_iter() and ->write_iter(), handle them |
2951 | * by looping over ->read() or ->write() manually. | |
31b51510 | 2952 | */ |
32960613 JA |
2953 | static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb, |
2954 | struct iov_iter *iter) | |
2955 | { | |
2956 | ssize_t ret = 0; | |
2957 | ||
2958 | /* | |
2959 | * Don't support polled IO through this interface, and we can't | |
2960 | * support non-blocking either. For the latter, this just causes | |
2961 | * the kiocb to be handled from an async context. | |
2962 | */ | |
2963 | if (kiocb->ki_flags & IOCB_HIPRI) | |
2964 | return -EOPNOTSUPP; | |
2965 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
2966 | return -EAGAIN; | |
2967 | ||
2968 | while (iov_iter_count(iter)) { | |
311ae9e1 | 2969 | struct iovec iovec; |
32960613 JA |
2970 | ssize_t nr; |
2971 | ||
311ae9e1 PB |
2972 | if (!iov_iter_is_bvec(iter)) { |
2973 | iovec = iov_iter_iovec(iter); | |
2974 | } else { | |
2975 | /* fixed buffers import bvec */ | |
2976 | iovec.iov_base = kmap(iter->bvec->bv_page) | |
2977 | + iter->iov_offset; | |
2978 | iovec.iov_len = min(iter->count, | |
2979 | iter->bvec->bv_len - iter->iov_offset); | |
2980 | } | |
2981 | ||
32960613 JA |
2982 | if (rw == READ) { |
2983 | nr = file->f_op->read(file, iovec.iov_base, | |
0fef9483 | 2984 | iovec.iov_len, io_kiocb_ppos(kiocb)); |
32960613 JA |
2985 | } else { |
2986 | nr = file->f_op->write(file, iovec.iov_base, | |
0fef9483 | 2987 | iovec.iov_len, io_kiocb_ppos(kiocb)); |
32960613 JA |
2988 | } |
2989 | ||
311ae9e1 PB |
2990 | if (iov_iter_is_bvec(iter)) |
2991 | kunmap(iter->bvec->bv_page); | |
2992 | ||
32960613 JA |
2993 | if (nr < 0) { |
2994 | if (!ret) | |
2995 | ret = nr; | |
2996 | break; | |
2997 | } | |
2998 | ret += nr; | |
2999 | if (nr != iovec.iov_len) | |
3000 | break; | |
3001 | iov_iter_advance(iter, nr); | |
3002 | } | |
3003 | ||
3004 | return ret; | |
3005 | } | |
3006 | ||
ff6165b2 JA |
3007 | static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec, |
3008 | const struct iovec *fast_iov, struct iov_iter *iter) | |
f67676d1 | 3009 | { |
b64e3444 PB |
3010 | struct io_async_rw *rw = &req->io->rw; |
3011 | ||
ff6165b2 JA |
3012 | memcpy(&rw->iter, iter, sizeof(*iter)); |
3013 | rw->free_iovec = NULL; | |
227c0c96 | 3014 | rw->bytes_done = 0; |
ff6165b2 JA |
3015 | /* can only be fixed buffers, no need to do anything */ |
3016 | if (iter->type == ITER_BVEC) | |
3017 | return; | |
b64e3444 | 3018 | if (!iovec) { |
ff6165b2 JA |
3019 | unsigned iov_off = 0; |
3020 | ||
3021 | rw->iter.iov = rw->fast_iov; | |
3022 | if (iter->iov != fast_iov) { | |
3023 | iov_off = iter->iov - fast_iov; | |
3024 | rw->iter.iov += iov_off; | |
3025 | } | |
3026 | if (rw->fast_iov != fast_iov) | |
3027 | memcpy(rw->fast_iov + iov_off, fast_iov + iov_off, | |
45097dae | 3028 | sizeof(struct iovec) * iter->nr_segs); |
99bc4c38 | 3029 | } else { |
ff6165b2 | 3030 | rw->free_iovec = iovec; |
99bc4c38 | 3031 | req->flags |= REQ_F_NEED_CLEANUP; |
f67676d1 JA |
3032 | } |
3033 | } | |
3034 | ||
3d9932a8 XW |
3035 | static inline int __io_alloc_async_ctx(struct io_kiocb *req) |
3036 | { | |
3037 | req->io = kmalloc(sizeof(*req->io), GFP_KERNEL); | |
3038 | return req->io == NULL; | |
3039 | } | |
3040 | ||
b7bb4f7d | 3041 | static int io_alloc_async_ctx(struct io_kiocb *req) |
f67676d1 | 3042 | { |
d3656344 JA |
3043 | if (!io_op_defs[req->opcode].async_ctx) |
3044 | return 0; | |
3d9932a8 XW |
3045 | |
3046 | return __io_alloc_async_ctx(req); | |
b7bb4f7d JA |
3047 | } |
3048 | ||
ff6165b2 JA |
3049 | static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec, |
3050 | const struct iovec *fast_iov, | |
227c0c96 | 3051 | struct iov_iter *iter, bool force) |
b7bb4f7d | 3052 | { |
227c0c96 | 3053 | if (!force && !io_op_defs[req->opcode].async_ctx) |
74566df3 | 3054 | return 0; |
5d204bcf | 3055 | if (!req->io) { |
3d9932a8 | 3056 | if (__io_alloc_async_ctx(req)) |
5d204bcf | 3057 | return -ENOMEM; |
b7bb4f7d | 3058 | |
ff6165b2 | 3059 | io_req_map_rw(req, iovec, fast_iov, iter); |
5d204bcf | 3060 | } |
b7bb4f7d | 3061 | return 0; |
f67676d1 JA |
3062 | } |
3063 | ||
c3e330a4 PB |
3064 | static inline int io_rw_prep_async(struct io_kiocb *req, int rw, |
3065 | bool force_nonblock) | |
3066 | { | |
ff6165b2 | 3067 | struct io_async_rw *iorw = &req->io->rw; |
c183edff | 3068 | struct iovec *iov; |
c3e330a4 PB |
3069 | ssize_t ret; |
3070 | ||
c183edff JA |
3071 | iorw->iter.iov = iov = iorw->fast_iov; |
3072 | ret = __io_import_iovec(rw, req, &iov, &iorw->iter, !force_nonblock); | |
c3e330a4 PB |
3073 | if (unlikely(ret < 0)) |
3074 | return ret; | |
3075 | ||
c183edff | 3076 | iorw->iter.iov = iov; |
ff6165b2 | 3077 | io_req_map_rw(req, iorw->iter.iov, iorw->fast_iov, &iorw->iter); |
c3e330a4 PB |
3078 | return 0; |
3079 | } | |
3080 | ||
3529d8c2 JA |
3081 | static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
3082 | bool force_nonblock) | |
f67676d1 JA |
3083 | { |
3084 | ssize_t ret; | |
3085 | ||
3529d8c2 JA |
3086 | ret = io_prep_rw(req, sqe, force_nonblock); |
3087 | if (ret) | |
3088 | return ret; | |
f67676d1 | 3089 | |
3529d8c2 JA |
3090 | if (unlikely(!(req->file->f_mode & FMODE_READ))) |
3091 | return -EBADF; | |
f67676d1 | 3092 | |
5f798bea PB |
3093 | /* either don't need iovec imported or already have it */ |
3094 | if (!req->io || req->flags & REQ_F_NEED_CLEANUP) | |
3529d8c2 | 3095 | return 0; |
c3e330a4 | 3096 | return io_rw_prep_async(req, READ, force_nonblock); |
f67676d1 JA |
3097 | } |
3098 | ||
c1dd91d1 JA |
3099 | /* |
3100 | * This is our waitqueue callback handler, registered through lock_page_async() | |
3101 | * when we initially tried to do the IO with the iocb armed our waitqueue. | |
3102 | * This gets called when the page is unlocked, and we generally expect that to | |
3103 | * happen when the page IO is completed and the page is now uptodate. This will | |
3104 | * queue a task_work based retry of the operation, attempting to copy the data | |
3105 | * again. If the latter fails because the page was NOT uptodate, then we will | |
3106 | * do a thread based blocking retry of the operation. That's the unexpected | |
3107 | * slow path. | |
3108 | */ | |
bcf5a063 JA |
3109 | static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode, |
3110 | int sync, void *arg) | |
3111 | { | |
3112 | struct wait_page_queue *wpq; | |
3113 | struct io_kiocb *req = wait->private; | |
bcf5a063 | 3114 | struct wait_page_key *key = arg; |
bcf5a063 JA |
3115 | int ret; |
3116 | ||
3117 | wpq = container_of(wait, struct wait_page_queue, wait); | |
3118 | ||
cdc8fcb4 LT |
3119 | if (!wake_page_match(wpq, key)) |
3120 | return 0; | |
3121 | ||
c8d317aa | 3122 | req->rw.kiocb.ki_flags &= ~IOCB_WAITQ; |
bcf5a063 JA |
3123 | list_del_init(&wait->entry); |
3124 | ||
e7375122 | 3125 | init_task_work(&req->task_work, io_req_task_submit); |
6d816e08 JA |
3126 | percpu_ref_get(&req->ctx->refs); |
3127 | ||
bcf5a063 JA |
3128 | /* submit ref gets dropped, acquire a new one */ |
3129 | refcount_inc(&req->refs); | |
fd7d6de2 | 3130 | ret = io_req_task_work_add(req, &req->task_work, true); |
bcf5a063 | 3131 | if (unlikely(ret)) { |
c2c4c83c JA |
3132 | struct task_struct *tsk; |
3133 | ||
bcf5a063 | 3134 | /* queue just for cancelation */ |
e7375122 | 3135 | init_task_work(&req->task_work, io_req_task_cancel); |
bcf5a063 | 3136 | tsk = io_wq_get_task(req->ctx->io_wq); |
e7375122 | 3137 | task_work_add(tsk, &req->task_work, 0); |
c2c4c83c | 3138 | wake_up_process(tsk); |
bcf5a063 | 3139 | } |
bcf5a063 JA |
3140 | return 1; |
3141 | } | |
3142 | ||
c1dd91d1 JA |
3143 | /* |
3144 | * This controls whether a given IO request should be armed for async page | |
3145 | * based retry. If we return false here, the request is handed to the async | |
3146 | * worker threads for retry. If we're doing buffered reads on a regular file, | |
3147 | * we prepare a private wait_page_queue entry and retry the operation. This | |
3148 | * will either succeed because the page is now uptodate and unlocked, or it | |
3149 | * will register a callback when the page is unlocked at IO completion. Through | |
3150 | * that callback, io_uring uses task_work to setup a retry of the operation. | |
3151 | * That retry will attempt the buffered read again. The retry will generally | |
3152 | * succeed, or in rare cases where it fails, we then fall back to using the | |
3153 | * async worker threads for a blocking retry. | |
3154 | */ | |
227c0c96 | 3155 | static bool io_rw_should_retry(struct io_kiocb *req) |
f67676d1 | 3156 | { |
3b2a4439 | 3157 | struct wait_page_queue *wait = &req->io->rw.wpq; |
bcf5a063 | 3158 | struct kiocb *kiocb = &req->rw.kiocb; |
f67676d1 | 3159 | |
bcf5a063 JA |
3160 | /* never retry for NOWAIT, we just complete with -EAGAIN */ |
3161 | if (req->flags & REQ_F_NOWAIT) | |
3162 | return false; | |
f67676d1 | 3163 | |
227c0c96 | 3164 | /* Only for buffered IO */ |
3b2a4439 | 3165 | if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI)) |
bcf5a063 | 3166 | return false; |
3b2a4439 | 3167 | |
bcf5a063 JA |
3168 | /* |
3169 | * just use poll if we can, and don't attempt if the fs doesn't | |
3170 | * support callback based unlocks | |
3171 | */ | |
3172 | if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC)) | |
3173 | return false; | |
f67676d1 | 3174 | |
3b2a4439 JA |
3175 | wait->wait.func = io_async_buf_func; |
3176 | wait->wait.private = req; | |
3177 | wait->wait.flags = 0; | |
3178 | INIT_LIST_HEAD(&wait->wait.entry); | |
3179 | kiocb->ki_flags |= IOCB_WAITQ; | |
c8d317aa | 3180 | kiocb->ki_flags &= ~IOCB_NOWAIT; |
3b2a4439 | 3181 | kiocb->ki_waitq = wait; |
3b2a4439 | 3182 | return true; |
bcf5a063 JA |
3183 | } |
3184 | ||
3185 | static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter) | |
3186 | { | |
3187 | if (req->file->f_op->read_iter) | |
3188 | return call_read_iter(req->file, &req->rw.kiocb, iter); | |
2dd2111d GH |
3189 | else if (req->file->f_op->read) |
3190 | return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter); | |
3191 | else | |
3192 | return -EINVAL; | |
f67676d1 JA |
3193 | } |
3194 | ||
a1d7c393 JA |
3195 | static int io_read(struct io_kiocb *req, bool force_nonblock, |
3196 | struct io_comp_state *cs) | |
2b188cc1 JA |
3197 | { |
3198 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
9adbd45d | 3199 | struct kiocb *kiocb = &req->rw.kiocb; |
ff6165b2 | 3200 | struct iov_iter __iter, *iter = &__iter; |
227c0c96 | 3201 | ssize_t io_size, ret, ret2; |
31b51510 | 3202 | size_t iov_count; |
f5cac8b1 | 3203 | bool no_async; |
ff6165b2 JA |
3204 | |
3205 | if (req->io) | |
3206 | iter = &req->io->rw.iter; | |
2b188cc1 | 3207 | |
ff6165b2 | 3208 | ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock); |
06b76d44 JA |
3209 | if (ret < 0) |
3210 | return ret; | |
eefdf30f | 3211 | iov_count = iov_iter_count(iter); |
fa15bafb PB |
3212 | io_size = ret; |
3213 | req->result = io_size; | |
227c0c96 | 3214 | ret = 0; |
2b188cc1 | 3215 | |
fd6c2e4c JA |
3216 | /* Ensure we clear previously set non-block flag */ |
3217 | if (!force_nonblock) | |
29de5f6a | 3218 | kiocb->ki_flags &= ~IOCB_NOWAIT; |
fd6c2e4c | 3219 | |
24c74678 | 3220 | /* If the file doesn't support async, just async punt */ |
f5cac8b1 JA |
3221 | no_async = force_nonblock && !io_file_supports_async(req->file, READ); |
3222 | if (no_async) | |
f67676d1 | 3223 | goto copy_iov; |
9e645e11 | 3224 | |
0fef9483 | 3225 | ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count); |
fa15bafb PB |
3226 | if (unlikely(ret)) |
3227 | goto out_free; | |
2b188cc1 | 3228 | |
227c0c96 | 3229 | ret = io_iter_do_read(req, iter); |
32960613 | 3230 | |
227c0c96 JA |
3231 | if (!ret) { |
3232 | goto done; | |
3233 | } else if (ret == -EIOCBQUEUED) { | |
3234 | ret = 0; | |
3235 | goto out_free; | |
3236 | } else if (ret == -EAGAIN) { | |
eefdf30f JA |
3237 | /* IOPOLL retry should happen for io-wq threads */ |
3238 | if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
f91daf56 | 3239 | goto done; |
355afaeb JA |
3240 | /* no retry on NONBLOCK marked file */ |
3241 | if (req->file->f_flags & O_NONBLOCK) | |
3242 | goto done; | |
84216315 JA |
3243 | /* some cases will consume bytes even on error returns */ |
3244 | iov_iter_revert(iter, iov_count - iov_iter_count(iter)); | |
f38c7e3a JA |
3245 | ret = 0; |
3246 | goto copy_iov; | |
227c0c96 | 3247 | } else if (ret < 0) { |
00d23d51 JA |
3248 | /* make sure -ERESTARTSYS -> -EINTR is done */ |
3249 | goto done; | |
227c0c96 JA |
3250 | } |
3251 | ||
3252 | /* read it all, or we did blocking attempt. no retry. */ | |
f91daf56 JA |
3253 | if (!iov_iter_count(iter) || !force_nonblock || |
3254 | (req->file->f_flags & O_NONBLOCK)) | |
227c0c96 JA |
3255 | goto done; |
3256 | ||
3257 | io_size -= ret; | |
3258 | copy_iov: | |
3259 | ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true); | |
3260 | if (ret2) { | |
3261 | ret = ret2; | |
3262 | goto out_free; | |
3263 | } | |
f5cac8b1 JA |
3264 | if (no_async) |
3265 | return -EAGAIN; | |
227c0c96 JA |
3266 | /* it's copied and will be cleaned with ->io */ |
3267 | iovec = NULL; | |
3268 | /* now use our persistent iterator, if we aren't already */ | |
3269 | iter = &req->io->rw.iter; | |
3270 | retry: | |
3271 | req->io->rw.bytes_done += ret; | |
3272 | /* if we can retry, do so with the callbacks armed */ | |
3273 | if (!io_rw_should_retry(req)) { | |
fa15bafb PB |
3274 | kiocb->ki_flags &= ~IOCB_WAITQ; |
3275 | return -EAGAIN; | |
2b188cc1 | 3276 | } |
227c0c96 JA |
3277 | |
3278 | /* | |
3279 | * Now retry read with the IOCB_WAITQ parts set in the iocb. If we | |
3280 | * get -EIOCBQUEUED, then we'll get a notification when the desired | |
3281 | * page gets unlocked. We can also get a partial read here, and if we | |
3282 | * do, then just retry at the new offset. | |
3283 | */ | |
3284 | ret = io_iter_do_read(req, iter); | |
3285 | if (ret == -EIOCBQUEUED) { | |
3286 | ret = 0; | |
3287 | goto out_free; | |
3288 | } else if (ret > 0 && ret < io_size) { | |
3289 | /* we got some bytes, but not all. retry. */ | |
3290 | goto retry; | |
3291 | } | |
3292 | done: | |
3293 | kiocb_done(kiocb, ret, cs); | |
3294 | ret = 0; | |
f67676d1 | 3295 | out_free: |
f261c168 | 3296 | /* it's reportedly faster than delegating the null check to kfree() */ |
252917c3 | 3297 | if (iovec) |
6f2cc166 | 3298 | kfree(iovec); |
2b188cc1 JA |
3299 | return ret; |
3300 | } | |
3301 | ||
3529d8c2 JA |
3302 | static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
3303 | bool force_nonblock) | |
f67676d1 JA |
3304 | { |
3305 | ssize_t ret; | |
3306 | ||
3529d8c2 JA |
3307 | ret = io_prep_rw(req, sqe, force_nonblock); |
3308 | if (ret) | |
3309 | return ret; | |
f67676d1 | 3310 | |
3529d8c2 JA |
3311 | if (unlikely(!(req->file->f_mode & FMODE_WRITE))) |
3312 | return -EBADF; | |
f67676d1 | 3313 | |
5f798bea PB |
3314 | /* either don't need iovec imported or already have it */ |
3315 | if (!req->io || req->flags & REQ_F_NEED_CLEANUP) | |
3529d8c2 | 3316 | return 0; |
c3e330a4 | 3317 | return io_rw_prep_async(req, WRITE, force_nonblock); |
f67676d1 JA |
3318 | } |
3319 | ||
a1d7c393 JA |
3320 | static int io_write(struct io_kiocb *req, bool force_nonblock, |
3321 | struct io_comp_state *cs) | |
2b188cc1 JA |
3322 | { |
3323 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
9adbd45d | 3324 | struct kiocb *kiocb = &req->rw.kiocb; |
ff6165b2 | 3325 | struct iov_iter __iter, *iter = &__iter; |
31b51510 | 3326 | size_t iov_count; |
fa15bafb | 3327 | ssize_t ret, ret2, io_size; |
2b188cc1 | 3328 | |
ff6165b2 JA |
3329 | if (req->io) |
3330 | iter = &req->io->rw.iter; | |
3331 | ||
3332 | ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock); | |
06b76d44 JA |
3333 | if (ret < 0) |
3334 | return ret; | |
eefdf30f | 3335 | iov_count = iov_iter_count(iter); |
fa15bafb PB |
3336 | io_size = ret; |
3337 | req->result = io_size; | |
2b188cc1 | 3338 | |
fd6c2e4c JA |
3339 | /* Ensure we clear previously set non-block flag */ |
3340 | if (!force_nonblock) | |
9adbd45d | 3341 | req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT; |
fd6c2e4c | 3342 | |
24c74678 | 3343 | /* If the file doesn't support async, just async punt */ |
af197f50 | 3344 | if (force_nonblock && !io_file_supports_async(req->file, WRITE)) |
f67676d1 | 3345 | goto copy_iov; |
31b51510 | 3346 | |
10d59345 JA |
3347 | /* file path doesn't support NOWAIT for non-direct_IO */ |
3348 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) && | |
3349 | (req->flags & REQ_F_ISREG)) | |
f67676d1 | 3350 | goto copy_iov; |
31b51510 | 3351 | |
0fef9483 | 3352 | ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count); |
fa15bafb PB |
3353 | if (unlikely(ret)) |
3354 | goto out_free; | |
4ed734b0 | 3355 | |
fa15bafb PB |
3356 | /* |
3357 | * Open-code file_start_write here to grab freeze protection, | |
3358 | * which will be released by another thread in | |
3359 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
3360 | * released so that it doesn't complain about the held lock when | |
3361 | * we return to userspace. | |
3362 | */ | |
3363 | if (req->flags & REQ_F_ISREG) { | |
3364 | __sb_start_write(file_inode(req->file)->i_sb, | |
3365 | SB_FREEZE_WRITE, true); | |
3366 | __sb_writers_release(file_inode(req->file)->i_sb, | |
3367 | SB_FREEZE_WRITE); | |
3368 | } | |
3369 | kiocb->ki_flags |= IOCB_WRITE; | |
4ed734b0 | 3370 | |
fa15bafb | 3371 | if (req->file->f_op->write_iter) |
ff6165b2 | 3372 | ret2 = call_write_iter(req->file, kiocb, iter); |
2dd2111d | 3373 | else if (req->file->f_op->write) |
ff6165b2 | 3374 | ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter); |
2dd2111d GH |
3375 | else |
3376 | ret2 = -EINVAL; | |
4ed734b0 | 3377 | |
fa15bafb PB |
3378 | /* |
3379 | * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just | |
3380 | * retry them without IOCB_NOWAIT. | |
3381 | */ | |
3382 | if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT)) | |
3383 | ret2 = -EAGAIN; | |
355afaeb JA |
3384 | /* no retry on NONBLOCK marked file */ |
3385 | if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK)) | |
3386 | goto done; | |
fa15bafb | 3387 | if (!force_nonblock || ret2 != -EAGAIN) { |
eefdf30f JA |
3388 | /* IOPOLL retry should happen for io-wq threads */ |
3389 | if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN) | |
3390 | goto copy_iov; | |
355afaeb | 3391 | done: |
fa15bafb PB |
3392 | kiocb_done(kiocb, ret2, cs); |
3393 | } else { | |
f67676d1 | 3394 | copy_iov: |
84216315 JA |
3395 | /* some cases will consume bytes even on error returns */ |
3396 | iov_iter_revert(iter, iov_count - iov_iter_count(iter)); | |
227c0c96 | 3397 | ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false); |
ff6165b2 JA |
3398 | if (!ret) |
3399 | return -EAGAIN; | |
2b188cc1 | 3400 | } |
31b51510 | 3401 | out_free: |
f261c168 | 3402 | /* it's reportedly faster than delegating the null check to kfree() */ |
252917c3 | 3403 | if (iovec) |
6f2cc166 | 3404 | kfree(iovec); |
2b188cc1 JA |
3405 | return ret; |
3406 | } | |
3407 | ||
f2a8d5c7 PB |
3408 | static int __io_splice_prep(struct io_kiocb *req, |
3409 | const struct io_uring_sqe *sqe) | |
7d67af2c PB |
3410 | { |
3411 | struct io_splice* sp = &req->splice; | |
3412 | unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL; | |
3413 | int ret; | |
3414 | ||
3415 | if (req->flags & REQ_F_NEED_CLEANUP) | |
3416 | return 0; | |
3232dd02 PB |
3417 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3418 | return -EINVAL; | |
7d67af2c PB |
3419 | |
3420 | sp->file_in = NULL; | |
7d67af2c PB |
3421 | sp->len = READ_ONCE(sqe->len); |
3422 | sp->flags = READ_ONCE(sqe->splice_flags); | |
3423 | ||
3424 | if (unlikely(sp->flags & ~valid_flags)) | |
3425 | return -EINVAL; | |
3426 | ||
3427 | ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in, | |
3428 | (sp->flags & SPLICE_F_FD_IN_FIXED)); | |
3429 | if (ret) | |
3430 | return ret; | |
3431 | req->flags |= REQ_F_NEED_CLEANUP; | |
3432 | ||
7cdaf587 XW |
3433 | if (!S_ISREG(file_inode(sp->file_in)->i_mode)) { |
3434 | /* | |
3435 | * Splice operation will be punted aync, and here need to | |
3436 | * modify io_wq_work.flags, so initialize io_wq_work firstly. | |
3437 | */ | |
3438 | io_req_init_async(req); | |
7d67af2c | 3439 | req->work.flags |= IO_WQ_WORK_UNBOUND; |
7cdaf587 | 3440 | } |
7d67af2c PB |
3441 | |
3442 | return 0; | |
3443 | } | |
3444 | ||
f2a8d5c7 PB |
3445 | static int io_tee_prep(struct io_kiocb *req, |
3446 | const struct io_uring_sqe *sqe) | |
3447 | { | |
3448 | if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off)) | |
3449 | return -EINVAL; | |
3450 | return __io_splice_prep(req, sqe); | |
3451 | } | |
3452 | ||
3453 | static int io_tee(struct io_kiocb *req, bool force_nonblock) | |
3454 | { | |
3455 | struct io_splice *sp = &req->splice; | |
3456 | struct file *in = sp->file_in; | |
3457 | struct file *out = sp->file_out; | |
3458 | unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED; | |
3459 | long ret = 0; | |
3460 | ||
3461 | if (force_nonblock) | |
3462 | return -EAGAIN; | |
3463 | if (sp->len) | |
3464 | ret = do_tee(in, out, sp->len, flags); | |
3465 | ||
3466 | io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED)); | |
3467 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
3468 | ||
f2a8d5c7 PB |
3469 | if (ret != sp->len) |
3470 | req_set_fail_links(req); | |
e1e16097 | 3471 | io_req_complete(req, ret); |
f2a8d5c7 PB |
3472 | return 0; |
3473 | } | |
3474 | ||
3475 | static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
3476 | { | |
3477 | struct io_splice* sp = &req->splice; | |
3478 | ||
3479 | sp->off_in = READ_ONCE(sqe->splice_off_in); | |
3480 | sp->off_out = READ_ONCE(sqe->off); | |
3481 | return __io_splice_prep(req, sqe); | |
3482 | } | |
3483 | ||
014db007 | 3484 | static int io_splice(struct io_kiocb *req, bool force_nonblock) |
7d67af2c PB |
3485 | { |
3486 | struct io_splice *sp = &req->splice; | |
3487 | struct file *in = sp->file_in; | |
3488 | struct file *out = sp->file_out; | |
3489 | unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED; | |
3490 | loff_t *poff_in, *poff_out; | |
c9687426 | 3491 | long ret = 0; |
7d67af2c | 3492 | |
2fb3e822 PB |
3493 | if (force_nonblock) |
3494 | return -EAGAIN; | |
7d67af2c PB |
3495 | |
3496 | poff_in = (sp->off_in == -1) ? NULL : &sp->off_in; | |
3497 | poff_out = (sp->off_out == -1) ? NULL : &sp->off_out; | |
c9687426 | 3498 | |
948a7749 | 3499 | if (sp->len) |
c9687426 | 3500 | ret = do_splice(in, poff_in, out, poff_out, sp->len, flags); |
7d67af2c PB |
3501 | |
3502 | io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED)); | |
3503 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
3504 | ||
7d67af2c PB |
3505 | if (ret != sp->len) |
3506 | req_set_fail_links(req); | |
e1e16097 | 3507 | io_req_complete(req, ret); |
7d67af2c PB |
3508 | return 0; |
3509 | } | |
3510 | ||
2b188cc1 JA |
3511 | /* |
3512 | * IORING_OP_NOP just posts a completion event, nothing else. | |
3513 | */ | |
229a7b63 | 3514 | static int io_nop(struct io_kiocb *req, struct io_comp_state *cs) |
2b188cc1 JA |
3515 | { |
3516 | struct io_ring_ctx *ctx = req->ctx; | |
2b188cc1 | 3517 | |
def596e9 JA |
3518 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
3519 | return -EINVAL; | |
3520 | ||
229a7b63 | 3521 | __io_req_complete(req, 0, 0, cs); |
2b188cc1 JA |
3522 | return 0; |
3523 | } | |
3524 | ||
3529d8c2 | 3525 | static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
c992fe29 | 3526 | { |
6b06314c | 3527 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 3528 | |
09bb8394 JA |
3529 | if (!req->file) |
3530 | return -EBADF; | |
c992fe29 | 3531 | |
6b06314c | 3532 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 3533 | return -EINVAL; |
edafccee | 3534 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
3535 | return -EINVAL; |
3536 | ||
8ed8d3c3 JA |
3537 | req->sync.flags = READ_ONCE(sqe->fsync_flags); |
3538 | if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC)) | |
3539 | return -EINVAL; | |
3540 | ||
3541 | req->sync.off = READ_ONCE(sqe->off); | |
3542 | req->sync.len = READ_ONCE(sqe->len); | |
c992fe29 CH |
3543 | return 0; |
3544 | } | |
3545 | ||
ac45abc0 | 3546 | static int io_fsync(struct io_kiocb *req, bool force_nonblock) |
8ed8d3c3 | 3547 | { |
8ed8d3c3 | 3548 | loff_t end = req->sync.off + req->sync.len; |
8ed8d3c3 JA |
3549 | int ret; |
3550 | ||
ac45abc0 PB |
3551 | /* fsync always requires a blocking context */ |
3552 | if (force_nonblock) | |
3553 | return -EAGAIN; | |
3554 | ||
9adbd45d | 3555 | ret = vfs_fsync_range(req->file, req->sync.off, |
8ed8d3c3 JA |
3556 | end > 0 ? end : LLONG_MAX, |
3557 | req->sync.flags & IORING_FSYNC_DATASYNC); | |
3558 | if (ret < 0) | |
3559 | req_set_fail_links(req); | |
e1e16097 | 3560 | io_req_complete(req, ret); |
c992fe29 CH |
3561 | return 0; |
3562 | } | |
3563 | ||
d63d1b5e JA |
3564 | static int io_fallocate_prep(struct io_kiocb *req, |
3565 | const struct io_uring_sqe *sqe) | |
3566 | { | |
3567 | if (sqe->ioprio || sqe->buf_index || sqe->rw_flags) | |
3568 | return -EINVAL; | |
3232dd02 PB |
3569 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3570 | return -EINVAL; | |
d63d1b5e JA |
3571 | |
3572 | req->sync.off = READ_ONCE(sqe->off); | |
3573 | req->sync.len = READ_ONCE(sqe->addr); | |
3574 | req->sync.mode = READ_ONCE(sqe->len); | |
3575 | return 0; | |
3576 | } | |
3577 | ||
014db007 | 3578 | static int io_fallocate(struct io_kiocb *req, bool force_nonblock) |
5d17b4a4 | 3579 | { |
ac45abc0 PB |
3580 | int ret; |
3581 | ||
d63d1b5e | 3582 | /* fallocate always requiring blocking context */ |
ac45abc0 | 3583 | if (force_nonblock) |
5d17b4a4 | 3584 | return -EAGAIN; |
ac45abc0 PB |
3585 | ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off, |
3586 | req->sync.len); | |
ac45abc0 PB |
3587 | if (ret < 0) |
3588 | req_set_fail_links(req); | |
e1e16097 | 3589 | io_req_complete(req, ret); |
5d17b4a4 JA |
3590 | return 0; |
3591 | } | |
3592 | ||
ec65fea5 | 3593 | static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
b7bb4f7d | 3594 | { |
f8748881 | 3595 | const char __user *fname; |
15b71abe | 3596 | int ret; |
b7bb4f7d | 3597 | |
ec65fea5 | 3598 | if (unlikely(sqe->ioprio || sqe->buf_index)) |
15b71abe | 3599 | return -EINVAL; |
ec65fea5 | 3600 | if (unlikely(req->flags & REQ_F_FIXED_FILE)) |
cf3040ca | 3601 | return -EBADF; |
03b1230c | 3602 | |
ec65fea5 PB |
3603 | /* open.how should be already initialised */ |
3604 | if (!(req->open.how.flags & O_PATH) && force_o_largefile()) | |
08a1d26e | 3605 | req->open.how.flags |= O_LARGEFILE; |
3529d8c2 | 3606 | |
25e72d10 PB |
3607 | req->open.dfd = READ_ONCE(sqe->fd); |
3608 | fname = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
f8748881 | 3609 | req->open.filename = getname(fname); |
15b71abe JA |
3610 | if (IS_ERR(req->open.filename)) { |
3611 | ret = PTR_ERR(req->open.filename); | |
3612 | req->open.filename = NULL; | |
3613 | return ret; | |
3614 | } | |
4022e7af | 3615 | req->open.nofile = rlimit(RLIMIT_NOFILE); |
8fef80bf | 3616 | req->flags |= REQ_F_NEED_CLEANUP; |
15b71abe | 3617 | return 0; |
03b1230c JA |
3618 | } |
3619 | ||
ec65fea5 PB |
3620 | static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
3621 | { | |
3622 | u64 flags, mode; | |
3623 | ||
4eb8dded JA |
3624 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) |
3625 | return -EINVAL; | |
ec65fea5 PB |
3626 | if (req->flags & REQ_F_NEED_CLEANUP) |
3627 | return 0; | |
3628 | mode = READ_ONCE(sqe->len); | |
3629 | flags = READ_ONCE(sqe->open_flags); | |
3630 | req->open.how = build_open_how(flags, mode); | |
3631 | return __io_openat_prep(req, sqe); | |
3632 | } | |
3633 | ||
cebdb986 | 3634 | static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
aa1fa28f | 3635 | { |
cebdb986 | 3636 | struct open_how __user *how; |
cebdb986 | 3637 | size_t len; |
0fa03c62 JA |
3638 | int ret; |
3639 | ||
4eb8dded JA |
3640 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) |
3641 | return -EINVAL; | |
0bdbdd08 PB |
3642 | if (req->flags & REQ_F_NEED_CLEANUP) |
3643 | return 0; | |
cebdb986 JA |
3644 | how = u64_to_user_ptr(READ_ONCE(sqe->addr2)); |
3645 | len = READ_ONCE(sqe->len); | |
cebdb986 JA |
3646 | if (len < OPEN_HOW_SIZE_VER0) |
3647 | return -EINVAL; | |
3529d8c2 | 3648 | |
cebdb986 JA |
3649 | ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how, |
3650 | len); | |
3651 | if (ret) | |
3652 | return ret; | |
3529d8c2 | 3653 | |
ec65fea5 | 3654 | return __io_openat_prep(req, sqe); |
cebdb986 JA |
3655 | } |
3656 | ||
014db007 | 3657 | static int io_openat2(struct io_kiocb *req, bool force_nonblock) |
15b71abe JA |
3658 | { |
3659 | struct open_flags op; | |
15b71abe JA |
3660 | struct file *file; |
3661 | int ret; | |
3662 | ||
f86cd20c | 3663 | if (force_nonblock) |
15b71abe | 3664 | return -EAGAIN; |
15b71abe | 3665 | |
cebdb986 | 3666 | ret = build_open_flags(&req->open.how, &op); |
15b71abe JA |
3667 | if (ret) |
3668 | goto err; | |
3669 | ||
4022e7af | 3670 | ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile); |
15b71abe JA |
3671 | if (ret < 0) |
3672 | goto err; | |
3673 | ||
3674 | file = do_filp_open(req->open.dfd, req->open.filename, &op); | |
3675 | if (IS_ERR(file)) { | |
3676 | put_unused_fd(ret); | |
3677 | ret = PTR_ERR(file); | |
3678 | } else { | |
3679 | fsnotify_open(file); | |
3680 | fd_install(ret, file); | |
3681 | } | |
3682 | err: | |
3683 | putname(req->open.filename); | |
8fef80bf | 3684 | req->flags &= ~REQ_F_NEED_CLEANUP; |
15b71abe JA |
3685 | if (ret < 0) |
3686 | req_set_fail_links(req); | |
e1e16097 | 3687 | io_req_complete(req, ret); |
15b71abe JA |
3688 | return 0; |
3689 | } | |
3690 | ||
014db007 | 3691 | static int io_openat(struct io_kiocb *req, bool force_nonblock) |
cebdb986 | 3692 | { |
014db007 | 3693 | return io_openat2(req, force_nonblock); |
cebdb986 JA |
3694 | } |
3695 | ||
067524e9 JA |
3696 | static int io_remove_buffers_prep(struct io_kiocb *req, |
3697 | const struct io_uring_sqe *sqe) | |
3698 | { | |
3699 | struct io_provide_buf *p = &req->pbuf; | |
3700 | u64 tmp; | |
3701 | ||
3702 | if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off) | |
3703 | return -EINVAL; | |
3704 | ||
3705 | tmp = READ_ONCE(sqe->fd); | |
3706 | if (!tmp || tmp > USHRT_MAX) | |
3707 | return -EINVAL; | |
3708 | ||
3709 | memset(p, 0, sizeof(*p)); | |
3710 | p->nbufs = tmp; | |
3711 | p->bgid = READ_ONCE(sqe->buf_group); | |
3712 | return 0; | |
3713 | } | |
3714 | ||
3715 | static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf, | |
3716 | int bgid, unsigned nbufs) | |
3717 | { | |
3718 | unsigned i = 0; | |
3719 | ||
3720 | /* shouldn't happen */ | |
3721 | if (!nbufs) | |
3722 | return 0; | |
3723 | ||
3724 | /* the head kbuf is the list itself */ | |
3725 | while (!list_empty(&buf->list)) { | |
3726 | struct io_buffer *nxt; | |
3727 | ||
3728 | nxt = list_first_entry(&buf->list, struct io_buffer, list); | |
3729 | list_del(&nxt->list); | |
3730 | kfree(nxt); | |
3731 | if (++i == nbufs) | |
3732 | return i; | |
3733 | } | |
3734 | i++; | |
3735 | kfree(buf); | |
3736 | idr_remove(&ctx->io_buffer_idr, bgid); | |
3737 | ||
3738 | return i; | |
3739 | } | |
3740 | ||
229a7b63 JA |
3741 | static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock, |
3742 | struct io_comp_state *cs) | |
067524e9 JA |
3743 | { |
3744 | struct io_provide_buf *p = &req->pbuf; | |
3745 | struct io_ring_ctx *ctx = req->ctx; | |
3746 | struct io_buffer *head; | |
3747 | int ret = 0; | |
3748 | ||
3749 | io_ring_submit_lock(ctx, !force_nonblock); | |
3750 | ||
3751 | lockdep_assert_held(&ctx->uring_lock); | |
3752 | ||
3753 | ret = -ENOENT; | |
3754 | head = idr_find(&ctx->io_buffer_idr, p->bgid); | |
3755 | if (head) | |
3756 | ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs); | |
3757 | ||
3758 | io_ring_submit_lock(ctx, !force_nonblock); | |
3759 | if (ret < 0) | |
3760 | req_set_fail_links(req); | |
229a7b63 | 3761 | __io_req_complete(req, ret, 0, cs); |
067524e9 JA |
3762 | return 0; |
3763 | } | |
3764 | ||
ddf0322d JA |
3765 | static int io_provide_buffers_prep(struct io_kiocb *req, |
3766 | const struct io_uring_sqe *sqe) | |
3767 | { | |
3768 | struct io_provide_buf *p = &req->pbuf; | |
3769 | u64 tmp; | |
3770 | ||
3771 | if (sqe->ioprio || sqe->rw_flags) | |
3772 | return -EINVAL; | |
3773 | ||
3774 | tmp = READ_ONCE(sqe->fd); | |
3775 | if (!tmp || tmp > USHRT_MAX) | |
3776 | return -E2BIG; | |
3777 | p->nbufs = tmp; | |
3778 | p->addr = READ_ONCE(sqe->addr); | |
3779 | p->len = READ_ONCE(sqe->len); | |
3780 | ||
efe68c1c | 3781 | if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs))) |
ddf0322d JA |
3782 | return -EFAULT; |
3783 | ||
3784 | p->bgid = READ_ONCE(sqe->buf_group); | |
3785 | tmp = READ_ONCE(sqe->off); | |
3786 | if (tmp > USHRT_MAX) | |
3787 | return -E2BIG; | |
3788 | p->bid = tmp; | |
3789 | return 0; | |
3790 | } | |
3791 | ||
3792 | static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head) | |
3793 | { | |
3794 | struct io_buffer *buf; | |
3795 | u64 addr = pbuf->addr; | |
3796 | int i, bid = pbuf->bid; | |
3797 | ||
3798 | for (i = 0; i < pbuf->nbufs; i++) { | |
3799 | buf = kmalloc(sizeof(*buf), GFP_KERNEL); | |
3800 | if (!buf) | |
3801 | break; | |
3802 | ||
3803 | buf->addr = addr; | |
3804 | buf->len = pbuf->len; | |
3805 | buf->bid = bid; | |
3806 | addr += pbuf->len; | |
3807 | bid++; | |
3808 | if (!*head) { | |
3809 | INIT_LIST_HEAD(&buf->list); | |
3810 | *head = buf; | |
3811 | } else { | |
3812 | list_add_tail(&buf->list, &(*head)->list); | |
3813 | } | |
3814 | } | |
3815 | ||
3816 | return i ? i : -ENOMEM; | |
3817 | } | |
3818 | ||
229a7b63 JA |
3819 | static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock, |
3820 | struct io_comp_state *cs) | |
ddf0322d JA |
3821 | { |
3822 | struct io_provide_buf *p = &req->pbuf; | |
3823 | struct io_ring_ctx *ctx = req->ctx; | |
3824 | struct io_buffer *head, *list; | |
3825 | int ret = 0; | |
3826 | ||
3827 | io_ring_submit_lock(ctx, !force_nonblock); | |
3828 | ||
3829 | lockdep_assert_held(&ctx->uring_lock); | |
3830 | ||
3831 | list = head = idr_find(&ctx->io_buffer_idr, p->bgid); | |
3832 | ||
3833 | ret = io_add_buffers(p, &head); | |
3834 | if (ret < 0) | |
3835 | goto out; | |
3836 | ||
3837 | if (!list) { | |
3838 | ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1, | |
3839 | GFP_KERNEL); | |
3840 | if (ret < 0) { | |
067524e9 | 3841 | __io_remove_buffers(ctx, head, p->bgid, -1U); |
ddf0322d JA |
3842 | goto out; |
3843 | } | |
3844 | } | |
3845 | out: | |
3846 | io_ring_submit_unlock(ctx, !force_nonblock); | |
3847 | if (ret < 0) | |
3848 | req_set_fail_links(req); | |
229a7b63 | 3849 | __io_req_complete(req, ret, 0, cs); |
ddf0322d | 3850 | return 0; |
cebdb986 JA |
3851 | } |
3852 | ||
3e4827b0 JA |
3853 | static int io_epoll_ctl_prep(struct io_kiocb *req, |
3854 | const struct io_uring_sqe *sqe) | |
3855 | { | |
3856 | #if defined(CONFIG_EPOLL) | |
3857 | if (sqe->ioprio || sqe->buf_index) | |
3858 | return -EINVAL; | |
6ca56f84 | 3859 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL))) |
3232dd02 | 3860 | return -EINVAL; |
3e4827b0 JA |
3861 | |
3862 | req->epoll.epfd = READ_ONCE(sqe->fd); | |
3863 | req->epoll.op = READ_ONCE(sqe->len); | |
3864 | req->epoll.fd = READ_ONCE(sqe->off); | |
3865 | ||
3866 | if (ep_op_has_event(req->epoll.op)) { | |
3867 | struct epoll_event __user *ev; | |
3868 | ||
3869 | ev = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
3870 | if (copy_from_user(&req->epoll.event, ev, sizeof(*ev))) | |
3871 | return -EFAULT; | |
3872 | } | |
3873 | ||
3874 | return 0; | |
3875 | #else | |
3876 | return -EOPNOTSUPP; | |
3877 | #endif | |
3878 | } | |
3879 | ||
229a7b63 JA |
3880 | static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock, |
3881 | struct io_comp_state *cs) | |
3e4827b0 JA |
3882 | { |
3883 | #if defined(CONFIG_EPOLL) | |
3884 | struct io_epoll *ie = &req->epoll; | |
3885 | int ret; | |
3886 | ||
3887 | ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock); | |
3888 | if (force_nonblock && ret == -EAGAIN) | |
3889 | return -EAGAIN; | |
3890 | ||
3891 | if (ret < 0) | |
3892 | req_set_fail_links(req); | |
229a7b63 | 3893 | __io_req_complete(req, ret, 0, cs); |
3e4827b0 JA |
3894 | return 0; |
3895 | #else | |
3896 | return -EOPNOTSUPP; | |
3897 | #endif | |
3898 | } | |
3899 | ||
c1ca757b JA |
3900 | static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
3901 | { | |
3902 | #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU) | |
3903 | if (sqe->ioprio || sqe->buf_index || sqe->off) | |
3904 | return -EINVAL; | |
3232dd02 PB |
3905 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3906 | return -EINVAL; | |
c1ca757b JA |
3907 | |
3908 | req->madvise.addr = READ_ONCE(sqe->addr); | |
3909 | req->madvise.len = READ_ONCE(sqe->len); | |
3910 | req->madvise.advice = READ_ONCE(sqe->fadvise_advice); | |
3911 | return 0; | |
3912 | #else | |
3913 | return -EOPNOTSUPP; | |
3914 | #endif | |
3915 | } | |
3916 | ||
014db007 | 3917 | static int io_madvise(struct io_kiocb *req, bool force_nonblock) |
c1ca757b JA |
3918 | { |
3919 | #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU) | |
3920 | struct io_madvise *ma = &req->madvise; | |
3921 | int ret; | |
3922 | ||
3923 | if (force_nonblock) | |
3924 | return -EAGAIN; | |
3925 | ||
3926 | ret = do_madvise(ma->addr, ma->len, ma->advice); | |
3927 | if (ret < 0) | |
3928 | req_set_fail_links(req); | |
e1e16097 | 3929 | io_req_complete(req, ret); |
c1ca757b JA |
3930 | return 0; |
3931 | #else | |
3932 | return -EOPNOTSUPP; | |
3933 | #endif | |
3934 | } | |
3935 | ||
4840e418 JA |
3936 | static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
3937 | { | |
3938 | if (sqe->ioprio || sqe->buf_index || sqe->addr) | |
3939 | return -EINVAL; | |
3232dd02 PB |
3940 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3941 | return -EINVAL; | |
4840e418 JA |
3942 | |
3943 | req->fadvise.offset = READ_ONCE(sqe->off); | |
3944 | req->fadvise.len = READ_ONCE(sqe->len); | |
3945 | req->fadvise.advice = READ_ONCE(sqe->fadvise_advice); | |
3946 | return 0; | |
3947 | } | |
3948 | ||
014db007 | 3949 | static int io_fadvise(struct io_kiocb *req, bool force_nonblock) |
4840e418 JA |
3950 | { |
3951 | struct io_fadvise *fa = &req->fadvise; | |
3952 | int ret; | |
3953 | ||
3e69426d JA |
3954 | if (force_nonblock) { |
3955 | switch (fa->advice) { | |
3956 | case POSIX_FADV_NORMAL: | |
3957 | case POSIX_FADV_RANDOM: | |
3958 | case POSIX_FADV_SEQUENTIAL: | |
3959 | break; | |
3960 | default: | |
3961 | return -EAGAIN; | |
3962 | } | |
3963 | } | |
4840e418 JA |
3964 | |
3965 | ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice); | |
3966 | if (ret < 0) | |
3967 | req_set_fail_links(req); | |
e1e16097 | 3968 | io_req_complete(req, ret); |
4840e418 JA |
3969 | return 0; |
3970 | } | |
3971 | ||
eddc7ef5 JA |
3972 | static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
3973 | { | |
6ca56f84 | 3974 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL))) |
3232dd02 | 3975 | return -EINVAL; |
eddc7ef5 JA |
3976 | if (sqe->ioprio || sqe->buf_index) |
3977 | return -EINVAL; | |
9c280f90 | 3978 | if (req->flags & REQ_F_FIXED_FILE) |
cf3040ca | 3979 | return -EBADF; |
eddc7ef5 | 3980 | |
1d9e1288 BM |
3981 | req->statx.dfd = READ_ONCE(sqe->fd); |
3982 | req->statx.mask = READ_ONCE(sqe->len); | |
e62753e4 | 3983 | req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
1d9e1288 BM |
3984 | req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2)); |
3985 | req->statx.flags = READ_ONCE(sqe->statx_flags); | |
eddc7ef5 JA |
3986 | |
3987 | return 0; | |
3988 | } | |
3989 | ||
014db007 | 3990 | static int io_statx(struct io_kiocb *req, bool force_nonblock) |
eddc7ef5 | 3991 | { |
1d9e1288 | 3992 | struct io_statx *ctx = &req->statx; |
eddc7ef5 JA |
3993 | int ret; |
3994 | ||
5b0bbee4 JA |
3995 | if (force_nonblock) { |
3996 | /* only need file table for an actual valid fd */ | |
3997 | if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD) | |
3998 | req->flags |= REQ_F_NO_FILE_TABLE; | |
eddc7ef5 | 3999 | return -EAGAIN; |
5b0bbee4 | 4000 | } |
eddc7ef5 | 4001 | |
e62753e4 BM |
4002 | ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask, |
4003 | ctx->buffer); | |
eddc7ef5 | 4004 | |
eddc7ef5 JA |
4005 | if (ret < 0) |
4006 | req_set_fail_links(req); | |
e1e16097 | 4007 | io_req_complete(req, ret); |
eddc7ef5 JA |
4008 | return 0; |
4009 | } | |
4010 | ||
b5dba59e JA |
4011 | static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4012 | { | |
4013 | /* | |
4014 | * If we queue this for async, it must not be cancellable. That would | |
7cdaf587 XW |
4015 | * leave the 'file' in an undeterminate state, and here need to modify |
4016 | * io_wq_work.flags, so initialize io_wq_work firstly. | |
b5dba59e | 4017 | */ |
7cdaf587 | 4018 | io_req_init_async(req); |
b5dba59e JA |
4019 | req->work.flags |= IO_WQ_WORK_NO_CANCEL; |
4020 | ||
3232dd02 PB |
4021 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) |
4022 | return -EINVAL; | |
b5dba59e JA |
4023 | if (sqe->ioprio || sqe->off || sqe->addr || sqe->len || |
4024 | sqe->rw_flags || sqe->buf_index) | |
4025 | return -EINVAL; | |
9c280f90 | 4026 | if (req->flags & REQ_F_FIXED_FILE) |
cf3040ca | 4027 | return -EBADF; |
b5dba59e JA |
4028 | |
4029 | req->close.fd = READ_ONCE(sqe->fd); | |
0f212204 | 4030 | if ((req->file && req->file->f_op == &io_uring_fops)) |
fd2206e4 | 4031 | return -EBADF; |
b5dba59e | 4032 | |
3af73b28 | 4033 | req->close.put_file = NULL; |
b5dba59e | 4034 | return 0; |
b5dba59e JA |
4035 | } |
4036 | ||
229a7b63 JA |
4037 | static int io_close(struct io_kiocb *req, bool force_nonblock, |
4038 | struct io_comp_state *cs) | |
b5dba59e | 4039 | { |
3af73b28 | 4040 | struct io_close *close = &req->close; |
b5dba59e JA |
4041 | int ret; |
4042 | ||
3af73b28 PB |
4043 | /* might be already done during nonblock submission */ |
4044 | if (!close->put_file) { | |
4045 | ret = __close_fd_get_file(close->fd, &close->put_file); | |
4046 | if (ret < 0) | |
4047 | return (ret == -ENOENT) ? -EBADF : ret; | |
4048 | } | |
b5dba59e JA |
4049 | |
4050 | /* if the file has a flush method, be safe and punt to async */ | |
3af73b28 | 4051 | if (close->put_file->f_op->flush && force_nonblock) { |
24c74678 PB |
4052 | /* was never set, but play safe */ |
4053 | req->flags &= ~REQ_F_NOWAIT; | |
0bf0eefd | 4054 | /* avoid grabbing files - we don't need the files */ |
24c74678 | 4055 | req->flags |= REQ_F_NO_FILE_TABLE; |
0bf0eefd | 4056 | return -EAGAIN; |
a2100672 | 4057 | } |
b5dba59e | 4058 | |
3af73b28 PB |
4059 | /* No ->flush() or already async, safely close from here */ |
4060 | ret = filp_close(close->put_file, req->work.files); | |
4061 | if (ret < 0) | |
4062 | req_set_fail_links(req); | |
3af73b28 PB |
4063 | fput(close->put_file); |
4064 | close->put_file = NULL; | |
229a7b63 | 4065 | __io_req_complete(req, ret, 0, cs); |
1a417f4e | 4066 | return 0; |
b5dba59e JA |
4067 | } |
4068 | ||
3529d8c2 | 4069 | static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
5d17b4a4 JA |
4070 | { |
4071 | struct io_ring_ctx *ctx = req->ctx; | |
5d17b4a4 JA |
4072 | |
4073 | if (!req->file) | |
4074 | return -EBADF; | |
5d17b4a4 JA |
4075 | |
4076 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
4077 | return -EINVAL; | |
4078 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
4079 | return -EINVAL; | |
4080 | ||
8ed8d3c3 JA |
4081 | req->sync.off = READ_ONCE(sqe->off); |
4082 | req->sync.len = READ_ONCE(sqe->len); | |
4083 | req->sync.flags = READ_ONCE(sqe->sync_range_flags); | |
8ed8d3c3 JA |
4084 | return 0; |
4085 | } | |
4086 | ||
ac45abc0 | 4087 | static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock) |
8ed8d3c3 | 4088 | { |
8ed8d3c3 JA |
4089 | int ret; |
4090 | ||
ac45abc0 PB |
4091 | /* sync_file_range always requires a blocking context */ |
4092 | if (force_nonblock) | |
4093 | return -EAGAIN; | |
4094 | ||
9adbd45d | 4095 | ret = sync_file_range(req->file, req->sync.off, req->sync.len, |
8ed8d3c3 JA |
4096 | req->sync.flags); |
4097 | if (ret < 0) | |
4098 | req_set_fail_links(req); | |
e1e16097 | 4099 | io_req_complete(req, ret); |
5d17b4a4 JA |
4100 | return 0; |
4101 | } | |
4102 | ||
469956e8 | 4103 | #if defined(CONFIG_NET) |
02d27d89 PB |
4104 | static int io_setup_async_msg(struct io_kiocb *req, |
4105 | struct io_async_msghdr *kmsg) | |
4106 | { | |
4107 | if (req->io) | |
4108 | return -EAGAIN; | |
4109 | if (io_alloc_async_ctx(req)) { | |
4110 | if (kmsg->iov != kmsg->fast_iov) | |
4111 | kfree(kmsg->iov); | |
4112 | return -ENOMEM; | |
4113 | } | |
4114 | req->flags |= REQ_F_NEED_CLEANUP; | |
4115 | memcpy(&req->io->msg, kmsg, sizeof(*kmsg)); | |
4116 | return -EAGAIN; | |
4117 | } | |
4118 | ||
2ae523ed PB |
4119 | static int io_sendmsg_copy_hdr(struct io_kiocb *req, |
4120 | struct io_async_msghdr *iomsg) | |
4121 | { | |
4122 | iomsg->iov = iomsg->fast_iov; | |
4123 | iomsg->msg.msg_name = &iomsg->addr; | |
4124 | return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg, | |
4125 | req->sr_msg.msg_flags, &iomsg->iov); | |
4126 | } | |
4127 | ||
3529d8c2 | 4128 | static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
03b1230c | 4129 | { |
e47293fd | 4130 | struct io_sr_msg *sr = &req->sr_msg; |
3529d8c2 | 4131 | struct io_async_ctx *io = req->io; |
99bc4c38 | 4132 | int ret; |
03b1230c | 4133 | |
d2b6f48b PB |
4134 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4135 | return -EINVAL; | |
4136 | ||
e47293fd | 4137 | sr->msg_flags = READ_ONCE(sqe->msg_flags); |
270a5940 | 4138 | sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
fddaface | 4139 | sr->len = READ_ONCE(sqe->len); |
3529d8c2 | 4140 | |
d8768362 JA |
4141 | #ifdef CONFIG_COMPAT |
4142 | if (req->ctx->compat) | |
4143 | sr->msg_flags |= MSG_CMSG_COMPAT; | |
4144 | #endif | |
4145 | ||
fddaface | 4146 | if (!io || req->opcode == IORING_OP_SEND) |
3529d8c2 | 4147 | return 0; |
5f798bea PB |
4148 | /* iovec is already imported */ |
4149 | if (req->flags & REQ_F_NEED_CLEANUP) | |
4150 | return 0; | |
3529d8c2 | 4151 | |
2ae523ed | 4152 | ret = io_sendmsg_copy_hdr(req, &io->msg); |
99bc4c38 PB |
4153 | if (!ret) |
4154 | req->flags |= REQ_F_NEED_CLEANUP; | |
4155 | return ret; | |
03b1230c JA |
4156 | } |
4157 | ||
229a7b63 JA |
4158 | static int io_sendmsg(struct io_kiocb *req, bool force_nonblock, |
4159 | struct io_comp_state *cs) | |
aa1fa28f | 4160 | { |
6b754c8b | 4161 | struct io_async_msghdr iomsg, *kmsg; |
0fa03c62 | 4162 | struct socket *sock; |
7a7cacba | 4163 | unsigned flags; |
0fa03c62 JA |
4164 | int ret; |
4165 | ||
0fa03c62 | 4166 | sock = sock_from_file(req->file, &ret); |
7a7cacba PB |
4167 | if (unlikely(!sock)) |
4168 | return ret; | |
3529d8c2 | 4169 | |
7a7cacba PB |
4170 | if (req->io) { |
4171 | kmsg = &req->io->msg; | |
4172 | kmsg->msg.msg_name = &req->io->msg.addr; | |
4173 | /* if iov is set, it's allocated already */ | |
4174 | if (!kmsg->iov) | |
4175 | kmsg->iov = kmsg->fast_iov; | |
4176 | kmsg->msg.msg_iter.iov = kmsg->iov; | |
4177 | } else { | |
4178 | ret = io_sendmsg_copy_hdr(req, &iomsg); | |
4179 | if (ret) | |
4180 | return ret; | |
4181 | kmsg = &iomsg; | |
0fa03c62 | 4182 | } |
0fa03c62 | 4183 | |
7a7cacba PB |
4184 | flags = req->sr_msg.msg_flags; |
4185 | if (flags & MSG_DONTWAIT) | |
4186 | req->flags |= REQ_F_NOWAIT; | |
4187 | else if (force_nonblock) | |
4188 | flags |= MSG_DONTWAIT; | |
e47293fd | 4189 | |
7a7cacba PB |
4190 | ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags); |
4191 | if (force_nonblock && ret == -EAGAIN) | |
4192 | return io_setup_async_msg(req, kmsg); | |
4193 | if (ret == -ERESTARTSYS) | |
4194 | ret = -EINTR; | |
0fa03c62 | 4195 | |
6b754c8b | 4196 | if (kmsg->iov != kmsg->fast_iov) |
0b416c3e | 4197 | kfree(kmsg->iov); |
99bc4c38 | 4198 | req->flags &= ~REQ_F_NEED_CLEANUP; |
4e88d6e7 JA |
4199 | if (ret < 0) |
4200 | req_set_fail_links(req); | |
229a7b63 | 4201 | __io_req_complete(req, ret, 0, cs); |
5d17b4a4 | 4202 | return 0; |
03b1230c | 4203 | } |
aa1fa28f | 4204 | |
229a7b63 JA |
4205 | static int io_send(struct io_kiocb *req, bool force_nonblock, |
4206 | struct io_comp_state *cs) | |
fddaface | 4207 | { |
7a7cacba PB |
4208 | struct io_sr_msg *sr = &req->sr_msg; |
4209 | struct msghdr msg; | |
4210 | struct iovec iov; | |
fddaface | 4211 | struct socket *sock; |
7a7cacba | 4212 | unsigned flags; |
fddaface JA |
4213 | int ret; |
4214 | ||
fddaface | 4215 | sock = sock_from_file(req->file, &ret); |
7a7cacba PB |
4216 | if (unlikely(!sock)) |
4217 | return ret; | |
fddaface | 4218 | |
7a7cacba PB |
4219 | ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter); |
4220 | if (unlikely(ret)) | |
14c32eee | 4221 | return ret;; |
fddaface | 4222 | |
7a7cacba PB |
4223 | msg.msg_name = NULL; |
4224 | msg.msg_control = NULL; | |
4225 | msg.msg_controllen = 0; | |
4226 | msg.msg_namelen = 0; | |
fddaface | 4227 | |
7a7cacba PB |
4228 | flags = req->sr_msg.msg_flags; |
4229 | if (flags & MSG_DONTWAIT) | |
4230 | req->flags |= REQ_F_NOWAIT; | |
4231 | else if (force_nonblock) | |
4232 | flags |= MSG_DONTWAIT; | |
fddaface | 4233 | |
7a7cacba PB |
4234 | msg.msg_flags = flags; |
4235 | ret = sock_sendmsg(sock, &msg); | |
4236 | if (force_nonblock && ret == -EAGAIN) | |
4237 | return -EAGAIN; | |
4238 | if (ret == -ERESTARTSYS) | |
4239 | ret = -EINTR; | |
fddaface | 4240 | |
fddaface JA |
4241 | if (ret < 0) |
4242 | req_set_fail_links(req); | |
229a7b63 | 4243 | __io_req_complete(req, ret, 0, cs); |
fddaface | 4244 | return 0; |
fddaface JA |
4245 | } |
4246 | ||
1400e697 PB |
4247 | static int __io_recvmsg_copy_hdr(struct io_kiocb *req, |
4248 | struct io_async_msghdr *iomsg) | |
52de1fe1 JA |
4249 | { |
4250 | struct io_sr_msg *sr = &req->sr_msg; | |
4251 | struct iovec __user *uiov; | |
4252 | size_t iov_len; | |
4253 | int ret; | |
4254 | ||
1400e697 PB |
4255 | ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg, |
4256 | &iomsg->uaddr, &uiov, &iov_len); | |
52de1fe1 JA |
4257 | if (ret) |
4258 | return ret; | |
4259 | ||
4260 | if (req->flags & REQ_F_BUFFER_SELECT) { | |
4261 | if (iov_len > 1) | |
4262 | return -EINVAL; | |
1400e697 | 4263 | if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov))) |
52de1fe1 | 4264 | return -EFAULT; |
1400e697 PB |
4265 | sr->len = iomsg->iov[0].iov_len; |
4266 | iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1, | |
52de1fe1 | 4267 | sr->len); |
1400e697 | 4268 | iomsg->iov = NULL; |
52de1fe1 JA |
4269 | } else { |
4270 | ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV, | |
1400e697 | 4271 | &iomsg->iov, &iomsg->msg.msg_iter); |
52de1fe1 JA |
4272 | if (ret > 0) |
4273 | ret = 0; | |
4274 | } | |
4275 | ||
4276 | return ret; | |
4277 | } | |
4278 | ||
4279 | #ifdef CONFIG_COMPAT | |
4280 | static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req, | |
1400e697 | 4281 | struct io_async_msghdr *iomsg) |
52de1fe1 JA |
4282 | { |
4283 | struct compat_msghdr __user *msg_compat; | |
4284 | struct io_sr_msg *sr = &req->sr_msg; | |
4285 | struct compat_iovec __user *uiov; | |
4286 | compat_uptr_t ptr; | |
4287 | compat_size_t len; | |
4288 | int ret; | |
4289 | ||
270a5940 | 4290 | msg_compat = (struct compat_msghdr __user *) sr->umsg; |
1400e697 | 4291 | ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr, |
52de1fe1 JA |
4292 | &ptr, &len); |
4293 | if (ret) | |
4294 | return ret; | |
4295 | ||
4296 | uiov = compat_ptr(ptr); | |
4297 | if (req->flags & REQ_F_BUFFER_SELECT) { | |
4298 | compat_ssize_t clen; | |
4299 | ||
4300 | if (len > 1) | |
4301 | return -EINVAL; | |
4302 | if (!access_ok(uiov, sizeof(*uiov))) | |
4303 | return -EFAULT; | |
4304 | if (__get_user(clen, &uiov->iov_len)) | |
4305 | return -EFAULT; | |
4306 | if (clen < 0) | |
4307 | return -EINVAL; | |
1400e697 PB |
4308 | sr->len = iomsg->iov[0].iov_len; |
4309 | iomsg->iov = NULL; | |
52de1fe1 JA |
4310 | } else { |
4311 | ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV, | |
1400e697 PB |
4312 | &iomsg->iov, |
4313 | &iomsg->msg.msg_iter); | |
52de1fe1 JA |
4314 | if (ret < 0) |
4315 | return ret; | |
4316 | } | |
4317 | ||
4318 | return 0; | |
4319 | } | |
4320 | #endif | |
4321 | ||
1400e697 PB |
4322 | static int io_recvmsg_copy_hdr(struct io_kiocb *req, |
4323 | struct io_async_msghdr *iomsg) | |
52de1fe1 | 4324 | { |
1400e697 PB |
4325 | iomsg->msg.msg_name = &iomsg->addr; |
4326 | iomsg->iov = iomsg->fast_iov; | |
52de1fe1 JA |
4327 | |
4328 | #ifdef CONFIG_COMPAT | |
4329 | if (req->ctx->compat) | |
1400e697 | 4330 | return __io_compat_recvmsg_copy_hdr(req, iomsg); |
fddaface | 4331 | #endif |
52de1fe1 | 4332 | |
1400e697 | 4333 | return __io_recvmsg_copy_hdr(req, iomsg); |
52de1fe1 JA |
4334 | } |
4335 | ||
bcda7baa | 4336 | static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req, |
7fbb1b54 | 4337 | bool needs_lock) |
bcda7baa JA |
4338 | { |
4339 | struct io_sr_msg *sr = &req->sr_msg; | |
4340 | struct io_buffer *kbuf; | |
4341 | ||
bcda7baa JA |
4342 | kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock); |
4343 | if (IS_ERR(kbuf)) | |
4344 | return kbuf; | |
4345 | ||
4346 | sr->kbuf = kbuf; | |
4347 | req->flags |= REQ_F_BUFFER_SELECTED; | |
bcda7baa | 4348 | return kbuf; |
fddaface JA |
4349 | } |
4350 | ||
7fbb1b54 PB |
4351 | static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req) |
4352 | { | |
4353 | return io_put_kbuf(req, req->sr_msg.kbuf); | |
4354 | } | |
4355 | ||
3529d8c2 JA |
4356 | static int io_recvmsg_prep(struct io_kiocb *req, |
4357 | const struct io_uring_sqe *sqe) | |
aa1fa28f | 4358 | { |
e47293fd | 4359 | struct io_sr_msg *sr = &req->sr_msg; |
3529d8c2 | 4360 | struct io_async_ctx *io = req->io; |
99bc4c38 | 4361 | int ret; |
3529d8c2 | 4362 | |
d2b6f48b PB |
4363 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4364 | return -EINVAL; | |
4365 | ||
3529d8c2 | 4366 | sr->msg_flags = READ_ONCE(sqe->msg_flags); |
270a5940 | 4367 | sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
0b7b21e4 | 4368 | sr->len = READ_ONCE(sqe->len); |
bcda7baa | 4369 | sr->bgid = READ_ONCE(sqe->buf_group); |
06b76d44 | 4370 | |
d8768362 JA |
4371 | #ifdef CONFIG_COMPAT |
4372 | if (req->ctx->compat) | |
4373 | sr->msg_flags |= MSG_CMSG_COMPAT; | |
4374 | #endif | |
4375 | ||
fddaface | 4376 | if (!io || req->opcode == IORING_OP_RECV) |
06b76d44 | 4377 | return 0; |
5f798bea PB |
4378 | /* iovec is already imported */ |
4379 | if (req->flags & REQ_F_NEED_CLEANUP) | |
4380 | return 0; | |
03b1230c | 4381 | |
1400e697 | 4382 | ret = io_recvmsg_copy_hdr(req, &io->msg); |
99bc4c38 PB |
4383 | if (!ret) |
4384 | req->flags |= REQ_F_NEED_CLEANUP; | |
4385 | return ret; | |
aa1fa28f JA |
4386 | } |
4387 | ||
229a7b63 JA |
4388 | static int io_recvmsg(struct io_kiocb *req, bool force_nonblock, |
4389 | struct io_comp_state *cs) | |
aa1fa28f | 4390 | { |
6b754c8b | 4391 | struct io_async_msghdr iomsg, *kmsg; |
03b1230c | 4392 | struct socket *sock; |
7fbb1b54 | 4393 | struct io_buffer *kbuf; |
7a7cacba | 4394 | unsigned flags; |
52de1fe1 | 4395 | int ret, cflags = 0; |
03b1230c | 4396 | |
03b1230c | 4397 | sock = sock_from_file(req->file, &ret); |
7a7cacba PB |
4398 | if (unlikely(!sock)) |
4399 | return ret; | |
3529d8c2 | 4400 | |
7a7cacba PB |
4401 | if (req->io) { |
4402 | kmsg = &req->io->msg; | |
4403 | kmsg->msg.msg_name = &req->io->msg.addr; | |
4404 | /* if iov is set, it's allocated already */ | |
4405 | if (!kmsg->iov) | |
4406 | kmsg->iov = kmsg->fast_iov; | |
4407 | kmsg->msg.msg_iter.iov = kmsg->iov; | |
4408 | } else { | |
4409 | ret = io_recvmsg_copy_hdr(req, &iomsg); | |
4410 | if (ret) | |
681fda8d | 4411 | return ret; |
7a7cacba PB |
4412 | kmsg = &iomsg; |
4413 | } | |
03b1230c | 4414 | |
bc02ef33 | 4415 | if (req->flags & REQ_F_BUFFER_SELECT) { |
7fbb1b54 | 4416 | kbuf = io_recv_buffer_select(req, !force_nonblock); |
bc02ef33 | 4417 | if (IS_ERR(kbuf)) |
52de1fe1 | 4418 | return PTR_ERR(kbuf); |
7a7cacba PB |
4419 | kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr); |
4420 | iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov, | |
4421 | 1, req->sr_msg.len); | |
4422 | } | |
52de1fe1 | 4423 | |
7a7cacba PB |
4424 | flags = req->sr_msg.msg_flags; |
4425 | if (flags & MSG_DONTWAIT) | |
4426 | req->flags |= REQ_F_NOWAIT; | |
4427 | else if (force_nonblock) | |
4428 | flags |= MSG_DONTWAIT; | |
e47293fd | 4429 | |
7a7cacba PB |
4430 | ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg, |
4431 | kmsg->uaddr, flags); | |
0e1b6fe3 PB |
4432 | if (force_nonblock && ret == -EAGAIN) |
4433 | return io_setup_async_msg(req, kmsg); | |
7a7cacba PB |
4434 | if (ret == -ERESTARTSYS) |
4435 | ret = -EINTR; | |
03b1230c | 4436 | |
7fbb1b54 PB |
4437 | if (req->flags & REQ_F_BUFFER_SELECTED) |
4438 | cflags = io_put_recv_kbuf(req); | |
6b754c8b | 4439 | if (kmsg->iov != kmsg->fast_iov) |
0b416c3e | 4440 | kfree(kmsg->iov); |
99bc4c38 | 4441 | req->flags &= ~REQ_F_NEED_CLEANUP; |
4e88d6e7 JA |
4442 | if (ret < 0) |
4443 | req_set_fail_links(req); | |
229a7b63 | 4444 | __io_req_complete(req, ret, cflags, cs); |
03b1230c | 4445 | return 0; |
0fa03c62 | 4446 | } |
5d17b4a4 | 4447 | |
229a7b63 JA |
4448 | static int io_recv(struct io_kiocb *req, bool force_nonblock, |
4449 | struct io_comp_state *cs) | |
fddaface | 4450 | { |
6b754c8b | 4451 | struct io_buffer *kbuf; |
7a7cacba PB |
4452 | struct io_sr_msg *sr = &req->sr_msg; |
4453 | struct msghdr msg; | |
4454 | void __user *buf = sr->buf; | |
fddaface | 4455 | struct socket *sock; |
7a7cacba PB |
4456 | struct iovec iov; |
4457 | unsigned flags; | |
bcda7baa | 4458 | int ret, cflags = 0; |
fddaface | 4459 | |
fddaface | 4460 | sock = sock_from_file(req->file, &ret); |
7a7cacba PB |
4461 | if (unlikely(!sock)) |
4462 | return ret; | |
fddaface | 4463 | |
bc02ef33 | 4464 | if (req->flags & REQ_F_BUFFER_SELECT) { |
7fbb1b54 | 4465 | kbuf = io_recv_buffer_select(req, !force_nonblock); |
bcda7baa JA |
4466 | if (IS_ERR(kbuf)) |
4467 | return PTR_ERR(kbuf); | |
7a7cacba | 4468 | buf = u64_to_user_ptr(kbuf->addr); |
bc02ef33 | 4469 | } |
bcda7baa | 4470 | |
7a7cacba | 4471 | ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter); |
14c32eee PB |
4472 | if (unlikely(ret)) |
4473 | goto out_free; | |
fddaface | 4474 | |
7a7cacba PB |
4475 | msg.msg_name = NULL; |
4476 | msg.msg_control = NULL; | |
4477 | msg.msg_controllen = 0; | |
4478 | msg.msg_namelen = 0; | |
4479 | msg.msg_iocb = NULL; | |
4480 | msg.msg_flags = 0; | |
fddaface | 4481 | |
7a7cacba PB |
4482 | flags = req->sr_msg.msg_flags; |
4483 | if (flags & MSG_DONTWAIT) | |
4484 | req->flags |= REQ_F_NOWAIT; | |
4485 | else if (force_nonblock) | |
4486 | flags |= MSG_DONTWAIT; | |
4487 | ||
4488 | ret = sock_recvmsg(sock, &msg, flags); | |
4489 | if (force_nonblock && ret == -EAGAIN) | |
4490 | return -EAGAIN; | |
4491 | if (ret == -ERESTARTSYS) | |
4492 | ret = -EINTR; | |
14c32eee | 4493 | out_free: |
7fbb1b54 PB |
4494 | if (req->flags & REQ_F_BUFFER_SELECTED) |
4495 | cflags = io_put_recv_kbuf(req); | |
fddaface JA |
4496 | if (ret < 0) |
4497 | req_set_fail_links(req); | |
229a7b63 | 4498 | __io_req_complete(req, ret, cflags, cs); |
fddaface | 4499 | return 0; |
fddaface JA |
4500 | } |
4501 | ||
3529d8c2 | 4502 | static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
17f2fe35 | 4503 | { |
8ed8d3c3 JA |
4504 | struct io_accept *accept = &req->accept; |
4505 | ||
17f2fe35 JA |
4506 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) |
4507 | return -EINVAL; | |
8042d6ce | 4508 | if (sqe->ioprio || sqe->len || sqe->buf_index) |
17f2fe35 JA |
4509 | return -EINVAL; |
4510 | ||
d55e5f5b JA |
4511 | accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
4512 | accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2)); | |
8ed8d3c3 | 4513 | accept->flags = READ_ONCE(sqe->accept_flags); |
09952e3e | 4514 | accept->nofile = rlimit(RLIMIT_NOFILE); |
8ed8d3c3 | 4515 | return 0; |
8ed8d3c3 | 4516 | } |
17f2fe35 | 4517 | |
229a7b63 JA |
4518 | static int io_accept(struct io_kiocb *req, bool force_nonblock, |
4519 | struct io_comp_state *cs) | |
8ed8d3c3 JA |
4520 | { |
4521 | struct io_accept *accept = &req->accept; | |
ac45abc0 | 4522 | unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0; |
8ed8d3c3 JA |
4523 | int ret; |
4524 | ||
e697deed JX |
4525 | if (req->file->f_flags & O_NONBLOCK) |
4526 | req->flags |= REQ_F_NOWAIT; | |
4527 | ||
8ed8d3c3 | 4528 | ret = __sys_accept4_file(req->file, file_flags, accept->addr, |
09952e3e JA |
4529 | accept->addr_len, accept->flags, |
4530 | accept->nofile); | |
8ed8d3c3 | 4531 | if (ret == -EAGAIN && force_nonblock) |
17f2fe35 | 4532 | return -EAGAIN; |
ac45abc0 PB |
4533 | if (ret < 0) { |
4534 | if (ret == -ERESTARTSYS) | |
4535 | ret = -EINTR; | |
4e88d6e7 | 4536 | req_set_fail_links(req); |
ac45abc0 | 4537 | } |
229a7b63 | 4538 | __io_req_complete(req, ret, 0, cs); |
17f2fe35 | 4539 | return 0; |
8ed8d3c3 JA |
4540 | } |
4541 | ||
3529d8c2 | 4542 | static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f499a021 | 4543 | { |
3529d8c2 JA |
4544 | struct io_connect *conn = &req->connect; |
4545 | struct io_async_ctx *io = req->io; | |
f499a021 | 4546 | |
3fbb51c1 JA |
4547 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) |
4548 | return -EINVAL; | |
4549 | if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags) | |
4550 | return -EINVAL; | |
4551 | ||
3529d8c2 JA |
4552 | conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
4553 | conn->addr_len = READ_ONCE(sqe->addr2); | |
4554 | ||
4555 | if (!io) | |
4556 | return 0; | |
4557 | ||
4558 | return move_addr_to_kernel(conn->addr, conn->addr_len, | |
3fbb51c1 | 4559 | &io->connect.address); |
f499a021 JA |
4560 | } |
4561 | ||
229a7b63 JA |
4562 | static int io_connect(struct io_kiocb *req, bool force_nonblock, |
4563 | struct io_comp_state *cs) | |
f8e85cf2 | 4564 | { |
f499a021 | 4565 | struct io_async_ctx __io, *io; |
f8e85cf2 | 4566 | unsigned file_flags; |
3fbb51c1 | 4567 | int ret; |
f8e85cf2 | 4568 | |
f499a021 JA |
4569 | if (req->io) { |
4570 | io = req->io; | |
4571 | } else { | |
3529d8c2 JA |
4572 | ret = move_addr_to_kernel(req->connect.addr, |
4573 | req->connect.addr_len, | |
4574 | &__io.connect.address); | |
f499a021 JA |
4575 | if (ret) |
4576 | goto out; | |
4577 | io = &__io; | |
4578 | } | |
4579 | ||
3fbb51c1 JA |
4580 | file_flags = force_nonblock ? O_NONBLOCK : 0; |
4581 | ||
4582 | ret = __sys_connect_file(req->file, &io->connect.address, | |
4583 | req->connect.addr_len, file_flags); | |
87f80d62 | 4584 | if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) { |
b7bb4f7d JA |
4585 | if (req->io) |
4586 | return -EAGAIN; | |
4587 | if (io_alloc_async_ctx(req)) { | |
f499a021 JA |
4588 | ret = -ENOMEM; |
4589 | goto out; | |
4590 | } | |
b7bb4f7d | 4591 | memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect)); |
f8e85cf2 | 4592 | return -EAGAIN; |
f499a021 | 4593 | } |
f8e85cf2 JA |
4594 | if (ret == -ERESTARTSYS) |
4595 | ret = -EINTR; | |
f499a021 | 4596 | out: |
4e88d6e7 JA |
4597 | if (ret < 0) |
4598 | req_set_fail_links(req); | |
229a7b63 | 4599 | __io_req_complete(req, ret, 0, cs); |
f8e85cf2 | 4600 | return 0; |
469956e8 Y |
4601 | } |
4602 | #else /* !CONFIG_NET */ | |
4603 | static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
4604 | { | |
f8e85cf2 | 4605 | return -EOPNOTSUPP; |
f8e85cf2 JA |
4606 | } |
4607 | ||
1e16c2f9 RD |
4608 | static int io_sendmsg(struct io_kiocb *req, bool force_nonblock, |
4609 | struct io_comp_state *cs) | |
469956e8 Y |
4610 | { |
4611 | return -EOPNOTSUPP; | |
4612 | } | |
4613 | ||
1e16c2f9 RD |
4614 | static int io_send(struct io_kiocb *req, bool force_nonblock, |
4615 | struct io_comp_state *cs) | |
469956e8 Y |
4616 | { |
4617 | return -EOPNOTSUPP; | |
4618 | } | |
4619 | ||
4620 | static int io_recvmsg_prep(struct io_kiocb *req, | |
4621 | const struct io_uring_sqe *sqe) | |
4622 | { | |
4623 | return -EOPNOTSUPP; | |
4624 | } | |
4625 | ||
1e16c2f9 RD |
4626 | static int io_recvmsg(struct io_kiocb *req, bool force_nonblock, |
4627 | struct io_comp_state *cs) | |
469956e8 Y |
4628 | { |
4629 | return -EOPNOTSUPP; | |
4630 | } | |
4631 | ||
1e16c2f9 RD |
4632 | static int io_recv(struct io_kiocb *req, bool force_nonblock, |
4633 | struct io_comp_state *cs) | |
469956e8 Y |
4634 | { |
4635 | return -EOPNOTSUPP; | |
4636 | } | |
4637 | ||
4638 | static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
4639 | { | |
4640 | return -EOPNOTSUPP; | |
4641 | } | |
4642 | ||
1e16c2f9 RD |
4643 | static int io_accept(struct io_kiocb *req, bool force_nonblock, |
4644 | struct io_comp_state *cs) | |
469956e8 Y |
4645 | { |
4646 | return -EOPNOTSUPP; | |
4647 | } | |
ce593a6c | 4648 | |
469956e8 Y |
4649 | static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4650 | { | |
4651 | return -EOPNOTSUPP; | |
4652 | } | |
4653 | ||
1e16c2f9 RD |
4654 | static int io_connect(struct io_kiocb *req, bool force_nonblock, |
4655 | struct io_comp_state *cs) | |
469956e8 | 4656 | { |
f8e85cf2 | 4657 | return -EOPNOTSUPP; |
ce593a6c | 4658 | } |
469956e8 | 4659 | #endif /* CONFIG_NET */ |
f8e85cf2 | 4660 | |
d7718a9d JA |
4661 | struct io_poll_table { |
4662 | struct poll_table_struct pt; | |
4663 | struct io_kiocb *req; | |
4664 | int error; | |
4665 | }; | |
ce593a6c | 4666 | |
d7718a9d JA |
4667 | static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll, |
4668 | __poll_t mask, task_work_func_t func) | |
4669 | { | |
fd7d6de2 | 4670 | bool twa_signal_ok; |
aa96bf8a | 4671 | int ret; |
d7718a9d JA |
4672 | |
4673 | /* for instances that support it check for an event match first: */ | |
4674 | if (mask && !(mask & poll->events)) | |
4675 | return 0; | |
4676 | ||
4677 | trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask); | |
4678 | ||
4679 | list_del_init(&poll->wait.entry); | |
4680 | ||
d7718a9d JA |
4681 | req->result = mask; |
4682 | init_task_work(&req->task_work, func); | |
6d816e08 JA |
4683 | percpu_ref_get(&req->ctx->refs); |
4684 | ||
fd7d6de2 JA |
4685 | /* |
4686 | * If we using the signalfd wait_queue_head for this wakeup, then | |
4687 | * it's not safe to use TWA_SIGNAL as we could be recursing on the | |
4688 | * tsk->sighand->siglock on doing the wakeup. Should not be needed | |
4689 | * either, as the normal wakeup will suffice. | |
4690 | */ | |
4691 | twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh); | |
4692 | ||
d7718a9d | 4693 | /* |
e3aabf95 JA |
4694 | * If this fails, then the task is exiting. When a task exits, the |
4695 | * work gets canceled, so just cancel this request as well instead | |
4696 | * of executing it. We can't safely execute it anyway, as we may not | |
4697 | * have the needed state needed for it anyway. | |
d7718a9d | 4698 | */ |
fd7d6de2 | 4699 | ret = io_req_task_work_add(req, &req->task_work, twa_signal_ok); |
aa96bf8a | 4700 | if (unlikely(ret)) { |
c2c4c83c JA |
4701 | struct task_struct *tsk; |
4702 | ||
e3aabf95 | 4703 | WRITE_ONCE(poll->canceled, true); |
aa96bf8a | 4704 | tsk = io_wq_get_task(req->ctx->io_wq); |
ce593a6c JA |
4705 | task_work_add(tsk, &req->task_work, 0); |
4706 | wake_up_process(tsk); | |
aa96bf8a | 4707 | } |
d7718a9d JA |
4708 | return 1; |
4709 | } | |
4710 | ||
74ce6ce4 JA |
4711 | static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll) |
4712 | __acquires(&req->ctx->completion_lock) | |
4713 | { | |
4714 | struct io_ring_ctx *ctx = req->ctx; | |
4715 | ||
4716 | if (!req->result && !READ_ONCE(poll->canceled)) { | |
4717 | struct poll_table_struct pt = { ._key = poll->events }; | |
4718 | ||
4719 | req->result = vfs_poll(req->file, &pt) & poll->events; | |
4720 | } | |
4721 | ||
4722 | spin_lock_irq(&ctx->completion_lock); | |
4723 | if (!req->result && !READ_ONCE(poll->canceled)) { | |
4724 | add_wait_queue(poll->head, &poll->wait); | |
4725 | return true; | |
4726 | } | |
4727 | ||
4728 | return false; | |
4729 | } | |
4730 | ||
d4e7cd36 | 4731 | static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req) |
18bceab1 | 4732 | { |
d4e7cd36 JA |
4733 | /* pure poll stashes this in ->io, poll driven retry elsewhere */ |
4734 | if (req->opcode == IORING_OP_POLL_ADD) | |
4735 | return (struct io_poll_iocb *) req->io; | |
4736 | return req->apoll->double_poll; | |
4737 | } | |
4738 | ||
4739 | static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req) | |
4740 | { | |
4741 | if (req->opcode == IORING_OP_POLL_ADD) | |
4742 | return &req->poll; | |
4743 | return &req->apoll->poll; | |
4744 | } | |
4745 | ||
4746 | static void io_poll_remove_double(struct io_kiocb *req) | |
4747 | { | |
4748 | struct io_poll_iocb *poll = io_poll_get_double(req); | |
18bceab1 JA |
4749 | |
4750 | lockdep_assert_held(&req->ctx->completion_lock); | |
4751 | ||
4752 | if (poll && poll->head) { | |
4753 | struct wait_queue_head *head = poll->head; | |
4754 | ||
4755 | spin_lock(&head->lock); | |
4756 | list_del_init(&poll->wait.entry); | |
4757 | if (poll->wait.private) | |
4758 | refcount_dec(&req->refs); | |
4759 | poll->head = NULL; | |
4760 | spin_unlock(&head->lock); | |
4761 | } | |
4762 | } | |
4763 | ||
4764 | static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error) | |
4765 | { | |
4766 | struct io_ring_ctx *ctx = req->ctx; | |
4767 | ||
d4e7cd36 | 4768 | io_poll_remove_double(req); |
18bceab1 JA |
4769 | req->poll.done = true; |
4770 | io_cqring_fill_event(req, error ? error : mangle_poll(mask)); | |
4771 | io_commit_cqring(ctx); | |
4772 | } | |
4773 | ||
4774 | static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt) | |
4775 | { | |
4776 | struct io_ring_ctx *ctx = req->ctx; | |
4777 | ||
4778 | if (io_poll_rewait(req, &req->poll)) { | |
4779 | spin_unlock_irq(&ctx->completion_lock); | |
4780 | return; | |
4781 | } | |
4782 | ||
4783 | hash_del(&req->hash_node); | |
4784 | io_poll_complete(req, req->result, 0); | |
4785 | req->flags |= REQ_F_COMP_LOCKED; | |
9b5f7bd9 | 4786 | *nxt = io_put_req_find_next(req); |
18bceab1 JA |
4787 | spin_unlock_irq(&ctx->completion_lock); |
4788 | ||
4789 | io_cqring_ev_posted(ctx); | |
4790 | } | |
4791 | ||
4792 | static void io_poll_task_func(struct callback_head *cb) | |
4793 | { | |
4794 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); | |
6d816e08 | 4795 | struct io_ring_ctx *ctx = req->ctx; |
18bceab1 JA |
4796 | struct io_kiocb *nxt = NULL; |
4797 | ||
4798 | io_poll_task_handler(req, &nxt); | |
ea1164e5 PB |
4799 | if (nxt) |
4800 | __io_req_task_submit(nxt); | |
6d816e08 | 4801 | percpu_ref_put(&ctx->refs); |
18bceab1 JA |
4802 | } |
4803 | ||
4804 | static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode, | |
4805 | int sync, void *key) | |
4806 | { | |
4807 | struct io_kiocb *req = wait->private; | |
d4e7cd36 | 4808 | struct io_poll_iocb *poll = io_poll_get_single(req); |
18bceab1 JA |
4809 | __poll_t mask = key_to_poll(key); |
4810 | ||
4811 | /* for instances that support it check for an event match first: */ | |
4812 | if (mask && !(mask & poll->events)) | |
4813 | return 0; | |
4814 | ||
8706e04e JA |
4815 | list_del_init(&wait->entry); |
4816 | ||
807abcb0 | 4817 | if (poll && poll->head) { |
18bceab1 JA |
4818 | bool done; |
4819 | ||
807abcb0 JA |
4820 | spin_lock(&poll->head->lock); |
4821 | done = list_empty(&poll->wait.entry); | |
18bceab1 | 4822 | if (!done) |
807abcb0 | 4823 | list_del_init(&poll->wait.entry); |
d4e7cd36 JA |
4824 | /* make sure double remove sees this as being gone */ |
4825 | wait->private = NULL; | |
807abcb0 | 4826 | spin_unlock(&poll->head->lock); |
18bceab1 JA |
4827 | if (!done) |
4828 | __io_async_wake(req, poll, mask, io_poll_task_func); | |
4829 | } | |
4830 | refcount_dec(&req->refs); | |
4831 | return 1; | |
4832 | } | |
4833 | ||
4834 | static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events, | |
4835 | wait_queue_func_t wake_func) | |
4836 | { | |
4837 | poll->head = NULL; | |
4838 | poll->done = false; | |
4839 | poll->canceled = false; | |
4840 | poll->events = events; | |
4841 | INIT_LIST_HEAD(&poll->wait.entry); | |
4842 | init_waitqueue_func_entry(&poll->wait, wake_func); | |
4843 | } | |
4844 | ||
4845 | static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt, | |
807abcb0 JA |
4846 | struct wait_queue_head *head, |
4847 | struct io_poll_iocb **poll_ptr) | |
18bceab1 JA |
4848 | { |
4849 | struct io_kiocb *req = pt->req; | |
4850 | ||
4851 | /* | |
4852 | * If poll->head is already set, it's because the file being polled | |
4853 | * uses multiple waitqueues for poll handling (eg one for read, one | |
4854 | * for write). Setup a separate io_poll_iocb if this happens. | |
4855 | */ | |
4856 | if (unlikely(poll->head)) { | |
4857 | /* already have a 2nd entry, fail a third attempt */ | |
807abcb0 | 4858 | if (*poll_ptr) { |
18bceab1 JA |
4859 | pt->error = -EINVAL; |
4860 | return; | |
4861 | } | |
4862 | poll = kmalloc(sizeof(*poll), GFP_ATOMIC); | |
4863 | if (!poll) { | |
4864 | pt->error = -ENOMEM; | |
4865 | return; | |
4866 | } | |
4867 | io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake); | |
4868 | refcount_inc(&req->refs); | |
4869 | poll->wait.private = req; | |
807abcb0 | 4870 | *poll_ptr = poll; |
18bceab1 JA |
4871 | } |
4872 | ||
4873 | pt->error = 0; | |
4874 | poll->head = head; | |
a31eb4a2 JX |
4875 | |
4876 | if (poll->events & EPOLLEXCLUSIVE) | |
4877 | add_wait_queue_exclusive(head, &poll->wait); | |
4878 | else | |
4879 | add_wait_queue(head, &poll->wait); | |
18bceab1 JA |
4880 | } |
4881 | ||
4882 | static void io_async_queue_proc(struct file *file, struct wait_queue_head *head, | |
4883 | struct poll_table_struct *p) | |
4884 | { | |
4885 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
807abcb0 | 4886 | struct async_poll *apoll = pt->req->apoll; |
18bceab1 | 4887 | |
807abcb0 | 4888 | __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll); |
18bceab1 JA |
4889 | } |
4890 | ||
d7718a9d JA |
4891 | static void io_async_task_func(struct callback_head *cb) |
4892 | { | |
4893 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); | |
4894 | struct async_poll *apoll = req->apoll; | |
4895 | struct io_ring_ctx *ctx = req->ctx; | |
4896 | ||
4897 | trace_io_uring_task_run(req->ctx, req->opcode, req->user_data); | |
4898 | ||
74ce6ce4 | 4899 | if (io_poll_rewait(req, &apoll->poll)) { |
d7718a9d | 4900 | spin_unlock_irq(&ctx->completion_lock); |
6d816e08 | 4901 | percpu_ref_put(&ctx->refs); |
74ce6ce4 | 4902 | return; |
d7718a9d JA |
4903 | } |
4904 | ||
31067255 | 4905 | /* If req is still hashed, it cannot have been canceled. Don't check. */ |
0be0b0e3 | 4906 | if (hash_hashed(&req->hash_node)) |
74ce6ce4 | 4907 | hash_del(&req->hash_node); |
2bae047e | 4908 | |
d4e7cd36 | 4909 | io_poll_remove_double(req); |
74ce6ce4 JA |
4910 | spin_unlock_irq(&ctx->completion_lock); |
4911 | ||
0be0b0e3 PB |
4912 | if (!READ_ONCE(apoll->poll.canceled)) |
4913 | __io_req_task_submit(req); | |
4914 | else | |
4915 | __io_req_task_cancel(req, -ECANCELED); | |
aa340845 | 4916 | |
6d816e08 | 4917 | percpu_ref_put(&ctx->refs); |
807abcb0 | 4918 | kfree(apoll->double_poll); |
31067255 | 4919 | kfree(apoll); |
d7718a9d JA |
4920 | } |
4921 | ||
4922 | static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
4923 | void *key) | |
4924 | { | |
4925 | struct io_kiocb *req = wait->private; | |
4926 | struct io_poll_iocb *poll = &req->apoll->poll; | |
4927 | ||
4928 | trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data, | |
4929 | key_to_poll(key)); | |
4930 | ||
4931 | return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func); | |
4932 | } | |
4933 | ||
4934 | static void io_poll_req_insert(struct io_kiocb *req) | |
4935 | { | |
4936 | struct io_ring_ctx *ctx = req->ctx; | |
4937 | struct hlist_head *list; | |
4938 | ||
4939 | list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)]; | |
4940 | hlist_add_head(&req->hash_node, list); | |
4941 | } | |
4942 | ||
4943 | static __poll_t __io_arm_poll_handler(struct io_kiocb *req, | |
4944 | struct io_poll_iocb *poll, | |
4945 | struct io_poll_table *ipt, __poll_t mask, | |
4946 | wait_queue_func_t wake_func) | |
4947 | __acquires(&ctx->completion_lock) | |
4948 | { | |
4949 | struct io_ring_ctx *ctx = req->ctx; | |
4950 | bool cancel = false; | |
4951 | ||
18bceab1 | 4952 | io_init_poll_iocb(poll, mask, wake_func); |
b90cd197 | 4953 | poll->file = req->file; |
18bceab1 | 4954 | poll->wait.private = req; |
d7718a9d JA |
4955 | |
4956 | ipt->pt._key = mask; | |
4957 | ipt->req = req; | |
4958 | ipt->error = -EINVAL; | |
4959 | ||
d7718a9d JA |
4960 | mask = vfs_poll(req->file, &ipt->pt) & poll->events; |
4961 | ||
4962 | spin_lock_irq(&ctx->completion_lock); | |
4963 | if (likely(poll->head)) { | |
4964 | spin_lock(&poll->head->lock); | |
4965 | if (unlikely(list_empty(&poll->wait.entry))) { | |
4966 | if (ipt->error) | |
4967 | cancel = true; | |
4968 | ipt->error = 0; | |
4969 | mask = 0; | |
4970 | } | |
4971 | if (mask || ipt->error) | |
4972 | list_del_init(&poll->wait.entry); | |
4973 | else if (cancel) | |
4974 | WRITE_ONCE(poll->canceled, true); | |
4975 | else if (!poll->done) /* actually waiting for an event */ | |
4976 | io_poll_req_insert(req); | |
4977 | spin_unlock(&poll->head->lock); | |
4978 | } | |
4979 | ||
4980 | return mask; | |
4981 | } | |
4982 | ||
4983 | static bool io_arm_poll_handler(struct io_kiocb *req) | |
4984 | { | |
4985 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
4986 | struct io_ring_ctx *ctx = req->ctx; | |
4987 | struct async_poll *apoll; | |
4988 | struct io_poll_table ipt; | |
4989 | __poll_t mask, ret; | |
9dab14b8 | 4990 | int rw; |
d7718a9d JA |
4991 | |
4992 | if (!req->file || !file_can_poll(req->file)) | |
4993 | return false; | |
24c74678 | 4994 | if (req->flags & REQ_F_POLLED) |
d7718a9d | 4995 | return false; |
9dab14b8 JA |
4996 | if (def->pollin) |
4997 | rw = READ; | |
4998 | else if (def->pollout) | |
4999 | rw = WRITE; | |
5000 | else | |
5001 | return false; | |
5002 | /* if we can't nonblock try, then no point in arming a poll handler */ | |
5003 | if (!io_file_supports_async(req->file, rw)) | |
d7718a9d JA |
5004 | return false; |
5005 | ||
5006 | apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC); | |
5007 | if (unlikely(!apoll)) | |
5008 | return false; | |
807abcb0 | 5009 | apoll->double_poll = NULL; |
d7718a9d JA |
5010 | |
5011 | req->flags |= REQ_F_POLLED; | |
d7718a9d JA |
5012 | req->apoll = apoll; |
5013 | INIT_HLIST_NODE(&req->hash_node); | |
5014 | ||
8755d97a | 5015 | mask = 0; |
d7718a9d | 5016 | if (def->pollin) |
8755d97a | 5017 | mask |= POLLIN | POLLRDNORM; |
d7718a9d JA |
5018 | if (def->pollout) |
5019 | mask |= POLLOUT | POLLWRNORM; | |
5020 | mask |= POLLERR | POLLPRI; | |
5021 | ||
5022 | ipt.pt._qproc = io_async_queue_proc; | |
5023 | ||
5024 | ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, | |
5025 | io_async_wake); | |
a36da65c | 5026 | if (ret || ipt.error) { |
d4e7cd36 | 5027 | io_poll_remove_double(req); |
d7718a9d | 5028 | spin_unlock_irq(&ctx->completion_lock); |
807abcb0 | 5029 | kfree(apoll->double_poll); |
d7718a9d JA |
5030 | kfree(apoll); |
5031 | return false; | |
5032 | } | |
5033 | spin_unlock_irq(&ctx->completion_lock); | |
5034 | trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask, | |
5035 | apoll->poll.events); | |
5036 | return true; | |
5037 | } | |
5038 | ||
5039 | static bool __io_poll_remove_one(struct io_kiocb *req, | |
5040 | struct io_poll_iocb *poll) | |
221c5eb2 | 5041 | { |
b41e9852 | 5042 | bool do_complete = false; |
221c5eb2 JA |
5043 | |
5044 | spin_lock(&poll->head->lock); | |
5045 | WRITE_ONCE(poll->canceled, true); | |
392edb45 JA |
5046 | if (!list_empty(&poll->wait.entry)) { |
5047 | list_del_init(&poll->wait.entry); | |
b41e9852 | 5048 | do_complete = true; |
221c5eb2 JA |
5049 | } |
5050 | spin_unlock(&poll->head->lock); | |
3bfa5bcb | 5051 | hash_del(&req->hash_node); |
d7718a9d JA |
5052 | return do_complete; |
5053 | } | |
5054 | ||
5055 | static bool io_poll_remove_one(struct io_kiocb *req) | |
5056 | { | |
5057 | bool do_complete; | |
5058 | ||
d4e7cd36 JA |
5059 | io_poll_remove_double(req); |
5060 | ||
d7718a9d JA |
5061 | if (req->opcode == IORING_OP_POLL_ADD) { |
5062 | do_complete = __io_poll_remove_one(req, &req->poll); | |
5063 | } else { | |
3bfa5bcb JA |
5064 | struct async_poll *apoll = req->apoll; |
5065 | ||
d7718a9d | 5066 | /* non-poll requests have submit ref still */ |
3bfa5bcb JA |
5067 | do_complete = __io_poll_remove_one(req, &apoll->poll); |
5068 | if (do_complete) { | |
d7718a9d | 5069 | io_put_req(req); |
807abcb0 | 5070 | kfree(apoll->double_poll); |
3bfa5bcb JA |
5071 | kfree(apoll); |
5072 | } | |
b1f573bd XW |
5073 | } |
5074 | ||
b41e9852 JA |
5075 | if (do_complete) { |
5076 | io_cqring_fill_event(req, -ECANCELED); | |
5077 | io_commit_cqring(req->ctx); | |
5078 | req->flags |= REQ_F_COMP_LOCKED; | |
f254ac04 | 5079 | req_set_fail_links(req); |
b41e9852 JA |
5080 | io_put_req(req); |
5081 | } | |
5082 | ||
5083 | return do_complete; | |
221c5eb2 JA |
5084 | } |
5085 | ||
76e1b642 JA |
5086 | /* |
5087 | * Returns true if we found and killed one or more poll requests | |
5088 | */ | |
5089 | static bool io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk) | |
221c5eb2 | 5090 | { |
78076bb6 | 5091 | struct hlist_node *tmp; |
221c5eb2 | 5092 | struct io_kiocb *req; |
8e2e1faf | 5093 | int posted = 0, i; |
221c5eb2 JA |
5094 | |
5095 | spin_lock_irq(&ctx->completion_lock); | |
78076bb6 JA |
5096 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { |
5097 | struct hlist_head *list; | |
5098 | ||
5099 | list = &ctx->cancel_hash[i]; | |
f3606e3a JA |
5100 | hlist_for_each_entry_safe(req, tmp, list, hash_node) { |
5101 | if (io_task_match(req, tsk)) | |
5102 | posted += io_poll_remove_one(req); | |
5103 | } | |
221c5eb2 JA |
5104 | } |
5105 | spin_unlock_irq(&ctx->completion_lock); | |
b41e9852 | 5106 | |
8e2e1faf JA |
5107 | if (posted) |
5108 | io_cqring_ev_posted(ctx); | |
76e1b642 JA |
5109 | |
5110 | return posted != 0; | |
221c5eb2 JA |
5111 | } |
5112 | ||
47f46768 JA |
5113 | static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr) |
5114 | { | |
78076bb6 | 5115 | struct hlist_head *list; |
47f46768 JA |
5116 | struct io_kiocb *req; |
5117 | ||
78076bb6 JA |
5118 | list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)]; |
5119 | hlist_for_each_entry(req, list, hash_node) { | |
b41e9852 JA |
5120 | if (sqe_addr != req->user_data) |
5121 | continue; | |
5122 | if (io_poll_remove_one(req)) | |
eac406c6 | 5123 | return 0; |
b41e9852 | 5124 | return -EALREADY; |
47f46768 JA |
5125 | } |
5126 | ||
5127 | return -ENOENT; | |
5128 | } | |
5129 | ||
3529d8c2 JA |
5130 | static int io_poll_remove_prep(struct io_kiocb *req, |
5131 | const struct io_uring_sqe *sqe) | |
0969e783 | 5132 | { |
0969e783 JA |
5133 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5134 | return -EINVAL; | |
5135 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index || | |
5136 | sqe->poll_events) | |
5137 | return -EINVAL; | |
5138 | ||
5139 | req->poll.addr = READ_ONCE(sqe->addr); | |
0969e783 JA |
5140 | return 0; |
5141 | } | |
5142 | ||
221c5eb2 JA |
5143 | /* |
5144 | * Find a running poll command that matches one specified in sqe->addr, | |
5145 | * and remove it if found. | |
5146 | */ | |
fc4df999 | 5147 | static int io_poll_remove(struct io_kiocb *req) |
221c5eb2 JA |
5148 | { |
5149 | struct io_ring_ctx *ctx = req->ctx; | |
0969e783 | 5150 | u64 addr; |
47f46768 | 5151 | int ret; |
221c5eb2 | 5152 | |
0969e783 | 5153 | addr = req->poll.addr; |
221c5eb2 | 5154 | spin_lock_irq(&ctx->completion_lock); |
0969e783 | 5155 | ret = io_poll_cancel(ctx, addr); |
221c5eb2 JA |
5156 | spin_unlock_irq(&ctx->completion_lock); |
5157 | ||
4e88d6e7 JA |
5158 | if (ret < 0) |
5159 | req_set_fail_links(req); | |
e1e16097 | 5160 | io_req_complete(req, ret); |
221c5eb2 JA |
5161 | return 0; |
5162 | } | |
5163 | ||
221c5eb2 JA |
5164 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, |
5165 | void *key) | |
5166 | { | |
c2f2eb7d JA |
5167 | struct io_kiocb *req = wait->private; |
5168 | struct io_poll_iocb *poll = &req->poll; | |
221c5eb2 | 5169 | |
d7718a9d | 5170 | return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func); |
221c5eb2 JA |
5171 | } |
5172 | ||
221c5eb2 JA |
5173 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, |
5174 | struct poll_table_struct *p) | |
5175 | { | |
5176 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
5177 | ||
807abcb0 | 5178 | __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->io); |
eac406c6 JA |
5179 | } |
5180 | ||
3529d8c2 | 5181 | static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
221c5eb2 JA |
5182 | { |
5183 | struct io_poll_iocb *poll = &req->poll; | |
5769a351 | 5184 | u32 events; |
221c5eb2 JA |
5185 | |
5186 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
5187 | return -EINVAL; | |
5188 | if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
5189 | return -EINVAL; | |
09bb8394 JA |
5190 | if (!poll->file) |
5191 | return -EBADF; | |
221c5eb2 | 5192 | |
5769a351 JX |
5193 | events = READ_ONCE(sqe->poll32_events); |
5194 | #ifdef __BIG_ENDIAN | |
5195 | events = swahw32(events); | |
5196 | #endif | |
a31eb4a2 JX |
5197 | poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP | |
5198 | (events & EPOLLEXCLUSIVE); | |
0969e783 JA |
5199 | return 0; |
5200 | } | |
5201 | ||
014db007 | 5202 | static int io_poll_add(struct io_kiocb *req) |
0969e783 JA |
5203 | { |
5204 | struct io_poll_iocb *poll = &req->poll; | |
5205 | struct io_ring_ctx *ctx = req->ctx; | |
5206 | struct io_poll_table ipt; | |
0969e783 | 5207 | __poll_t mask; |
0969e783 | 5208 | |
78076bb6 | 5209 | INIT_HLIST_NODE(&req->hash_node); |
d7718a9d | 5210 | ipt.pt._qproc = io_poll_queue_proc; |
36703247 | 5211 | |
d7718a9d JA |
5212 | mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events, |
5213 | io_poll_wake); | |
221c5eb2 | 5214 | |
8c838788 | 5215 | if (mask) { /* no async, we'd stolen it */ |
221c5eb2 | 5216 | ipt.error = 0; |
b0dd8a41 | 5217 | io_poll_complete(req, mask, 0); |
221c5eb2 | 5218 | } |
221c5eb2 JA |
5219 | spin_unlock_irq(&ctx->completion_lock); |
5220 | ||
8c838788 JA |
5221 | if (mask) { |
5222 | io_cqring_ev_posted(ctx); | |
014db007 | 5223 | io_put_req(req); |
221c5eb2 | 5224 | } |
8c838788 | 5225 | return ipt.error; |
221c5eb2 JA |
5226 | } |
5227 | ||
5262f567 JA |
5228 | static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer) |
5229 | { | |
ad8a48ac JA |
5230 | struct io_timeout_data *data = container_of(timer, |
5231 | struct io_timeout_data, timer); | |
5232 | struct io_kiocb *req = data->req; | |
5233 | struct io_ring_ctx *ctx = req->ctx; | |
5262f567 JA |
5234 | unsigned long flags; |
5235 | ||
5262f567 | 5236 | spin_lock_irqsave(&ctx->completion_lock, flags); |
01cec8c1 PB |
5237 | atomic_set(&req->ctx->cq_timeouts, |
5238 | atomic_read(&req->ctx->cq_timeouts) + 1); | |
5239 | ||
ef03681a | 5240 | /* |
11365043 JA |
5241 | * We could be racing with timeout deletion. If the list is empty, |
5242 | * then timeout lookup already found it and will be handling it. | |
ef03681a | 5243 | */ |
135fcde8 PB |
5244 | if (!list_empty(&req->timeout.list)) |
5245 | list_del_init(&req->timeout.list); | |
5262f567 | 5246 | |
78e19bbe | 5247 | io_cqring_fill_event(req, -ETIME); |
5262f567 JA |
5248 | io_commit_cqring(ctx); |
5249 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
5250 | ||
5251 | io_cqring_ev_posted(ctx); | |
4e88d6e7 | 5252 | req_set_fail_links(req); |
5262f567 JA |
5253 | io_put_req(req); |
5254 | return HRTIMER_NORESTART; | |
5255 | } | |
5256 | ||
f254ac04 JA |
5257 | static int __io_timeout_cancel(struct io_kiocb *req) |
5258 | { | |
5259 | int ret; | |
5260 | ||
5261 | list_del_init(&req->timeout.list); | |
5262 | ||
5263 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); | |
5264 | if (ret == -1) | |
5265 | return -EALREADY; | |
5266 | ||
5267 | req_set_fail_links(req); | |
5268 | req->flags |= REQ_F_COMP_LOCKED; | |
5269 | io_cqring_fill_event(req, -ECANCELED); | |
5270 | io_put_req(req); | |
5271 | return 0; | |
5272 | } | |
5273 | ||
47f46768 JA |
5274 | static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data) |
5275 | { | |
5276 | struct io_kiocb *req; | |
5277 | int ret = -ENOENT; | |
5278 | ||
135fcde8 | 5279 | list_for_each_entry(req, &ctx->timeout_list, timeout.list) { |
47f46768 | 5280 | if (user_data == req->user_data) { |
47f46768 JA |
5281 | ret = 0; |
5282 | break; | |
5283 | } | |
5284 | } | |
5285 | ||
5286 | if (ret == -ENOENT) | |
5287 | return ret; | |
5288 | ||
f254ac04 | 5289 | return __io_timeout_cancel(req); |
47f46768 JA |
5290 | } |
5291 | ||
3529d8c2 JA |
5292 | static int io_timeout_remove_prep(struct io_kiocb *req, |
5293 | const struct io_uring_sqe *sqe) | |
b29472ee | 5294 | { |
b29472ee JA |
5295 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5296 | return -EINVAL; | |
61710e43 DA |
5297 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) |
5298 | return -EINVAL; | |
5299 | if (sqe->ioprio || sqe->buf_index || sqe->len) | |
b29472ee JA |
5300 | return -EINVAL; |
5301 | ||
5302 | req->timeout.addr = READ_ONCE(sqe->addr); | |
5303 | req->timeout.flags = READ_ONCE(sqe->timeout_flags); | |
5304 | if (req->timeout.flags) | |
5305 | return -EINVAL; | |
5306 | ||
b29472ee JA |
5307 | return 0; |
5308 | } | |
5309 | ||
11365043 JA |
5310 | /* |
5311 | * Remove or update an existing timeout command | |
5312 | */ | |
fc4df999 | 5313 | static int io_timeout_remove(struct io_kiocb *req) |
11365043 JA |
5314 | { |
5315 | struct io_ring_ctx *ctx = req->ctx; | |
47f46768 | 5316 | int ret; |
11365043 | 5317 | |
11365043 | 5318 | spin_lock_irq(&ctx->completion_lock); |
b29472ee | 5319 | ret = io_timeout_cancel(ctx, req->timeout.addr); |
11365043 | 5320 | |
47f46768 | 5321 | io_cqring_fill_event(req, ret); |
11365043 JA |
5322 | io_commit_cqring(ctx); |
5323 | spin_unlock_irq(&ctx->completion_lock); | |
5262f567 | 5324 | io_cqring_ev_posted(ctx); |
4e88d6e7 JA |
5325 | if (ret < 0) |
5326 | req_set_fail_links(req); | |
ec9c02ad | 5327 | io_put_req(req); |
11365043 | 5328 | return 0; |
5262f567 JA |
5329 | } |
5330 | ||
3529d8c2 | 5331 | static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2d28390a | 5332 | bool is_timeout_link) |
5262f567 | 5333 | { |
ad8a48ac | 5334 | struct io_timeout_data *data; |
a41525ab | 5335 | unsigned flags; |
56080b02 | 5336 | u32 off = READ_ONCE(sqe->off); |
5262f567 | 5337 | |
ad8a48ac | 5338 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5262f567 | 5339 | return -EINVAL; |
ad8a48ac | 5340 | if (sqe->ioprio || sqe->buf_index || sqe->len != 1) |
a41525ab | 5341 | return -EINVAL; |
56080b02 | 5342 | if (off && is_timeout_link) |
2d28390a | 5343 | return -EINVAL; |
a41525ab JA |
5344 | flags = READ_ONCE(sqe->timeout_flags); |
5345 | if (flags & ~IORING_TIMEOUT_ABS) | |
5262f567 | 5346 | return -EINVAL; |
bdf20073 | 5347 | |
bfe68a22 | 5348 | req->timeout.off = off; |
26a61679 | 5349 | |
3529d8c2 | 5350 | if (!req->io && io_alloc_async_ctx(req)) |
26a61679 JA |
5351 | return -ENOMEM; |
5352 | ||
5353 | data = &req->io->timeout; | |
ad8a48ac | 5354 | data->req = req; |
ad8a48ac JA |
5355 | |
5356 | if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr))) | |
5262f567 JA |
5357 | return -EFAULT; |
5358 | ||
11365043 | 5359 | if (flags & IORING_TIMEOUT_ABS) |
ad8a48ac | 5360 | data->mode = HRTIMER_MODE_ABS; |
11365043 | 5361 | else |
ad8a48ac | 5362 | data->mode = HRTIMER_MODE_REL; |
11365043 | 5363 | |
ad8a48ac JA |
5364 | hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode); |
5365 | return 0; | |
5366 | } | |
5367 | ||
fc4df999 | 5368 | static int io_timeout(struct io_kiocb *req) |
ad8a48ac | 5369 | { |
ad8a48ac | 5370 | struct io_ring_ctx *ctx = req->ctx; |
bfe68a22 | 5371 | struct io_timeout_data *data = &req->io->timeout; |
ad8a48ac | 5372 | struct list_head *entry; |
bfe68a22 | 5373 | u32 tail, off = req->timeout.off; |
ad8a48ac | 5374 | |
733f5c95 | 5375 | spin_lock_irq(&ctx->completion_lock); |
93bd25bb | 5376 | |
5262f567 JA |
5377 | /* |
5378 | * sqe->off holds how many events that need to occur for this | |
93bd25bb JA |
5379 | * timeout event to be satisfied. If it isn't set, then this is |
5380 | * a pure timeout request, sequence isn't used. | |
5262f567 | 5381 | */ |
8eb7e2d0 | 5382 | if (io_is_timeout_noseq(req)) { |
93bd25bb JA |
5383 | entry = ctx->timeout_list.prev; |
5384 | goto add; | |
5385 | } | |
5262f567 | 5386 | |
bfe68a22 PB |
5387 | tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts); |
5388 | req->timeout.target_seq = tail + off; | |
5262f567 JA |
5389 | |
5390 | /* | |
5391 | * Insertion sort, ensuring the first entry in the list is always | |
5392 | * the one we need first. | |
5393 | */ | |
5262f567 | 5394 | list_for_each_prev(entry, &ctx->timeout_list) { |
135fcde8 PB |
5395 | struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, |
5396 | timeout.list); | |
5262f567 | 5397 | |
8eb7e2d0 | 5398 | if (io_is_timeout_noseq(nxt)) |
93bd25bb | 5399 | continue; |
bfe68a22 PB |
5400 | /* nxt.seq is behind @tail, otherwise would've been completed */ |
5401 | if (off >= nxt->timeout.target_seq - tail) | |
5262f567 JA |
5402 | break; |
5403 | } | |
93bd25bb | 5404 | add: |
135fcde8 | 5405 | list_add(&req->timeout.list, entry); |
ad8a48ac JA |
5406 | data->timer.function = io_timeout_fn; |
5407 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode); | |
5262f567 | 5408 | spin_unlock_irq(&ctx->completion_lock); |
5262f567 JA |
5409 | return 0; |
5410 | } | |
5262f567 | 5411 | |
62755e35 JA |
5412 | static bool io_cancel_cb(struct io_wq_work *work, void *data) |
5413 | { | |
5414 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
5415 | ||
5416 | return req->user_data == (unsigned long) data; | |
5417 | } | |
5418 | ||
e977d6d3 | 5419 | static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr) |
62755e35 | 5420 | { |
62755e35 | 5421 | enum io_wq_cancel cancel_ret; |
62755e35 JA |
5422 | int ret = 0; |
5423 | ||
4f26bda1 | 5424 | cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false); |
62755e35 JA |
5425 | switch (cancel_ret) { |
5426 | case IO_WQ_CANCEL_OK: | |
5427 | ret = 0; | |
5428 | break; | |
5429 | case IO_WQ_CANCEL_RUNNING: | |
5430 | ret = -EALREADY; | |
5431 | break; | |
5432 | case IO_WQ_CANCEL_NOTFOUND: | |
5433 | ret = -ENOENT; | |
5434 | break; | |
5435 | } | |
5436 | ||
e977d6d3 JA |
5437 | return ret; |
5438 | } | |
5439 | ||
47f46768 JA |
5440 | static void io_async_find_and_cancel(struct io_ring_ctx *ctx, |
5441 | struct io_kiocb *req, __u64 sqe_addr, | |
014db007 | 5442 | int success_ret) |
47f46768 JA |
5443 | { |
5444 | unsigned long flags; | |
5445 | int ret; | |
5446 | ||
5447 | ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr); | |
5448 | if (ret != -ENOENT) { | |
5449 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
5450 | goto done; | |
5451 | } | |
5452 | ||
5453 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
5454 | ret = io_timeout_cancel(ctx, sqe_addr); | |
5455 | if (ret != -ENOENT) | |
5456 | goto done; | |
5457 | ret = io_poll_cancel(ctx, sqe_addr); | |
5458 | done: | |
b0dd8a41 JA |
5459 | if (!ret) |
5460 | ret = success_ret; | |
47f46768 JA |
5461 | io_cqring_fill_event(req, ret); |
5462 | io_commit_cqring(ctx); | |
5463 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
5464 | io_cqring_ev_posted(ctx); | |
5465 | ||
4e88d6e7 JA |
5466 | if (ret < 0) |
5467 | req_set_fail_links(req); | |
014db007 | 5468 | io_put_req(req); |
47f46768 JA |
5469 | } |
5470 | ||
3529d8c2 JA |
5471 | static int io_async_cancel_prep(struct io_kiocb *req, |
5472 | const struct io_uring_sqe *sqe) | |
e977d6d3 | 5473 | { |
fbf23849 | 5474 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
e977d6d3 | 5475 | return -EINVAL; |
61710e43 DA |
5476 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) |
5477 | return -EINVAL; | |
5478 | if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags) | |
e977d6d3 JA |
5479 | return -EINVAL; |
5480 | ||
fbf23849 JA |
5481 | req->cancel.addr = READ_ONCE(sqe->addr); |
5482 | return 0; | |
5483 | } | |
5484 | ||
014db007 | 5485 | static int io_async_cancel(struct io_kiocb *req) |
fbf23849 JA |
5486 | { |
5487 | struct io_ring_ctx *ctx = req->ctx; | |
fbf23849 | 5488 | |
014db007 | 5489 | io_async_find_and_cancel(ctx, req, req->cancel.addr, 0); |
5262f567 JA |
5490 | return 0; |
5491 | } | |
5492 | ||
05f3fb3c JA |
5493 | static int io_files_update_prep(struct io_kiocb *req, |
5494 | const struct io_uring_sqe *sqe) | |
5495 | { | |
6ca56f84 JA |
5496 | if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL)) |
5497 | return -EINVAL; | |
61710e43 DA |
5498 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) |
5499 | return -EINVAL; | |
5500 | if (sqe->ioprio || sqe->rw_flags) | |
05f3fb3c JA |
5501 | return -EINVAL; |
5502 | ||
5503 | req->files_update.offset = READ_ONCE(sqe->off); | |
5504 | req->files_update.nr_args = READ_ONCE(sqe->len); | |
5505 | if (!req->files_update.nr_args) | |
5506 | return -EINVAL; | |
5507 | req->files_update.arg = READ_ONCE(sqe->addr); | |
5508 | return 0; | |
5509 | } | |
5510 | ||
229a7b63 JA |
5511 | static int io_files_update(struct io_kiocb *req, bool force_nonblock, |
5512 | struct io_comp_state *cs) | |
fbf23849 JA |
5513 | { |
5514 | struct io_ring_ctx *ctx = req->ctx; | |
05f3fb3c JA |
5515 | struct io_uring_files_update up; |
5516 | int ret; | |
fbf23849 | 5517 | |
f86cd20c | 5518 | if (force_nonblock) |
05f3fb3c | 5519 | return -EAGAIN; |
05f3fb3c JA |
5520 | |
5521 | up.offset = req->files_update.offset; | |
5522 | up.fds = req->files_update.arg; | |
5523 | ||
5524 | mutex_lock(&ctx->uring_lock); | |
5525 | ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args); | |
5526 | mutex_unlock(&ctx->uring_lock); | |
5527 | ||
5528 | if (ret < 0) | |
5529 | req_set_fail_links(req); | |
229a7b63 | 5530 | __io_req_complete(req, ret, 0, cs); |
5262f567 JA |
5531 | return 0; |
5532 | } | |
5533 | ||
3529d8c2 JA |
5534 | static int io_req_defer_prep(struct io_kiocb *req, |
5535 | const struct io_uring_sqe *sqe) | |
f67676d1 | 5536 | { |
e781573e | 5537 | ssize_t ret = 0; |
f67676d1 | 5538 | |
f1d96a8f PB |
5539 | if (!sqe) |
5540 | return 0; | |
5541 | ||
327d6d96 PB |
5542 | if (io_alloc_async_ctx(req)) |
5543 | return -EAGAIN; | |
f56040b8 PB |
5544 | ret = io_prep_work_files(req); |
5545 | if (unlikely(ret)) | |
5546 | return ret; | |
cccf0ee8 | 5547 | |
202700e1 JA |
5548 | io_prep_async_work(req); |
5549 | ||
d625c6ee | 5550 | switch (req->opcode) { |
e781573e JA |
5551 | case IORING_OP_NOP: |
5552 | break; | |
f67676d1 JA |
5553 | case IORING_OP_READV: |
5554 | case IORING_OP_READ_FIXED: | |
3a6820f2 | 5555 | case IORING_OP_READ: |
3529d8c2 | 5556 | ret = io_read_prep(req, sqe, true); |
f67676d1 JA |
5557 | break; |
5558 | case IORING_OP_WRITEV: | |
5559 | case IORING_OP_WRITE_FIXED: | |
3a6820f2 | 5560 | case IORING_OP_WRITE: |
3529d8c2 | 5561 | ret = io_write_prep(req, sqe, true); |
f67676d1 | 5562 | break; |
0969e783 | 5563 | case IORING_OP_POLL_ADD: |
3529d8c2 | 5564 | ret = io_poll_add_prep(req, sqe); |
0969e783 JA |
5565 | break; |
5566 | case IORING_OP_POLL_REMOVE: | |
3529d8c2 | 5567 | ret = io_poll_remove_prep(req, sqe); |
0969e783 | 5568 | break; |
8ed8d3c3 | 5569 | case IORING_OP_FSYNC: |
3529d8c2 | 5570 | ret = io_prep_fsync(req, sqe); |
8ed8d3c3 JA |
5571 | break; |
5572 | case IORING_OP_SYNC_FILE_RANGE: | |
3529d8c2 | 5573 | ret = io_prep_sfr(req, sqe); |
8ed8d3c3 | 5574 | break; |
03b1230c | 5575 | case IORING_OP_SENDMSG: |
fddaface | 5576 | case IORING_OP_SEND: |
3529d8c2 | 5577 | ret = io_sendmsg_prep(req, sqe); |
03b1230c JA |
5578 | break; |
5579 | case IORING_OP_RECVMSG: | |
fddaface | 5580 | case IORING_OP_RECV: |
3529d8c2 | 5581 | ret = io_recvmsg_prep(req, sqe); |
03b1230c | 5582 | break; |
f499a021 | 5583 | case IORING_OP_CONNECT: |
3529d8c2 | 5584 | ret = io_connect_prep(req, sqe); |
f499a021 | 5585 | break; |
2d28390a | 5586 | case IORING_OP_TIMEOUT: |
3529d8c2 | 5587 | ret = io_timeout_prep(req, sqe, false); |
b7bb4f7d | 5588 | break; |
b29472ee | 5589 | case IORING_OP_TIMEOUT_REMOVE: |
3529d8c2 | 5590 | ret = io_timeout_remove_prep(req, sqe); |
b29472ee | 5591 | break; |
fbf23849 | 5592 | case IORING_OP_ASYNC_CANCEL: |
3529d8c2 | 5593 | ret = io_async_cancel_prep(req, sqe); |
fbf23849 | 5594 | break; |
2d28390a | 5595 | case IORING_OP_LINK_TIMEOUT: |
3529d8c2 | 5596 | ret = io_timeout_prep(req, sqe, true); |
b7bb4f7d | 5597 | break; |
8ed8d3c3 | 5598 | case IORING_OP_ACCEPT: |
3529d8c2 | 5599 | ret = io_accept_prep(req, sqe); |
8ed8d3c3 | 5600 | break; |
d63d1b5e JA |
5601 | case IORING_OP_FALLOCATE: |
5602 | ret = io_fallocate_prep(req, sqe); | |
5603 | break; | |
15b71abe JA |
5604 | case IORING_OP_OPENAT: |
5605 | ret = io_openat_prep(req, sqe); | |
5606 | break; | |
b5dba59e JA |
5607 | case IORING_OP_CLOSE: |
5608 | ret = io_close_prep(req, sqe); | |
5609 | break; | |
05f3fb3c JA |
5610 | case IORING_OP_FILES_UPDATE: |
5611 | ret = io_files_update_prep(req, sqe); | |
5612 | break; | |
eddc7ef5 JA |
5613 | case IORING_OP_STATX: |
5614 | ret = io_statx_prep(req, sqe); | |
5615 | break; | |
4840e418 JA |
5616 | case IORING_OP_FADVISE: |
5617 | ret = io_fadvise_prep(req, sqe); | |
5618 | break; | |
c1ca757b JA |
5619 | case IORING_OP_MADVISE: |
5620 | ret = io_madvise_prep(req, sqe); | |
5621 | break; | |
cebdb986 JA |
5622 | case IORING_OP_OPENAT2: |
5623 | ret = io_openat2_prep(req, sqe); | |
5624 | break; | |
3e4827b0 JA |
5625 | case IORING_OP_EPOLL_CTL: |
5626 | ret = io_epoll_ctl_prep(req, sqe); | |
5627 | break; | |
7d67af2c PB |
5628 | case IORING_OP_SPLICE: |
5629 | ret = io_splice_prep(req, sqe); | |
5630 | break; | |
ddf0322d JA |
5631 | case IORING_OP_PROVIDE_BUFFERS: |
5632 | ret = io_provide_buffers_prep(req, sqe); | |
5633 | break; | |
067524e9 JA |
5634 | case IORING_OP_REMOVE_BUFFERS: |
5635 | ret = io_remove_buffers_prep(req, sqe); | |
5636 | break; | |
f2a8d5c7 PB |
5637 | case IORING_OP_TEE: |
5638 | ret = io_tee_prep(req, sqe); | |
5639 | break; | |
f67676d1 | 5640 | default: |
e781573e JA |
5641 | printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n", |
5642 | req->opcode); | |
5643 | ret = -EINVAL; | |
b7bb4f7d | 5644 | break; |
f67676d1 JA |
5645 | } |
5646 | ||
b7bb4f7d | 5647 | return ret; |
f67676d1 JA |
5648 | } |
5649 | ||
9cf7c104 PB |
5650 | static u32 io_get_sequence(struct io_kiocb *req) |
5651 | { | |
5652 | struct io_kiocb *pos; | |
5653 | struct io_ring_ctx *ctx = req->ctx; | |
5654 | u32 total_submitted, nr_reqs = 1; | |
5655 | ||
5656 | if (req->flags & REQ_F_LINK_HEAD) | |
5657 | list_for_each_entry(pos, &req->link_list, link_list) | |
5658 | nr_reqs++; | |
5659 | ||
5660 | total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped; | |
5661 | return total_submitted - nr_reqs; | |
5662 | } | |
5663 | ||
3529d8c2 | 5664 | static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
de0617e4 | 5665 | { |
a197f664 | 5666 | struct io_ring_ctx *ctx = req->ctx; |
27dc8338 | 5667 | struct io_defer_entry *de; |
f67676d1 | 5668 | int ret; |
9cf7c104 | 5669 | u32 seq; |
de0617e4 | 5670 | |
9d858b21 | 5671 | /* Still need defer if there is pending req in defer list. */ |
9cf7c104 PB |
5672 | if (likely(list_empty_careful(&ctx->defer_list) && |
5673 | !(req->flags & REQ_F_IO_DRAIN))) | |
5674 | return 0; | |
5675 | ||
5676 | seq = io_get_sequence(req); | |
5677 | /* Still a chance to pass the sequence check */ | |
5678 | if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) | |
de0617e4 JA |
5679 | return 0; |
5680 | ||
650b5481 | 5681 | if (!req->io) { |
650b5481 | 5682 | ret = io_req_defer_prep(req, sqe); |
327d6d96 | 5683 | if (ret) |
650b5481 PB |
5684 | return ret; |
5685 | } | |
cbdcb435 | 5686 | io_prep_async_link(req); |
27dc8338 PB |
5687 | de = kmalloc(sizeof(*de), GFP_KERNEL); |
5688 | if (!de) | |
5689 | return -ENOMEM; | |
2d28390a | 5690 | |
de0617e4 | 5691 | spin_lock_irq(&ctx->completion_lock); |
9cf7c104 | 5692 | if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { |
de0617e4 | 5693 | spin_unlock_irq(&ctx->completion_lock); |
27dc8338 | 5694 | kfree(de); |
ae34817b PB |
5695 | io_queue_async_work(req); |
5696 | return -EIOCBQUEUED; | |
de0617e4 JA |
5697 | } |
5698 | ||
915967f6 | 5699 | trace_io_uring_defer(ctx, req, req->user_data); |
27dc8338 | 5700 | de->req = req; |
9cf7c104 | 5701 | de->seq = seq; |
27dc8338 | 5702 | list_add_tail(&de->list, &ctx->defer_list); |
de0617e4 JA |
5703 | spin_unlock_irq(&ctx->completion_lock); |
5704 | return -EIOCBQUEUED; | |
5705 | } | |
5706 | ||
f573d384 JA |
5707 | static void io_req_drop_files(struct io_kiocb *req) |
5708 | { | |
5709 | struct io_ring_ctx *ctx = req->ctx; | |
5710 | unsigned long flags; | |
5711 | ||
5712 | spin_lock_irqsave(&ctx->inflight_lock, flags); | |
5713 | list_del(&req->inflight_entry); | |
5714 | if (waitqueue_active(&ctx->inflight_wait)) | |
5715 | wake_up(&ctx->inflight_wait); | |
5716 | spin_unlock_irqrestore(&ctx->inflight_lock, flags); | |
5717 | req->flags &= ~REQ_F_INFLIGHT; | |
0f212204 | 5718 | put_files_struct(req->work.files); |
9b828492 | 5719 | put_nsproxy(req->work.nsproxy); |
f573d384 JA |
5720 | req->work.files = NULL; |
5721 | } | |
5722 | ||
3ca405eb | 5723 | static void __io_clean_op(struct io_kiocb *req) |
99bc4c38 PB |
5724 | { |
5725 | struct io_async_ctx *io = req->io; | |
5726 | ||
0e1b6fe3 PB |
5727 | if (req->flags & REQ_F_BUFFER_SELECTED) { |
5728 | switch (req->opcode) { | |
5729 | case IORING_OP_READV: | |
5730 | case IORING_OP_READ_FIXED: | |
5731 | case IORING_OP_READ: | |
bcda7baa | 5732 | kfree((void *)(unsigned long)req->rw.addr); |
0e1b6fe3 PB |
5733 | break; |
5734 | case IORING_OP_RECVMSG: | |
5735 | case IORING_OP_RECV: | |
bcda7baa | 5736 | kfree(req->sr_msg.kbuf); |
0e1b6fe3 PB |
5737 | break; |
5738 | } | |
5739 | req->flags &= ~REQ_F_BUFFER_SELECTED; | |
99bc4c38 PB |
5740 | } |
5741 | ||
0e1b6fe3 PB |
5742 | if (req->flags & REQ_F_NEED_CLEANUP) { |
5743 | switch (req->opcode) { | |
5744 | case IORING_OP_READV: | |
5745 | case IORING_OP_READ_FIXED: | |
5746 | case IORING_OP_READ: | |
5747 | case IORING_OP_WRITEV: | |
5748 | case IORING_OP_WRITE_FIXED: | |
5749 | case IORING_OP_WRITE: | |
ff6165b2 JA |
5750 | if (io->rw.free_iovec) |
5751 | kfree(io->rw.free_iovec); | |
0e1b6fe3 PB |
5752 | break; |
5753 | case IORING_OP_RECVMSG: | |
5754 | case IORING_OP_SENDMSG: | |
5755 | if (io->msg.iov != io->msg.fast_iov) | |
5756 | kfree(io->msg.iov); | |
5757 | break; | |
5758 | case IORING_OP_SPLICE: | |
5759 | case IORING_OP_TEE: | |
5760 | io_put_file(req, req->splice.file_in, | |
5761 | (req->splice.flags & SPLICE_F_FD_IN_FIXED)); | |
5762 | break; | |
f3cd4850 JA |
5763 | case IORING_OP_OPENAT: |
5764 | case IORING_OP_OPENAT2: | |
5765 | if (req->open.filename) | |
5766 | putname(req->open.filename); | |
5767 | break; | |
0e1b6fe3 PB |
5768 | } |
5769 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
99bc4c38 | 5770 | } |
bb175342 | 5771 | |
f573d384 JA |
5772 | if (req->flags & REQ_F_INFLIGHT) |
5773 | io_req_drop_files(req); | |
99bc4c38 PB |
5774 | } |
5775 | ||
3529d8c2 | 5776 | static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
f13fad7b | 5777 | bool force_nonblock, struct io_comp_state *cs) |
2b188cc1 | 5778 | { |
a197f664 | 5779 | struct io_ring_ctx *ctx = req->ctx; |
d625c6ee | 5780 | int ret; |
2b188cc1 | 5781 | |
d625c6ee | 5782 | switch (req->opcode) { |
2b188cc1 | 5783 | case IORING_OP_NOP: |
229a7b63 | 5784 | ret = io_nop(req, cs); |
2b188cc1 JA |
5785 | break; |
5786 | case IORING_OP_READV: | |
edafccee | 5787 | case IORING_OP_READ_FIXED: |
3a6820f2 | 5788 | case IORING_OP_READ: |
3529d8c2 JA |
5789 | if (sqe) { |
5790 | ret = io_read_prep(req, sqe, force_nonblock); | |
5791 | if (ret < 0) | |
5792 | break; | |
5793 | } | |
a1d7c393 | 5794 | ret = io_read(req, force_nonblock, cs); |
edafccee | 5795 | break; |
3529d8c2 | 5796 | case IORING_OP_WRITEV: |
edafccee | 5797 | case IORING_OP_WRITE_FIXED: |
3a6820f2 | 5798 | case IORING_OP_WRITE: |
3529d8c2 JA |
5799 | if (sqe) { |
5800 | ret = io_write_prep(req, sqe, force_nonblock); | |
5801 | if (ret < 0) | |
5802 | break; | |
5803 | } | |
a1d7c393 | 5804 | ret = io_write(req, force_nonblock, cs); |
2b188cc1 | 5805 | break; |
c992fe29 | 5806 | case IORING_OP_FSYNC: |
3529d8c2 JA |
5807 | if (sqe) { |
5808 | ret = io_prep_fsync(req, sqe); | |
5809 | if (ret < 0) | |
5810 | break; | |
5811 | } | |
014db007 | 5812 | ret = io_fsync(req, force_nonblock); |
c992fe29 | 5813 | break; |
221c5eb2 | 5814 | case IORING_OP_POLL_ADD: |
3529d8c2 JA |
5815 | if (sqe) { |
5816 | ret = io_poll_add_prep(req, sqe); | |
5817 | if (ret) | |
5818 | break; | |
5819 | } | |
014db007 | 5820 | ret = io_poll_add(req); |
221c5eb2 JA |
5821 | break; |
5822 | case IORING_OP_POLL_REMOVE: | |
3529d8c2 JA |
5823 | if (sqe) { |
5824 | ret = io_poll_remove_prep(req, sqe); | |
5825 | if (ret < 0) | |
5826 | break; | |
5827 | } | |
fc4df999 | 5828 | ret = io_poll_remove(req); |
221c5eb2 | 5829 | break; |
5d17b4a4 | 5830 | case IORING_OP_SYNC_FILE_RANGE: |
3529d8c2 JA |
5831 | if (sqe) { |
5832 | ret = io_prep_sfr(req, sqe); | |
5833 | if (ret < 0) | |
5834 | break; | |
5835 | } | |
014db007 | 5836 | ret = io_sync_file_range(req, force_nonblock); |
5d17b4a4 | 5837 | break; |
0fa03c62 | 5838 | case IORING_OP_SENDMSG: |
fddaface | 5839 | case IORING_OP_SEND: |
3529d8c2 JA |
5840 | if (sqe) { |
5841 | ret = io_sendmsg_prep(req, sqe); | |
5842 | if (ret < 0) | |
5843 | break; | |
5844 | } | |
fddaface | 5845 | if (req->opcode == IORING_OP_SENDMSG) |
229a7b63 | 5846 | ret = io_sendmsg(req, force_nonblock, cs); |
fddaface | 5847 | else |
229a7b63 | 5848 | ret = io_send(req, force_nonblock, cs); |
0fa03c62 | 5849 | break; |
aa1fa28f | 5850 | case IORING_OP_RECVMSG: |
fddaface | 5851 | case IORING_OP_RECV: |
3529d8c2 JA |
5852 | if (sqe) { |
5853 | ret = io_recvmsg_prep(req, sqe); | |
5854 | if (ret) | |
5855 | break; | |
5856 | } | |
fddaface | 5857 | if (req->opcode == IORING_OP_RECVMSG) |
229a7b63 | 5858 | ret = io_recvmsg(req, force_nonblock, cs); |
fddaface | 5859 | else |
229a7b63 | 5860 | ret = io_recv(req, force_nonblock, cs); |
aa1fa28f | 5861 | break; |
5262f567 | 5862 | case IORING_OP_TIMEOUT: |
3529d8c2 JA |
5863 | if (sqe) { |
5864 | ret = io_timeout_prep(req, sqe, false); | |
5865 | if (ret) | |
5866 | break; | |
5867 | } | |
fc4df999 | 5868 | ret = io_timeout(req); |
5262f567 | 5869 | break; |
11365043 | 5870 | case IORING_OP_TIMEOUT_REMOVE: |
3529d8c2 JA |
5871 | if (sqe) { |
5872 | ret = io_timeout_remove_prep(req, sqe); | |
5873 | if (ret) | |
5874 | break; | |
5875 | } | |
fc4df999 | 5876 | ret = io_timeout_remove(req); |
11365043 | 5877 | break; |
17f2fe35 | 5878 | case IORING_OP_ACCEPT: |
3529d8c2 JA |
5879 | if (sqe) { |
5880 | ret = io_accept_prep(req, sqe); | |
5881 | if (ret) | |
5882 | break; | |
5883 | } | |
229a7b63 | 5884 | ret = io_accept(req, force_nonblock, cs); |
17f2fe35 | 5885 | break; |
f8e85cf2 | 5886 | case IORING_OP_CONNECT: |
3529d8c2 JA |
5887 | if (sqe) { |
5888 | ret = io_connect_prep(req, sqe); | |
5889 | if (ret) | |
5890 | break; | |
5891 | } | |
229a7b63 | 5892 | ret = io_connect(req, force_nonblock, cs); |
f8e85cf2 | 5893 | break; |
62755e35 | 5894 | case IORING_OP_ASYNC_CANCEL: |
3529d8c2 JA |
5895 | if (sqe) { |
5896 | ret = io_async_cancel_prep(req, sqe); | |
5897 | if (ret) | |
5898 | break; | |
5899 | } | |
014db007 | 5900 | ret = io_async_cancel(req); |
62755e35 | 5901 | break; |
d63d1b5e JA |
5902 | case IORING_OP_FALLOCATE: |
5903 | if (sqe) { | |
5904 | ret = io_fallocate_prep(req, sqe); | |
5905 | if (ret) | |
5906 | break; | |
5907 | } | |
014db007 | 5908 | ret = io_fallocate(req, force_nonblock); |
d63d1b5e | 5909 | break; |
15b71abe JA |
5910 | case IORING_OP_OPENAT: |
5911 | if (sqe) { | |
5912 | ret = io_openat_prep(req, sqe); | |
5913 | if (ret) | |
5914 | break; | |
5915 | } | |
014db007 | 5916 | ret = io_openat(req, force_nonblock); |
15b71abe | 5917 | break; |
b5dba59e JA |
5918 | case IORING_OP_CLOSE: |
5919 | if (sqe) { | |
5920 | ret = io_close_prep(req, sqe); | |
5921 | if (ret) | |
5922 | break; | |
5923 | } | |
229a7b63 | 5924 | ret = io_close(req, force_nonblock, cs); |
b5dba59e | 5925 | break; |
05f3fb3c JA |
5926 | case IORING_OP_FILES_UPDATE: |
5927 | if (sqe) { | |
5928 | ret = io_files_update_prep(req, sqe); | |
5929 | if (ret) | |
5930 | break; | |
5931 | } | |
229a7b63 | 5932 | ret = io_files_update(req, force_nonblock, cs); |
05f3fb3c | 5933 | break; |
eddc7ef5 JA |
5934 | case IORING_OP_STATX: |
5935 | if (sqe) { | |
5936 | ret = io_statx_prep(req, sqe); | |
5937 | if (ret) | |
5938 | break; | |
5939 | } | |
014db007 | 5940 | ret = io_statx(req, force_nonblock); |
eddc7ef5 | 5941 | break; |
4840e418 JA |
5942 | case IORING_OP_FADVISE: |
5943 | if (sqe) { | |
5944 | ret = io_fadvise_prep(req, sqe); | |
5945 | if (ret) | |
5946 | break; | |
5947 | } | |
014db007 | 5948 | ret = io_fadvise(req, force_nonblock); |
4840e418 | 5949 | break; |
c1ca757b JA |
5950 | case IORING_OP_MADVISE: |
5951 | if (sqe) { | |
5952 | ret = io_madvise_prep(req, sqe); | |
5953 | if (ret) | |
5954 | break; | |
5955 | } | |
014db007 | 5956 | ret = io_madvise(req, force_nonblock); |
c1ca757b | 5957 | break; |
cebdb986 JA |
5958 | case IORING_OP_OPENAT2: |
5959 | if (sqe) { | |
5960 | ret = io_openat2_prep(req, sqe); | |
5961 | if (ret) | |
5962 | break; | |
5963 | } | |
014db007 | 5964 | ret = io_openat2(req, force_nonblock); |
cebdb986 | 5965 | break; |
3e4827b0 JA |
5966 | case IORING_OP_EPOLL_CTL: |
5967 | if (sqe) { | |
5968 | ret = io_epoll_ctl_prep(req, sqe); | |
5969 | if (ret) | |
5970 | break; | |
5971 | } | |
229a7b63 | 5972 | ret = io_epoll_ctl(req, force_nonblock, cs); |
3e4827b0 | 5973 | break; |
7d67af2c PB |
5974 | case IORING_OP_SPLICE: |
5975 | if (sqe) { | |
5976 | ret = io_splice_prep(req, sqe); | |
5977 | if (ret < 0) | |
5978 | break; | |
5979 | } | |
014db007 | 5980 | ret = io_splice(req, force_nonblock); |
7d67af2c | 5981 | break; |
ddf0322d JA |
5982 | case IORING_OP_PROVIDE_BUFFERS: |
5983 | if (sqe) { | |
5984 | ret = io_provide_buffers_prep(req, sqe); | |
5985 | if (ret) | |
5986 | break; | |
5987 | } | |
229a7b63 | 5988 | ret = io_provide_buffers(req, force_nonblock, cs); |
ddf0322d | 5989 | break; |
067524e9 JA |
5990 | case IORING_OP_REMOVE_BUFFERS: |
5991 | if (sqe) { | |
5992 | ret = io_remove_buffers_prep(req, sqe); | |
5993 | if (ret) | |
5994 | break; | |
5995 | } | |
229a7b63 | 5996 | ret = io_remove_buffers(req, force_nonblock, cs); |
3e4827b0 | 5997 | break; |
f2a8d5c7 PB |
5998 | case IORING_OP_TEE: |
5999 | if (sqe) { | |
6000 | ret = io_tee_prep(req, sqe); | |
6001 | if (ret < 0) | |
6002 | break; | |
6003 | } | |
6004 | ret = io_tee(req, force_nonblock); | |
6005 | break; | |
2b188cc1 JA |
6006 | default: |
6007 | ret = -EINVAL; | |
6008 | break; | |
6009 | } | |
6010 | ||
def596e9 JA |
6011 | if (ret) |
6012 | return ret; | |
6013 | ||
b532576e JA |
6014 | /* If the op doesn't have a file, we're not polling for it */ |
6015 | if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) { | |
11ba820b JA |
6016 | const bool in_async = io_wq_current_is_worker(); |
6017 | ||
11ba820b JA |
6018 | /* workqueue context doesn't hold uring_lock, grab it now */ |
6019 | if (in_async) | |
6020 | mutex_lock(&ctx->uring_lock); | |
6021 | ||
def596e9 | 6022 | io_iopoll_req_issued(req); |
11ba820b JA |
6023 | |
6024 | if (in_async) | |
6025 | mutex_unlock(&ctx->uring_lock); | |
def596e9 JA |
6026 | } |
6027 | ||
6028 | return 0; | |
2b188cc1 JA |
6029 | } |
6030 | ||
f4db7182 | 6031 | static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work) |
2b188cc1 JA |
6032 | { |
6033 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
6df1db6b | 6034 | struct io_kiocb *timeout; |
561fb04a | 6035 | int ret = 0; |
2b188cc1 | 6036 | |
6df1db6b PB |
6037 | timeout = io_prep_linked_timeout(req); |
6038 | if (timeout) | |
6039 | io_queue_linked_timeout(timeout); | |
d4c81f38 | 6040 | |
0c9d5ccd JA |
6041 | /* if NO_CANCEL is set, we must still run the work */ |
6042 | if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) == | |
6043 | IO_WQ_WORK_CANCEL) { | |
561fb04a | 6044 | ret = -ECANCELED; |
0c9d5ccd | 6045 | } |
31b51510 | 6046 | |
561fb04a | 6047 | if (!ret) { |
561fb04a | 6048 | do { |
f13fad7b | 6049 | ret = io_issue_sqe(req, NULL, false, NULL); |
561fb04a JA |
6050 | /* |
6051 | * We can get EAGAIN for polled IO even though we're | |
6052 | * forcing a sync submission from here, since we can't | |
6053 | * wait for request slots on the block side. | |
6054 | */ | |
6055 | if (ret != -EAGAIN) | |
6056 | break; | |
6057 | cond_resched(); | |
6058 | } while (1); | |
6059 | } | |
31b51510 | 6060 | |
561fb04a | 6061 | if (ret) { |
4e88d6e7 | 6062 | req_set_fail_links(req); |
e1e16097 | 6063 | io_req_complete(req, ret); |
edafccee | 6064 | } |
2b188cc1 | 6065 | |
f4db7182 | 6066 | return io_steal_work(req); |
2b188cc1 JA |
6067 | } |
6068 | ||
65e19f54 JA |
6069 | static inline struct file *io_file_from_index(struct io_ring_ctx *ctx, |
6070 | int index) | |
6071 | { | |
6072 | struct fixed_file_table *table; | |
6073 | ||
05f3fb3c | 6074 | table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT]; |
84695089 | 6075 | return table->files[index & IORING_FILE_TABLE_MASK]; |
65e19f54 JA |
6076 | } |
6077 | ||
8da11c19 PB |
6078 | static int io_file_get(struct io_submit_state *state, struct io_kiocb *req, |
6079 | int fd, struct file **out_file, bool fixed) | |
09bb8394 | 6080 | { |
a197f664 | 6081 | struct io_ring_ctx *ctx = req->ctx; |
8da11c19 | 6082 | struct file *file; |
09bb8394 | 6083 | |
8da11c19 | 6084 | if (fixed) { |
05f3fb3c | 6085 | if (unlikely(!ctx->file_data || |
09bb8394 JA |
6086 | (unsigned) fd >= ctx->nr_user_files)) |
6087 | return -EBADF; | |
b7620121 | 6088 | fd = array_index_nospec(fd, ctx->nr_user_files); |
8da11c19 | 6089 | file = io_file_from_index(ctx, fd); |
fd2206e4 JA |
6090 | if (file) { |
6091 | req->fixed_file_refs = ctx->file_data->cur_refs; | |
6092 | percpu_ref_get(req->fixed_file_refs); | |
6093 | } | |
09bb8394 | 6094 | } else { |
c826bd7a | 6095 | trace_io_uring_file_get(ctx, fd); |
8da11c19 | 6096 | file = __io_file_get(state, fd); |
09bb8394 JA |
6097 | } |
6098 | ||
fd2206e4 JA |
6099 | if (file || io_op_defs[req->opcode].needs_file_no_error) { |
6100 | *out_file = file; | |
6101 | return 0; | |
6102 | } | |
6103 | return -EBADF; | |
09bb8394 JA |
6104 | } |
6105 | ||
8da11c19 | 6106 | static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req, |
63ff8223 | 6107 | int fd) |
8da11c19 | 6108 | { |
8da11c19 PB |
6109 | bool fixed; |
6110 | ||
63ff8223 | 6111 | fixed = (req->flags & REQ_F_FIXED_FILE) != 0; |
0cdaf760 | 6112 | if (unlikely(!fixed && io_async_submit(req->ctx))) |
8da11c19 PB |
6113 | return -EBADF; |
6114 | ||
6115 | return io_file_get(state, req, fd, &req->file, fixed); | |
6116 | } | |
6117 | ||
a197f664 | 6118 | static int io_grab_files(struct io_kiocb *req) |
fcb323cc | 6119 | { |
a197f664 | 6120 | struct io_ring_ctx *ctx = req->ctx; |
fcb323cc | 6121 | |
f56040b8 PB |
6122 | io_req_init_async(req); |
6123 | ||
5b0bbee4 | 6124 | if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE)) |
f86cd20c | 6125 | return 0; |
b5dba59e | 6126 | |
0f212204 | 6127 | req->work.files = get_files_struct(current); |
9b828492 JA |
6128 | get_nsproxy(current->nsproxy); |
6129 | req->work.nsproxy = current->nsproxy; | |
0f212204 JA |
6130 | req->flags |= REQ_F_INFLIGHT; |
6131 | ||
fcb323cc | 6132 | spin_lock_irq(&ctx->inflight_lock); |
0f212204 | 6133 | list_add(&req->inflight_entry, &ctx->inflight_list); |
fcb323cc | 6134 | spin_unlock_irq(&ctx->inflight_lock); |
0f212204 | 6135 | return 0; |
fcb323cc JA |
6136 | } |
6137 | ||
f56040b8 PB |
6138 | static inline int io_prep_work_files(struct io_kiocb *req) |
6139 | { | |
6140 | if (!io_op_defs[req->opcode].file_table) | |
6141 | return 0; | |
6142 | return io_grab_files(req); | |
6143 | } | |
6144 | ||
2665abfd | 6145 | static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer) |
2b188cc1 | 6146 | { |
ad8a48ac JA |
6147 | struct io_timeout_data *data = container_of(timer, |
6148 | struct io_timeout_data, timer); | |
6149 | struct io_kiocb *req = data->req; | |
2665abfd JA |
6150 | struct io_ring_ctx *ctx = req->ctx; |
6151 | struct io_kiocb *prev = NULL; | |
6152 | unsigned long flags; | |
2665abfd JA |
6153 | |
6154 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
6155 | ||
6156 | /* | |
6157 | * We don't expect the list to be empty, that will only happen if we | |
6158 | * race with the completion of the linked work. | |
6159 | */ | |
4493233e PB |
6160 | if (!list_empty(&req->link_list)) { |
6161 | prev = list_entry(req->link_list.prev, struct io_kiocb, | |
6162 | link_list); | |
5d960724 | 6163 | if (refcount_inc_not_zero(&prev->refs)) { |
4493233e | 6164 | list_del_init(&req->link_list); |
5d960724 JA |
6165 | prev->flags &= ~REQ_F_LINK_TIMEOUT; |
6166 | } else | |
76a46e06 | 6167 | prev = NULL; |
2665abfd JA |
6168 | } |
6169 | ||
6170 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
6171 | ||
6172 | if (prev) { | |
4e88d6e7 | 6173 | req_set_fail_links(prev); |
014db007 | 6174 | io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME); |
76a46e06 | 6175 | io_put_req(prev); |
47f46768 | 6176 | } else { |
e1e16097 | 6177 | io_req_complete(req, -ETIME); |
2665abfd | 6178 | } |
2665abfd JA |
6179 | return HRTIMER_NORESTART; |
6180 | } | |
6181 | ||
7271ef3a | 6182 | static void __io_queue_linked_timeout(struct io_kiocb *req) |
2665abfd | 6183 | { |
76a46e06 JA |
6184 | /* |
6185 | * If the list is now empty, then our linked request finished before | |
6186 | * we got a chance to setup the timer | |
6187 | */ | |
4493233e | 6188 | if (!list_empty(&req->link_list)) { |
2d28390a | 6189 | struct io_timeout_data *data = &req->io->timeout; |
94ae5e77 | 6190 | |
ad8a48ac JA |
6191 | data->timer.function = io_link_timeout_fn; |
6192 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), | |
6193 | data->mode); | |
2665abfd | 6194 | } |
7271ef3a JA |
6195 | } |
6196 | ||
6197 | static void io_queue_linked_timeout(struct io_kiocb *req) | |
6198 | { | |
6199 | struct io_ring_ctx *ctx = req->ctx; | |
6200 | ||
6201 | spin_lock_irq(&ctx->completion_lock); | |
6202 | __io_queue_linked_timeout(req); | |
76a46e06 | 6203 | spin_unlock_irq(&ctx->completion_lock); |
2665abfd | 6204 | |
2665abfd | 6205 | /* drop submission reference */ |
76a46e06 JA |
6206 | io_put_req(req); |
6207 | } | |
2665abfd | 6208 | |
ad8a48ac | 6209 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) |
2665abfd JA |
6210 | { |
6211 | struct io_kiocb *nxt; | |
6212 | ||
dea3b49c | 6213 | if (!(req->flags & REQ_F_LINK_HEAD)) |
2665abfd | 6214 | return NULL; |
6df1db6b | 6215 | if (req->flags & REQ_F_LINK_TIMEOUT) |
d7718a9d | 6216 | return NULL; |
2665abfd | 6217 | |
4493233e PB |
6218 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, |
6219 | link_list); | |
d625c6ee | 6220 | if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT) |
76a46e06 | 6221 | return NULL; |
2665abfd | 6222 | |
76a46e06 | 6223 | req->flags |= REQ_F_LINK_TIMEOUT; |
76a46e06 | 6224 | return nxt; |
2665abfd JA |
6225 | } |
6226 | ||
f13fad7b JA |
6227 | static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
6228 | struct io_comp_state *cs) | |
2b188cc1 | 6229 | { |
4a0a7a18 | 6230 | struct io_kiocb *linked_timeout; |
4bc4494e | 6231 | struct io_kiocb *nxt; |
193155c8 | 6232 | const struct cred *old_creds = NULL; |
e0c5c576 | 6233 | int ret; |
2b188cc1 | 6234 | |
4a0a7a18 JA |
6235 | again: |
6236 | linked_timeout = io_prep_linked_timeout(req); | |
6237 | ||
7cdaf587 XW |
6238 | if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds && |
6239 | req->work.creds != current_cred()) { | |
193155c8 JA |
6240 | if (old_creds) |
6241 | revert_creds(old_creds); | |
6242 | if (old_creds == req->work.creds) | |
6243 | old_creds = NULL; /* restored original creds */ | |
6244 | else | |
6245 | old_creds = override_creds(req->work.creds); | |
6246 | } | |
6247 | ||
f13fad7b | 6248 | ret = io_issue_sqe(req, sqe, true, cs); |
491381ce JA |
6249 | |
6250 | /* | |
6251 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
6252 | * doesn't support non-blocking read/write attempts | |
6253 | */ | |
24c74678 | 6254 | if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) { |
f063c547 | 6255 | if (!io_arm_poll_handler(req)) { |
86a761f8 | 6256 | punt: |
f063c547 PB |
6257 | ret = io_prep_work_files(req); |
6258 | if (unlikely(ret)) | |
bbad27b2 | 6259 | goto err; |
f063c547 PB |
6260 | /* |
6261 | * Queued up for async execution, worker will release | |
6262 | * submit reference when the iocb is actually submitted. | |
6263 | */ | |
6264 | io_queue_async_work(req); | |
2b188cc1 | 6265 | } |
bbad27b2 | 6266 | |
f063c547 PB |
6267 | if (linked_timeout) |
6268 | io_queue_linked_timeout(linked_timeout); | |
4bc4494e | 6269 | goto exit; |
2b188cc1 | 6270 | } |
e65ef56d | 6271 | |
652532ad | 6272 | if (unlikely(ret)) { |
fcb323cc | 6273 | err: |
652532ad PB |
6274 | /* un-prep timeout, so it'll be killed as any other linked */ |
6275 | req->flags &= ~REQ_F_LINK_TIMEOUT; | |
4e88d6e7 | 6276 | req_set_fail_links(req); |
e65ef56d | 6277 | io_put_req(req); |
e1e16097 | 6278 | io_req_complete(req, ret); |
652532ad | 6279 | goto exit; |
9e645e11 | 6280 | } |
652532ad PB |
6281 | |
6282 | /* drop submission reference */ | |
6283 | nxt = io_put_req_find_next(req); | |
6284 | if (linked_timeout) | |
6285 | io_queue_linked_timeout(linked_timeout); | |
6286 | ||
4a0a7a18 JA |
6287 | if (nxt) { |
6288 | req = nxt; | |
86a761f8 PB |
6289 | |
6290 | if (req->flags & REQ_F_FORCE_ASYNC) | |
6291 | goto punt; | |
4a0a7a18 JA |
6292 | goto again; |
6293 | } | |
4bc4494e | 6294 | exit: |
193155c8 JA |
6295 | if (old_creds) |
6296 | revert_creds(old_creds); | |
2b188cc1 JA |
6297 | } |
6298 | ||
f13fad7b JA |
6299 | static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
6300 | struct io_comp_state *cs) | |
4fe2c963 JL |
6301 | { |
6302 | int ret; | |
6303 | ||
3529d8c2 | 6304 | ret = io_req_defer(req, sqe); |
4fe2c963 JL |
6305 | if (ret) { |
6306 | if (ret != -EIOCBQUEUED) { | |
1118591a | 6307 | fail_req: |
4e88d6e7 | 6308 | req_set_fail_links(req); |
e1e16097 JA |
6309 | io_put_req(req); |
6310 | io_req_complete(req, ret); | |
4fe2c963 | 6311 | } |
2550878f | 6312 | } else if (req->flags & REQ_F_FORCE_ASYNC) { |
bd2ab18a | 6313 | if (!req->io) { |
bd2ab18a | 6314 | ret = io_req_defer_prep(req, sqe); |
327d6d96 | 6315 | if (unlikely(ret)) |
bd2ab18a PB |
6316 | goto fail_req; |
6317 | } | |
6318 | ||
ce35a47a JA |
6319 | /* |
6320 | * Never try inline submit of IOSQE_ASYNC is set, go straight | |
6321 | * to async execution. | |
6322 | */ | |
3e863ea3 | 6323 | io_req_init_async(req); |
ce35a47a JA |
6324 | req->work.flags |= IO_WQ_WORK_CONCURRENT; |
6325 | io_queue_async_work(req); | |
6326 | } else { | |
f13fad7b | 6327 | __io_queue_sqe(req, sqe, cs); |
ce35a47a | 6328 | } |
4fe2c963 JL |
6329 | } |
6330 | ||
f13fad7b JA |
6331 | static inline void io_queue_link_head(struct io_kiocb *req, |
6332 | struct io_comp_state *cs) | |
4fe2c963 | 6333 | { |
94ae5e77 | 6334 | if (unlikely(req->flags & REQ_F_FAIL_LINK)) { |
e1e16097 JA |
6335 | io_put_req(req); |
6336 | io_req_complete(req, -ECANCELED); | |
1b4a51b6 | 6337 | } else |
f13fad7b | 6338 | io_queue_sqe(req, NULL, cs); |
4fe2c963 JL |
6339 | } |
6340 | ||
1d4240cc | 6341 | static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
f13fad7b | 6342 | struct io_kiocb **link, struct io_comp_state *cs) |
9e645e11 | 6343 | { |
a197f664 | 6344 | struct io_ring_ctx *ctx = req->ctx; |
ef4ff581 | 6345 | int ret; |
9e645e11 | 6346 | |
9e645e11 JA |
6347 | /* |
6348 | * If we already have a head request, queue this one for async | |
6349 | * submittal once the head completes. If we don't have a head but | |
6350 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
6351 | * submitted sync once the chain is complete. If none of those | |
6352 | * conditions are true (normal request), then just queue it. | |
6353 | */ | |
6354 | if (*link) { | |
9d76377f | 6355 | struct io_kiocb *head = *link; |
4e88d6e7 | 6356 | |
8cdf2193 PB |
6357 | /* |
6358 | * Taking sequential execution of a link, draining both sides | |
6359 | * of the link also fullfils IOSQE_IO_DRAIN semantics for all | |
6360 | * requests in the link. So, it drains the head and the | |
6361 | * next after the link request. The last one is done via | |
6362 | * drain_next flag to persist the effect across calls. | |
6363 | */ | |
ef4ff581 | 6364 | if (req->flags & REQ_F_IO_DRAIN) { |
711be031 PB |
6365 | head->flags |= REQ_F_IO_DRAIN; |
6366 | ctx->drain_next = 1; | |
6367 | } | |
3529d8c2 | 6368 | ret = io_req_defer_prep(req, sqe); |
327d6d96 | 6369 | if (unlikely(ret)) { |
4e88d6e7 | 6370 | /* fail even hard links since we don't submit */ |
9d76377f | 6371 | head->flags |= REQ_F_FAIL_LINK; |
1d4240cc | 6372 | return ret; |
2d28390a | 6373 | } |
9d76377f PB |
6374 | trace_io_uring_link(ctx, req, head); |
6375 | list_add_tail(&req->link_list, &head->link_list); | |
32fe525b PB |
6376 | |
6377 | /* last request of a link, enqueue the link */ | |
ef4ff581 | 6378 | if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) { |
f13fad7b | 6379 | io_queue_link_head(head, cs); |
32fe525b PB |
6380 | *link = NULL; |
6381 | } | |
9e645e11 | 6382 | } else { |
711be031 PB |
6383 | if (unlikely(ctx->drain_next)) { |
6384 | req->flags |= REQ_F_IO_DRAIN; | |
ef4ff581 | 6385 | ctx->drain_next = 0; |
711be031 | 6386 | } |
ef4ff581 | 6387 | if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) { |
dea3b49c | 6388 | req->flags |= REQ_F_LINK_HEAD; |
711be031 | 6389 | INIT_LIST_HEAD(&req->link_list); |
f1d96a8f | 6390 | |
711be031 | 6391 | ret = io_req_defer_prep(req, sqe); |
327d6d96 | 6392 | if (unlikely(ret)) |
711be031 PB |
6393 | req->flags |= REQ_F_FAIL_LINK; |
6394 | *link = req; | |
6395 | } else { | |
f13fad7b | 6396 | io_queue_sqe(req, sqe, cs); |
711be031 | 6397 | } |
9e645e11 | 6398 | } |
2e6e1fde | 6399 | |
1d4240cc | 6400 | return 0; |
9e645e11 JA |
6401 | } |
6402 | ||
9a56a232 JA |
6403 | /* |
6404 | * Batched submission is done, ensure local IO is flushed out. | |
6405 | */ | |
6406 | static void io_submit_state_end(struct io_submit_state *state) | |
6407 | { | |
f13fad7b JA |
6408 | if (!list_empty(&state->comp.list)) |
6409 | io_submit_flush_completions(&state->comp); | |
9a56a232 | 6410 | blk_finish_plug(&state->plug); |
9f13c35b | 6411 | io_state_file_put(state); |
2579f913 | 6412 | if (state->free_reqs) |
6c8a3134 | 6413 | kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs); |
9a56a232 JA |
6414 | } |
6415 | ||
6416 | /* | |
6417 | * Start submission side cache. | |
6418 | */ | |
6419 | static void io_submit_state_start(struct io_submit_state *state, | |
013538bd | 6420 | struct io_ring_ctx *ctx, unsigned int max_ios) |
9a56a232 JA |
6421 | { |
6422 | blk_start_plug(&state->plug); | |
013538bd JA |
6423 | state->comp.nr = 0; |
6424 | INIT_LIST_HEAD(&state->comp.list); | |
6425 | state->comp.ctx = ctx; | |
2579f913 | 6426 | state->free_reqs = 0; |
9a56a232 JA |
6427 | state->file = NULL; |
6428 | state->ios_left = max_ios; | |
6429 | } | |
6430 | ||
2b188cc1 JA |
6431 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
6432 | { | |
75b28aff | 6433 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 6434 | |
caf582c6 PB |
6435 | /* |
6436 | * Ensure any loads from the SQEs are done at this point, | |
6437 | * since once we write the new head, the application could | |
6438 | * write new data to them. | |
6439 | */ | |
6440 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
6441 | } |
6442 | ||
2b188cc1 | 6443 | /* |
3529d8c2 | 6444 | * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory |
2b188cc1 JA |
6445 | * that is mapped by userspace. This means that care needs to be taken to |
6446 | * ensure that reads are stable, as we cannot rely on userspace always | |
6447 | * being a good citizen. If members of the sqe are validated and then later | |
6448 | * used, it's important that those reads are done through READ_ONCE() to | |
6449 | * prevent a re-load down the line. | |
6450 | */ | |
709b302f | 6451 | static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx) |
2b188cc1 | 6452 | { |
75b28aff | 6453 | u32 *sq_array = ctx->sq_array; |
2b188cc1 JA |
6454 | unsigned head; |
6455 | ||
6456 | /* | |
6457 | * The cached sq head (or cq tail) serves two purposes: | |
6458 | * | |
6459 | * 1) allows us to batch the cost of updating the user visible | |
6460 | * head updates. | |
6461 | * 2) allows the kernel side to track the head on its own, even | |
6462 | * though the application is the one updating it. | |
6463 | */ | |
ee7d46d9 | 6464 | head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]); |
709b302f PB |
6465 | if (likely(head < ctx->sq_entries)) |
6466 | return &ctx->sq_sqes[head]; | |
2b188cc1 JA |
6467 | |
6468 | /* drop invalid entries */ | |
498ccd9e | 6469 | ctx->cached_sq_dropped++; |
ee7d46d9 | 6470 | WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped); |
709b302f PB |
6471 | return NULL; |
6472 | } | |
6473 | ||
6474 | static inline void io_consume_sqe(struct io_ring_ctx *ctx) | |
6475 | { | |
6476 | ctx->cached_sq_head++; | |
2b188cc1 JA |
6477 | } |
6478 | ||
21b55dbc SG |
6479 | /* |
6480 | * Check SQE restrictions (opcode and flags). | |
6481 | * | |
6482 | * Returns 'true' if SQE is allowed, 'false' otherwise. | |
6483 | */ | |
6484 | static inline bool io_check_restriction(struct io_ring_ctx *ctx, | |
6485 | struct io_kiocb *req, | |
6486 | unsigned int sqe_flags) | |
6487 | { | |
6488 | if (!ctx->restricted) | |
6489 | return true; | |
6490 | ||
6491 | if (!test_bit(req->opcode, ctx->restrictions.sqe_op)) | |
6492 | return false; | |
6493 | ||
6494 | if ((sqe_flags & ctx->restrictions.sqe_flags_required) != | |
6495 | ctx->restrictions.sqe_flags_required) | |
6496 | return false; | |
6497 | ||
6498 | if (sqe_flags & ~(ctx->restrictions.sqe_flags_allowed | | |
6499 | ctx->restrictions.sqe_flags_required)) | |
6500 | return false; | |
6501 | ||
6502 | return true; | |
6503 | } | |
6504 | ||
ef4ff581 PB |
6505 | #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \ |
6506 | IOSQE_IO_HARDLINK | IOSQE_ASYNC | \ | |
6507 | IOSQE_BUFFER_SELECT) | |
6508 | ||
6509 | static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, | |
6510 | const struct io_uring_sqe *sqe, | |
0cdaf760 | 6511 | struct io_submit_state *state) |
0553b8bd | 6512 | { |
ef4ff581 | 6513 | unsigned int sqe_flags; |
63ff8223 | 6514 | int id; |
ef4ff581 | 6515 | |
0553b8bd PB |
6516 | req->opcode = READ_ONCE(sqe->opcode); |
6517 | req->user_data = READ_ONCE(sqe->user_data); | |
6518 | req->io = NULL; | |
6519 | req->file = NULL; | |
6520 | req->ctx = ctx; | |
6521 | req->flags = 0; | |
6522 | /* one is dropped after submission, the other at completion */ | |
6523 | refcount_set(&req->refs, 2); | |
4dd2824d | 6524 | req->task = current; |
e3bc8e9d | 6525 | get_task_struct(req->task); |
0f212204 | 6526 | atomic_long_inc(&req->task->io_uring->req_issue); |
0553b8bd | 6527 | req->result = 0; |
ef4ff581 PB |
6528 | |
6529 | if (unlikely(req->opcode >= IORING_OP_LAST)) | |
6530 | return -EINVAL; | |
6531 | ||
9d8426a0 JA |
6532 | if (unlikely(io_sq_thread_acquire_mm(ctx, req))) |
6533 | return -EFAULT; | |
ef4ff581 PB |
6534 | |
6535 | sqe_flags = READ_ONCE(sqe->flags); | |
6536 | /* enforce forwards compatibility on users */ | |
6537 | if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) | |
6538 | return -EINVAL; | |
6539 | ||
21b55dbc SG |
6540 | if (unlikely(!io_check_restriction(ctx, req, sqe_flags))) |
6541 | return -EACCES; | |
6542 | ||
ef4ff581 PB |
6543 | if ((sqe_flags & IOSQE_BUFFER_SELECT) && |
6544 | !io_op_defs[req->opcode].buffer_select) | |
6545 | return -EOPNOTSUPP; | |
6546 | ||
6547 | id = READ_ONCE(sqe->personality); | |
6548 | if (id) { | |
7cdaf587 | 6549 | io_req_init_async(req); |
ef4ff581 PB |
6550 | req->work.creds = idr_find(&ctx->personality_idr, id); |
6551 | if (unlikely(!req->work.creds)) | |
6552 | return -EINVAL; | |
6553 | get_cred(req->work.creds); | |
6554 | } | |
6555 | ||
6556 | /* same numerical values with corresponding REQ_F_*, safe to copy */ | |
c11368a5 | 6557 | req->flags |= sqe_flags; |
ef4ff581 | 6558 | |
63ff8223 JA |
6559 | if (!io_op_defs[req->opcode].needs_file) |
6560 | return 0; | |
6561 | ||
6562 | return io_req_set_file(state, req, READ_ONCE(sqe->fd)); | |
0553b8bd PB |
6563 | } |
6564 | ||
0f212204 | 6565 | static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr) |
6c271ce2 | 6566 | { |
ac8691c4 | 6567 | struct io_submit_state state; |
9e645e11 | 6568 | struct io_kiocb *link = NULL; |
9e645e11 | 6569 | int i, submitted = 0; |
6c271ce2 | 6570 | |
c4a2ed72 | 6571 | /* if we have a backlog and couldn't flush it all, return BUSY */ |
ad3eb2c8 JA |
6572 | if (test_bit(0, &ctx->sq_check_overflow)) { |
6573 | if (!list_empty(&ctx->cq_overflow_list) && | |
e6c8aa9a | 6574 | !io_cqring_overflow_flush(ctx, false, NULL, NULL)) |
ad3eb2c8 JA |
6575 | return -EBUSY; |
6576 | } | |
6c271ce2 | 6577 | |
ee7d46d9 PB |
6578 | /* make sure SQ entry isn't read before tail */ |
6579 | nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx)); | |
9ef4f124 | 6580 | |
2b85edfc PB |
6581 | if (!percpu_ref_tryget_many(&ctx->refs, nr)) |
6582 | return -EAGAIN; | |
6c271ce2 | 6583 | |
013538bd | 6584 | io_submit_state_start(&state, ctx, nr); |
6c271ce2 JA |
6585 | |
6586 | for (i = 0; i < nr; i++) { | |
3529d8c2 | 6587 | const struct io_uring_sqe *sqe; |
196be95c | 6588 | struct io_kiocb *req; |
1cb1edb2 | 6589 | int err; |
fb5ccc98 | 6590 | |
b1e50e54 PB |
6591 | sqe = io_get_sqe(ctx); |
6592 | if (unlikely(!sqe)) { | |
6593 | io_consume_sqe(ctx); | |
6594 | break; | |
6595 | } | |
ac8691c4 | 6596 | req = io_alloc_req(ctx, &state); |
196be95c PB |
6597 | if (unlikely(!req)) { |
6598 | if (!submitted) | |
6599 | submitted = -EAGAIN; | |
fb5ccc98 | 6600 | break; |
196be95c | 6601 | } |
fb5ccc98 | 6602 | |
ac8691c4 | 6603 | err = io_init_req(ctx, req, sqe, &state); |
709b302f | 6604 | io_consume_sqe(ctx); |
d3656344 JA |
6605 | /* will complete beyond this point, count as submitted */ |
6606 | submitted++; | |
6607 | ||
ef4ff581 | 6608 | if (unlikely(err)) { |
1cb1edb2 | 6609 | fail_req: |
e1e16097 JA |
6610 | io_put_req(req); |
6611 | io_req_complete(req, err); | |
196be95c PB |
6612 | break; |
6613 | } | |
fb5ccc98 | 6614 | |
354420f7 | 6615 | trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data, |
0cdaf760 | 6616 | true, io_async_submit(ctx)); |
f13fad7b | 6617 | err = io_submit_sqe(req, sqe, &link, &state.comp); |
1d4240cc PB |
6618 | if (err) |
6619 | goto fail_req; | |
6c271ce2 JA |
6620 | } |
6621 | ||
9466f437 PB |
6622 | if (unlikely(submitted != nr)) { |
6623 | int ref_used = (submitted == -EAGAIN) ? 0 : submitted; | |
6624 | ||
6625 | percpu_ref_put_many(&ctx->refs, nr - ref_used); | |
6626 | } | |
9e645e11 | 6627 | if (link) |
f13fad7b | 6628 | io_queue_link_head(link, &state.comp); |
ac8691c4 | 6629 | io_submit_state_end(&state); |
6c271ce2 | 6630 | |
ae9428ca PB |
6631 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
6632 | io_commit_sqring(ctx); | |
6633 | ||
6c271ce2 JA |
6634 | return submitted; |
6635 | } | |
6636 | ||
23b3628e XW |
6637 | static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx) |
6638 | { | |
6639 | /* Tell userspace we may need a wakeup call */ | |
6640 | spin_lock_irq(&ctx->completion_lock); | |
6641 | ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP; | |
6642 | spin_unlock_irq(&ctx->completion_lock); | |
6643 | } | |
6644 | ||
6645 | static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx) | |
6646 | { | |
6647 | spin_lock_irq(&ctx->completion_lock); | |
6648 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; | |
6649 | spin_unlock_irq(&ctx->completion_lock); | |
6650 | } | |
6651 | ||
3f0e64d0 JA |
6652 | static int io_sq_wake_function(struct wait_queue_entry *wqe, unsigned mode, |
6653 | int sync, void *key) | |
6654 | { | |
6655 | struct io_ring_ctx *ctx = container_of(wqe, struct io_ring_ctx, sqo_wait_entry); | |
6656 | int ret; | |
6657 | ||
6658 | ret = autoremove_wake_function(wqe, mode, sync, key); | |
6659 | if (ret) { | |
6660 | unsigned long flags; | |
6661 | ||
6662 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
6663 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; | |
6664 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
6665 | } | |
6666 | return ret; | |
6667 | } | |
6668 | ||
c8d1ba58 JA |
6669 | enum sq_ret { |
6670 | SQT_IDLE = 1, | |
6671 | SQT_SPIN = 2, | |
6672 | SQT_DID_WORK = 4, | |
6673 | }; | |
6674 | ||
6675 | static enum sq_ret __io_sq_thread(struct io_ring_ctx *ctx, | |
6676 | unsigned long start_jiffies) | |
6c271ce2 | 6677 | { |
c8d1ba58 | 6678 | unsigned long timeout = start_jiffies + ctx->sq_thread_idle; |
534ca6d6 | 6679 | struct io_sq_data *sqd = ctx->sq_data; |
c8d1ba58 | 6680 | unsigned int to_submit; |
bdcd3eab | 6681 | int ret = 0; |
6c271ce2 | 6682 | |
c8d1ba58 JA |
6683 | again: |
6684 | if (!list_empty(&ctx->iopoll_list)) { | |
6685 | unsigned nr_events = 0; | |
6a779382 | 6686 | |
c8d1ba58 JA |
6687 | mutex_lock(&ctx->uring_lock); |
6688 | if (!list_empty(&ctx->iopoll_list) && !need_resched()) | |
6689 | io_do_iopoll(ctx, &nr_events, 0); | |
6690 | mutex_unlock(&ctx->uring_lock); | |
6691 | } | |
a4c0b3de | 6692 | |
c8d1ba58 | 6693 | to_submit = io_sqring_entries(ctx); |
6c271ce2 | 6694 | |
c8d1ba58 JA |
6695 | /* |
6696 | * If submit got -EBUSY, flag us as needing the application | |
6697 | * to enter the kernel to reap and flush events. | |
6698 | */ | |
6699 | if (!to_submit || ret == -EBUSY || need_resched()) { | |
6700 | /* | |
6701 | * Drop cur_mm before scheduling, we can't hold it for | |
6702 | * long periods (or over schedule()). Do this before | |
6703 | * adding ourselves to the waitqueue, as the unuse/drop | |
6704 | * may sleep. | |
6705 | */ | |
6706 | io_sq_thread_drop_mm(); | |
6707 | ||
6708 | /* | |
6709 | * We're polling. If we're within the defined idle | |
6710 | * period, then let us spin without work before going | |
6711 | * to sleep. The exception is if we got EBUSY doing | |
6712 | * more IO, we should wait for the application to | |
6713 | * reap events and wake us up. | |
6714 | */ | |
6715 | if (!list_empty(&ctx->iopoll_list) || need_resched() || | |
6716 | (!time_after(jiffies, timeout) && ret != -EBUSY && | |
6717 | !percpu_ref_is_dying(&ctx->refs))) | |
6718 | return SQT_SPIN; | |
6c271ce2 | 6719 | |
534ca6d6 | 6720 | prepare_to_wait(&sqd->wait, &ctx->sqo_wait_entry, |
c8d1ba58 | 6721 | TASK_INTERRUPTIBLE); |
6c271ce2 | 6722 | |
c8d1ba58 JA |
6723 | /* |
6724 | * While doing polled IO, before going to sleep, we need | |
6725 | * to check if there are new reqs added to iopoll_list, | |
6726 | * it is because reqs may have been punted to io worker | |
6727 | * and will be added to iopoll_list later, hence check | |
6728 | * the iopoll_list again. | |
6729 | */ | |
6730 | if ((ctx->flags & IORING_SETUP_IOPOLL) && | |
6731 | !list_empty_careful(&ctx->iopoll_list)) { | |
534ca6d6 | 6732 | finish_wait(&sqd->wait, &ctx->sqo_wait_entry); |
c8d1ba58 | 6733 | goto again; |
6c271ce2 JA |
6734 | } |
6735 | ||
fb5ccc98 | 6736 | to_submit = io_sqring_entries(ctx); |
c8d1ba58 JA |
6737 | if (!to_submit || ret == -EBUSY) |
6738 | return SQT_IDLE; | |
6739 | } | |
c1edbf5f | 6740 | |
534ca6d6 | 6741 | finish_wait(&sqd->wait, &ctx->sqo_wait_entry); |
c8d1ba58 | 6742 | io_ring_clear_wakeup_flag(ctx); |
7143b5ac | 6743 | |
c8d1ba58 JA |
6744 | mutex_lock(&ctx->uring_lock); |
6745 | if (likely(!percpu_ref_is_dying(&ctx->refs))) | |
6746 | ret = io_submit_sqes(ctx, to_submit); | |
6747 | mutex_unlock(&ctx->uring_lock); | |
90554200 JA |
6748 | |
6749 | if (!io_sqring_full(ctx) && wq_has_sleeper(&ctx->sqo_sq_wait)) | |
6750 | wake_up(&ctx->sqo_sq_wait); | |
6751 | ||
c8d1ba58 JA |
6752 | return SQT_DID_WORK; |
6753 | } | |
6c271ce2 | 6754 | |
69fb2131 JA |
6755 | static void io_sqd_init_new(struct io_sq_data *sqd) |
6756 | { | |
6757 | struct io_ring_ctx *ctx; | |
6758 | ||
6759 | while (!list_empty(&sqd->ctx_new_list)) { | |
6760 | ctx = list_first_entry(&sqd->ctx_new_list, struct io_ring_ctx, sqd_list); | |
6761 | init_wait(&ctx->sqo_wait_entry); | |
6762 | ctx->sqo_wait_entry.func = io_sq_wake_function; | |
6763 | list_move_tail(&ctx->sqd_list, &sqd->ctx_list); | |
6764 | complete(&ctx->sq_thread_comp); | |
6765 | } | |
6766 | } | |
6767 | ||
c8d1ba58 JA |
6768 | static int io_sq_thread(void *data) |
6769 | { | |
69fb2131 JA |
6770 | const struct cred *old_cred = NULL; |
6771 | struct io_sq_data *sqd = data; | |
6772 | struct io_ring_ctx *ctx; | |
c8d1ba58 | 6773 | unsigned long start_jiffies; |
6c271ce2 | 6774 | |
69fb2131 JA |
6775 | start_jiffies = jiffies; |
6776 | while (!kthread_should_stop()) { | |
6777 | enum sq_ret ret = 0; | |
6c271ce2 | 6778 | |
69fb2131 JA |
6779 | /* |
6780 | * Any changes to the sqd lists are synchronized through the | |
6781 | * kthread parking. This synchronizes the thread vs users, | |
6782 | * the users are synchronized on the sqd->ctx_lock. | |
6783 | */ | |
6784 | if (kthread_should_park()) | |
6785 | kthread_parkme(); | |
6c271ce2 | 6786 | |
69fb2131 JA |
6787 | if (unlikely(!list_empty(&sqd->ctx_new_list))) |
6788 | io_sqd_init_new(sqd); | |
c8d1ba58 | 6789 | |
69fb2131 JA |
6790 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) { |
6791 | if (current->cred != ctx->creds) { | |
6792 | if (old_cred) | |
6793 | revert_creds(old_cred); | |
6794 | old_cred = override_creds(ctx->creds); | |
6795 | } | |
c8d1ba58 | 6796 | |
69fb2131 JA |
6797 | ret |= __io_sq_thread(ctx, start_jiffies); |
6798 | ||
6799 | io_sq_thread_drop_mm(); | |
6800 | } | |
6801 | ||
6802 | if (ret & SQT_SPIN) { | |
c8d1ba58 JA |
6803 | io_run_task_work(); |
6804 | cond_resched(); | |
69fb2131 JA |
6805 | } else if (ret == SQT_IDLE) { |
6806 | if (kthread_should_park()) | |
6807 | continue; | |
6808 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) | |
6809 | io_ring_set_wakeup_flag(ctx); | |
6810 | schedule(); | |
6811 | start_jiffies = jiffies; | |
6812 | list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) | |
6813 | io_ring_clear_wakeup_flag(ctx); | |
c8d1ba58 | 6814 | } |
6c271ce2 JA |
6815 | } |
6816 | ||
4c6e277c | 6817 | io_run_task_work(); |
b41e9852 | 6818 | |
69fb2131 JA |
6819 | if (old_cred) |
6820 | revert_creds(old_cred); | |
06058632 | 6821 | |
2bbcd6d3 | 6822 | kthread_parkme(); |
06058632 | 6823 | |
6c271ce2 JA |
6824 | return 0; |
6825 | } | |
6826 | ||
bda52162 JA |
6827 | struct io_wait_queue { |
6828 | struct wait_queue_entry wq; | |
6829 | struct io_ring_ctx *ctx; | |
6830 | unsigned to_wait; | |
6831 | unsigned nr_timeouts; | |
6832 | }; | |
6833 | ||
1d7bb1d5 | 6834 | static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush) |
bda52162 JA |
6835 | { |
6836 | struct io_ring_ctx *ctx = iowq->ctx; | |
6837 | ||
6838 | /* | |
d195a66e | 6839 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
6840 | * started waiting. For timeouts, we always want to return to userspace, |
6841 | * regardless of event count. | |
6842 | */ | |
1d7bb1d5 | 6843 | return io_cqring_events(ctx, noflush) >= iowq->to_wait || |
bda52162 JA |
6844 | atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
6845 | } | |
6846 | ||
6847 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
6848 | int wake_flags, void *key) | |
6849 | { | |
6850 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
6851 | wq); | |
6852 | ||
1d7bb1d5 JA |
6853 | /* use noflush == true, as we can't safely rely on locking context */ |
6854 | if (!io_should_wake(iowq, true)) | |
bda52162 JA |
6855 | return -1; |
6856 | ||
6857 | return autoremove_wake_function(curr, mode, wake_flags, key); | |
6858 | } | |
6859 | ||
2b188cc1 JA |
6860 | /* |
6861 | * Wait until events become available, if we don't already have some. The | |
6862 | * application must reap them itself, as they reside on the shared cq ring. | |
6863 | */ | |
6864 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
6865 | const sigset_t __user *sig, size_t sigsz) | |
6866 | { | |
bda52162 JA |
6867 | struct io_wait_queue iowq = { |
6868 | .wq = { | |
6869 | .private = current, | |
6870 | .func = io_wake_function, | |
6871 | .entry = LIST_HEAD_INIT(iowq.wq.entry), | |
6872 | }, | |
6873 | .ctx = ctx, | |
6874 | .to_wait = min_events, | |
6875 | }; | |
75b28aff | 6876 | struct io_rings *rings = ctx->rings; |
e9ffa5c2 | 6877 | int ret = 0; |
2b188cc1 | 6878 | |
b41e9852 JA |
6879 | do { |
6880 | if (io_cqring_events(ctx, false) >= min_events) | |
6881 | return 0; | |
4c6e277c | 6882 | if (!io_run_task_work()) |
b41e9852 | 6883 | break; |
b41e9852 | 6884 | } while (1); |
2b188cc1 JA |
6885 | |
6886 | if (sig) { | |
9e75ad5d AB |
6887 | #ifdef CONFIG_COMPAT |
6888 | if (in_compat_syscall()) | |
6889 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 6890 | sigsz); |
9e75ad5d AB |
6891 | else |
6892 | #endif | |
b772434b | 6893 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 6894 | |
2b188cc1 JA |
6895 | if (ret) |
6896 | return ret; | |
6897 | } | |
6898 | ||
bda52162 | 6899 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
c826bd7a | 6900 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 JA |
6901 | do { |
6902 | prepare_to_wait_exclusive(&ctx->wait, &iowq.wq, | |
6903 | TASK_INTERRUPTIBLE); | |
ce593a6c | 6904 | /* make sure we run task_work before checking for signals */ |
4c6e277c JA |
6905 | if (io_run_task_work()) |
6906 | continue; | |
bda52162 | 6907 | if (signal_pending(current)) { |
b7db41c9 JA |
6908 | if (current->jobctl & JOBCTL_TASK_WORK) { |
6909 | spin_lock_irq(¤t->sighand->siglock); | |
6910 | current->jobctl &= ~JOBCTL_TASK_WORK; | |
6911 | recalc_sigpending(); | |
6912 | spin_unlock_irq(¤t->sighand->siglock); | |
6913 | continue; | |
6914 | } | |
6915 | ret = -EINTR; | |
bda52162 JA |
6916 | break; |
6917 | } | |
ce593a6c JA |
6918 | if (io_should_wake(&iowq, false)) |
6919 | break; | |
6920 | schedule(); | |
bda52162 JA |
6921 | } while (1); |
6922 | finish_wait(&ctx->wait, &iowq.wq); | |
6923 | ||
b7db41c9 | 6924 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 6925 | |
75b28aff | 6926 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
6927 | } |
6928 | ||
6b06314c JA |
6929 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
6930 | { | |
6931 | #if defined(CONFIG_UNIX) | |
6932 | if (ctx->ring_sock) { | |
6933 | struct sock *sock = ctx->ring_sock->sk; | |
6934 | struct sk_buff *skb; | |
6935 | ||
6936 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
6937 | kfree_skb(skb); | |
6938 | } | |
6939 | #else | |
6940 | int i; | |
6941 | ||
65e19f54 JA |
6942 | for (i = 0; i < ctx->nr_user_files; i++) { |
6943 | struct file *file; | |
6944 | ||
6945 | file = io_file_from_index(ctx, i); | |
6946 | if (file) | |
6947 | fput(file); | |
6948 | } | |
6b06314c JA |
6949 | #endif |
6950 | } | |
6951 | ||
05f3fb3c JA |
6952 | static void io_file_ref_kill(struct percpu_ref *ref) |
6953 | { | |
6954 | struct fixed_file_data *data; | |
6955 | ||
6956 | data = container_of(ref, struct fixed_file_data, refs); | |
6957 | complete(&data->done); | |
6958 | } | |
6959 | ||
6b06314c JA |
6960 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) |
6961 | { | |
05f3fb3c | 6962 | struct fixed_file_data *data = ctx->file_data; |
05589553 | 6963 | struct fixed_file_ref_node *ref_node = NULL; |
65e19f54 JA |
6964 | unsigned nr_tables, i; |
6965 | ||
05f3fb3c | 6966 | if (!data) |
6b06314c JA |
6967 | return -ENXIO; |
6968 | ||
6a4d07cd | 6969 | spin_lock(&data->lock); |
05589553 XW |
6970 | if (!list_empty(&data->ref_list)) |
6971 | ref_node = list_first_entry(&data->ref_list, | |
6972 | struct fixed_file_ref_node, node); | |
6a4d07cd | 6973 | spin_unlock(&data->lock); |
05589553 XW |
6974 | if (ref_node) |
6975 | percpu_ref_kill(&ref_node->refs); | |
6976 | ||
6977 | percpu_ref_kill(&data->refs); | |
6978 | ||
6979 | /* wait for all refs nodes to complete */ | |
4a38aed2 | 6980 | flush_delayed_work(&ctx->file_put_work); |
2faf852d | 6981 | wait_for_completion(&data->done); |
05f3fb3c | 6982 | |
6b06314c | 6983 | __io_sqe_files_unregister(ctx); |
65e19f54 JA |
6984 | nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE); |
6985 | for (i = 0; i < nr_tables; i++) | |
05f3fb3c JA |
6986 | kfree(data->table[i].files); |
6987 | kfree(data->table); | |
05589553 XW |
6988 | percpu_ref_exit(&data->refs); |
6989 | kfree(data); | |
05f3fb3c | 6990 | ctx->file_data = NULL; |
6b06314c JA |
6991 | ctx->nr_user_files = 0; |
6992 | return 0; | |
6993 | } | |
6994 | ||
534ca6d6 | 6995 | static void io_put_sq_data(struct io_sq_data *sqd) |
6c271ce2 | 6996 | { |
534ca6d6 | 6997 | if (refcount_dec_and_test(&sqd->refs)) { |
2bbcd6d3 RP |
6998 | /* |
6999 | * The park is a bit of a work-around, without it we get | |
7000 | * warning spews on shutdown with SQPOLL set and affinity | |
7001 | * set to a single CPU. | |
7002 | */ | |
534ca6d6 JA |
7003 | if (sqd->thread) { |
7004 | kthread_park(sqd->thread); | |
7005 | kthread_stop(sqd->thread); | |
7006 | } | |
7007 | ||
7008 | kfree(sqd); | |
7009 | } | |
7010 | } | |
7011 | ||
aa06165d JA |
7012 | static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p) |
7013 | { | |
7014 | struct io_ring_ctx *ctx_attach; | |
7015 | struct io_sq_data *sqd; | |
7016 | struct fd f; | |
7017 | ||
7018 | f = fdget(p->wq_fd); | |
7019 | if (!f.file) | |
7020 | return ERR_PTR(-ENXIO); | |
7021 | if (f.file->f_op != &io_uring_fops) { | |
7022 | fdput(f); | |
7023 | return ERR_PTR(-EINVAL); | |
7024 | } | |
7025 | ||
7026 | ctx_attach = f.file->private_data; | |
7027 | sqd = ctx_attach->sq_data; | |
7028 | if (!sqd) { | |
7029 | fdput(f); | |
7030 | return ERR_PTR(-EINVAL); | |
7031 | } | |
7032 | ||
7033 | refcount_inc(&sqd->refs); | |
7034 | fdput(f); | |
7035 | return sqd; | |
7036 | } | |
7037 | ||
534ca6d6 JA |
7038 | static struct io_sq_data *io_get_sq_data(struct io_uring_params *p) |
7039 | { | |
7040 | struct io_sq_data *sqd; | |
7041 | ||
aa06165d JA |
7042 | if (p->flags & IORING_SETUP_ATTACH_WQ) |
7043 | return io_attach_sq_data(p); | |
7044 | ||
534ca6d6 JA |
7045 | sqd = kzalloc(sizeof(*sqd), GFP_KERNEL); |
7046 | if (!sqd) | |
7047 | return ERR_PTR(-ENOMEM); | |
7048 | ||
7049 | refcount_set(&sqd->refs, 1); | |
69fb2131 JA |
7050 | INIT_LIST_HEAD(&sqd->ctx_list); |
7051 | INIT_LIST_HEAD(&sqd->ctx_new_list); | |
7052 | mutex_init(&sqd->ctx_lock); | |
7053 | mutex_init(&sqd->lock); | |
534ca6d6 JA |
7054 | init_waitqueue_head(&sqd->wait); |
7055 | return sqd; | |
7056 | } | |
7057 | ||
69fb2131 JA |
7058 | static void io_sq_thread_unpark(struct io_sq_data *sqd) |
7059 | __releases(&sqd->lock) | |
7060 | { | |
7061 | if (!sqd->thread) | |
7062 | return; | |
7063 | kthread_unpark(sqd->thread); | |
7064 | mutex_unlock(&sqd->lock); | |
7065 | } | |
7066 | ||
7067 | static void io_sq_thread_park(struct io_sq_data *sqd) | |
7068 | __acquires(&sqd->lock) | |
7069 | { | |
7070 | if (!sqd->thread) | |
7071 | return; | |
7072 | mutex_lock(&sqd->lock); | |
7073 | kthread_park(sqd->thread); | |
7074 | } | |
7075 | ||
534ca6d6 JA |
7076 | static void io_sq_thread_stop(struct io_ring_ctx *ctx) |
7077 | { | |
7078 | struct io_sq_data *sqd = ctx->sq_data; | |
7079 | ||
7080 | if (sqd) { | |
7081 | if (sqd->thread) { | |
7082 | /* | |
7083 | * We may arrive here from the error branch in | |
7084 | * io_sq_offload_create() where the kthread is created | |
7085 | * without being waked up, thus wake it up now to make | |
7086 | * sure the wait will complete. | |
7087 | */ | |
7088 | wake_up_process(sqd->thread); | |
7089 | wait_for_completion(&ctx->sq_thread_comp); | |
69fb2131 JA |
7090 | |
7091 | io_sq_thread_park(sqd); | |
7092 | } | |
7093 | ||
7094 | mutex_lock(&sqd->ctx_lock); | |
7095 | list_del(&ctx->sqd_list); | |
7096 | mutex_unlock(&sqd->ctx_lock); | |
7097 | ||
7098 | if (sqd->thread) { | |
7099 | finish_wait(&sqd->wait, &ctx->sqo_wait_entry); | |
7100 | io_sq_thread_unpark(sqd); | |
534ca6d6 JA |
7101 | } |
7102 | ||
7103 | io_put_sq_data(sqd); | |
7104 | ctx->sq_data = NULL; | |
6c271ce2 JA |
7105 | } |
7106 | } | |
7107 | ||
6b06314c JA |
7108 | static void io_finish_async(struct io_ring_ctx *ctx) |
7109 | { | |
6c271ce2 JA |
7110 | io_sq_thread_stop(ctx); |
7111 | ||
561fb04a JA |
7112 | if (ctx->io_wq) { |
7113 | io_wq_destroy(ctx->io_wq); | |
7114 | ctx->io_wq = NULL; | |
6b06314c JA |
7115 | } |
7116 | } | |
7117 | ||
7118 | #if defined(CONFIG_UNIX) | |
6b06314c JA |
7119 | /* |
7120 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
7121 | * the io_uring can be safely unregistered on process exit, even if we have | |
7122 | * loops in the file referencing. | |
7123 | */ | |
7124 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
7125 | { | |
7126 | struct sock *sk = ctx->ring_sock->sk; | |
7127 | struct scm_fp_list *fpl; | |
7128 | struct sk_buff *skb; | |
08a45173 | 7129 | int i, nr_files; |
6b06314c | 7130 | |
6b06314c JA |
7131 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); |
7132 | if (!fpl) | |
7133 | return -ENOMEM; | |
7134 | ||
7135 | skb = alloc_skb(0, GFP_KERNEL); | |
7136 | if (!skb) { | |
7137 | kfree(fpl); | |
7138 | return -ENOMEM; | |
7139 | } | |
7140 | ||
7141 | skb->sk = sk; | |
6b06314c | 7142 | |
08a45173 | 7143 | nr_files = 0; |
6b06314c JA |
7144 | fpl->user = get_uid(ctx->user); |
7145 | for (i = 0; i < nr; i++) { | |
65e19f54 JA |
7146 | struct file *file = io_file_from_index(ctx, i + offset); |
7147 | ||
7148 | if (!file) | |
08a45173 | 7149 | continue; |
65e19f54 | 7150 | fpl->fp[nr_files] = get_file(file); |
08a45173 JA |
7151 | unix_inflight(fpl->user, fpl->fp[nr_files]); |
7152 | nr_files++; | |
6b06314c JA |
7153 | } |
7154 | ||
08a45173 JA |
7155 | if (nr_files) { |
7156 | fpl->max = SCM_MAX_FD; | |
7157 | fpl->count = nr_files; | |
7158 | UNIXCB(skb).fp = fpl; | |
05f3fb3c | 7159 | skb->destructor = unix_destruct_scm; |
08a45173 JA |
7160 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); |
7161 | skb_queue_head(&sk->sk_receive_queue, skb); | |
6b06314c | 7162 | |
08a45173 JA |
7163 | for (i = 0; i < nr_files; i++) |
7164 | fput(fpl->fp[i]); | |
7165 | } else { | |
7166 | kfree_skb(skb); | |
7167 | kfree(fpl); | |
7168 | } | |
6b06314c JA |
7169 | |
7170 | return 0; | |
7171 | } | |
7172 | ||
7173 | /* | |
7174 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
7175 | * causes regular reference counting to break down. We rely on the UNIX | |
7176 | * garbage collection to take care of this problem for us. | |
7177 | */ | |
7178 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
7179 | { | |
7180 | unsigned left, total; | |
7181 | int ret = 0; | |
7182 | ||
7183 | total = 0; | |
7184 | left = ctx->nr_user_files; | |
7185 | while (left) { | |
7186 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
7187 | |
7188 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
7189 | if (ret) | |
7190 | break; | |
7191 | left -= this_files; | |
7192 | total += this_files; | |
7193 | } | |
7194 | ||
7195 | if (!ret) | |
7196 | return 0; | |
7197 | ||
7198 | while (total < ctx->nr_user_files) { | |
65e19f54 JA |
7199 | struct file *file = io_file_from_index(ctx, total); |
7200 | ||
7201 | if (file) | |
7202 | fput(file); | |
6b06314c JA |
7203 | total++; |
7204 | } | |
7205 | ||
7206 | return ret; | |
7207 | } | |
7208 | #else | |
7209 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
7210 | { | |
7211 | return 0; | |
7212 | } | |
7213 | #endif | |
7214 | ||
65e19f54 JA |
7215 | static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables, |
7216 | unsigned nr_files) | |
7217 | { | |
7218 | int i; | |
7219 | ||
7220 | for (i = 0; i < nr_tables; i++) { | |
05f3fb3c | 7221 | struct fixed_file_table *table = &ctx->file_data->table[i]; |
65e19f54 JA |
7222 | unsigned this_files; |
7223 | ||
7224 | this_files = min(nr_files, IORING_MAX_FILES_TABLE); | |
7225 | table->files = kcalloc(this_files, sizeof(struct file *), | |
7226 | GFP_KERNEL); | |
7227 | if (!table->files) | |
7228 | break; | |
7229 | nr_files -= this_files; | |
7230 | } | |
7231 | ||
7232 | if (i == nr_tables) | |
7233 | return 0; | |
7234 | ||
7235 | for (i = 0; i < nr_tables; i++) { | |
05f3fb3c | 7236 | struct fixed_file_table *table = &ctx->file_data->table[i]; |
65e19f54 JA |
7237 | kfree(table->files); |
7238 | } | |
7239 | return 1; | |
7240 | } | |
7241 | ||
05f3fb3c JA |
7242 | static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file) |
7243 | { | |
7244 | #if defined(CONFIG_UNIX) | |
7245 | struct sock *sock = ctx->ring_sock->sk; | |
7246 | struct sk_buff_head list, *head = &sock->sk_receive_queue; | |
7247 | struct sk_buff *skb; | |
7248 | int i; | |
7249 | ||
7250 | __skb_queue_head_init(&list); | |
7251 | ||
7252 | /* | |
7253 | * Find the skb that holds this file in its SCM_RIGHTS. When found, | |
7254 | * remove this entry and rearrange the file array. | |
7255 | */ | |
7256 | skb = skb_dequeue(head); | |
7257 | while (skb) { | |
7258 | struct scm_fp_list *fp; | |
7259 | ||
7260 | fp = UNIXCB(skb).fp; | |
7261 | for (i = 0; i < fp->count; i++) { | |
7262 | int left; | |
7263 | ||
7264 | if (fp->fp[i] != file) | |
7265 | continue; | |
7266 | ||
7267 | unix_notinflight(fp->user, fp->fp[i]); | |
7268 | left = fp->count - 1 - i; | |
7269 | if (left) { | |
7270 | memmove(&fp->fp[i], &fp->fp[i + 1], | |
7271 | left * sizeof(struct file *)); | |
7272 | } | |
7273 | fp->count--; | |
7274 | if (!fp->count) { | |
7275 | kfree_skb(skb); | |
7276 | skb = NULL; | |
7277 | } else { | |
7278 | __skb_queue_tail(&list, skb); | |
7279 | } | |
7280 | fput(file); | |
7281 | file = NULL; | |
7282 | break; | |
7283 | } | |
7284 | ||
7285 | if (!file) | |
7286 | break; | |
7287 | ||
7288 | __skb_queue_tail(&list, skb); | |
7289 | ||
7290 | skb = skb_dequeue(head); | |
7291 | } | |
7292 | ||
7293 | if (skb_peek(&list)) { | |
7294 | spin_lock_irq(&head->lock); | |
7295 | while ((skb = __skb_dequeue(&list)) != NULL) | |
7296 | __skb_queue_tail(head, skb); | |
7297 | spin_unlock_irq(&head->lock); | |
7298 | } | |
7299 | #else | |
7300 | fput(file); | |
7301 | #endif | |
7302 | } | |
7303 | ||
7304 | struct io_file_put { | |
05589553 | 7305 | struct list_head list; |
05f3fb3c | 7306 | struct file *file; |
05f3fb3c JA |
7307 | }; |
7308 | ||
4a38aed2 | 7309 | static void __io_file_put_work(struct fixed_file_ref_node *ref_node) |
65e19f54 | 7310 | { |
4a38aed2 JA |
7311 | struct fixed_file_data *file_data = ref_node->file_data; |
7312 | struct io_ring_ctx *ctx = file_data->ctx; | |
05f3fb3c | 7313 | struct io_file_put *pfile, *tmp; |
05589553 XW |
7314 | |
7315 | list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) { | |
6a4d07cd | 7316 | list_del(&pfile->list); |
05589553 XW |
7317 | io_ring_file_put(ctx, pfile->file); |
7318 | kfree(pfile); | |
65e19f54 | 7319 | } |
05589553 | 7320 | |
6a4d07cd JA |
7321 | spin_lock(&file_data->lock); |
7322 | list_del(&ref_node->node); | |
7323 | spin_unlock(&file_data->lock); | |
05589553 XW |
7324 | |
7325 | percpu_ref_exit(&ref_node->refs); | |
7326 | kfree(ref_node); | |
7327 | percpu_ref_put(&file_data->refs); | |
2faf852d | 7328 | } |
65e19f54 | 7329 | |
4a38aed2 JA |
7330 | static void io_file_put_work(struct work_struct *work) |
7331 | { | |
7332 | struct io_ring_ctx *ctx; | |
7333 | struct llist_node *node; | |
7334 | ||
7335 | ctx = container_of(work, struct io_ring_ctx, file_put_work.work); | |
7336 | node = llist_del_all(&ctx->file_put_llist); | |
7337 | ||
7338 | while (node) { | |
7339 | struct fixed_file_ref_node *ref_node; | |
7340 | struct llist_node *next = node->next; | |
7341 | ||
7342 | ref_node = llist_entry(node, struct fixed_file_ref_node, llist); | |
7343 | __io_file_put_work(ref_node); | |
7344 | node = next; | |
7345 | } | |
7346 | } | |
7347 | ||
05589553 | 7348 | static void io_file_data_ref_zero(struct percpu_ref *ref) |
2faf852d | 7349 | { |
05589553 | 7350 | struct fixed_file_ref_node *ref_node; |
4a38aed2 JA |
7351 | struct io_ring_ctx *ctx; |
7352 | bool first_add; | |
7353 | int delay = HZ; | |
65e19f54 | 7354 | |
05589553 | 7355 | ref_node = container_of(ref, struct fixed_file_ref_node, refs); |
4a38aed2 | 7356 | ctx = ref_node->file_data->ctx; |
05589553 | 7357 | |
4a38aed2 JA |
7358 | if (percpu_ref_is_dying(&ctx->file_data->refs)) |
7359 | delay = 0; | |
05589553 | 7360 | |
4a38aed2 JA |
7361 | first_add = llist_add(&ref_node->llist, &ctx->file_put_llist); |
7362 | if (!delay) | |
7363 | mod_delayed_work(system_wq, &ctx->file_put_work, 0); | |
7364 | else if (first_add) | |
7365 | queue_delayed_work(system_wq, &ctx->file_put_work, delay); | |
05f3fb3c | 7366 | } |
65e19f54 | 7367 | |
05589553 XW |
7368 | static struct fixed_file_ref_node *alloc_fixed_file_ref_node( |
7369 | struct io_ring_ctx *ctx) | |
05f3fb3c | 7370 | { |
05589553 | 7371 | struct fixed_file_ref_node *ref_node; |
05f3fb3c | 7372 | |
05589553 XW |
7373 | ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL); |
7374 | if (!ref_node) | |
7375 | return ERR_PTR(-ENOMEM); | |
05f3fb3c | 7376 | |
05589553 XW |
7377 | if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero, |
7378 | 0, GFP_KERNEL)) { | |
7379 | kfree(ref_node); | |
7380 | return ERR_PTR(-ENOMEM); | |
7381 | } | |
7382 | INIT_LIST_HEAD(&ref_node->node); | |
7383 | INIT_LIST_HEAD(&ref_node->file_list); | |
05589553 XW |
7384 | ref_node->file_data = ctx->file_data; |
7385 | return ref_node; | |
05589553 XW |
7386 | } |
7387 | ||
7388 | static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node) | |
7389 | { | |
7390 | percpu_ref_exit(&ref_node->refs); | |
7391 | kfree(ref_node); | |
65e19f54 JA |
7392 | } |
7393 | ||
6b06314c JA |
7394 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, |
7395 | unsigned nr_args) | |
7396 | { | |
7397 | __s32 __user *fds = (__s32 __user *) arg; | |
65e19f54 | 7398 | unsigned nr_tables; |
05f3fb3c | 7399 | struct file *file; |
6b06314c JA |
7400 | int fd, ret = 0; |
7401 | unsigned i; | |
05589553 | 7402 | struct fixed_file_ref_node *ref_node; |
6b06314c | 7403 | |
05f3fb3c | 7404 | if (ctx->file_data) |
6b06314c JA |
7405 | return -EBUSY; |
7406 | if (!nr_args) | |
7407 | return -EINVAL; | |
7408 | if (nr_args > IORING_MAX_FIXED_FILES) | |
7409 | return -EMFILE; | |
7410 | ||
05f3fb3c JA |
7411 | ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL); |
7412 | if (!ctx->file_data) | |
7413 | return -ENOMEM; | |
7414 | ctx->file_data->ctx = ctx; | |
7415 | init_completion(&ctx->file_data->done); | |
05589553 | 7416 | INIT_LIST_HEAD(&ctx->file_data->ref_list); |
f7fe9346 | 7417 | spin_lock_init(&ctx->file_data->lock); |
05f3fb3c | 7418 | |
65e19f54 | 7419 | nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE); |
05f3fb3c JA |
7420 | ctx->file_data->table = kcalloc(nr_tables, |
7421 | sizeof(struct fixed_file_table), | |
65e19f54 | 7422 | GFP_KERNEL); |
05f3fb3c JA |
7423 | if (!ctx->file_data->table) { |
7424 | kfree(ctx->file_data); | |
7425 | ctx->file_data = NULL; | |
6b06314c | 7426 | return -ENOMEM; |
05f3fb3c JA |
7427 | } |
7428 | ||
05589553 | 7429 | if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill, |
05f3fb3c JA |
7430 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) { |
7431 | kfree(ctx->file_data->table); | |
7432 | kfree(ctx->file_data); | |
7433 | ctx->file_data = NULL; | |
6b06314c | 7434 | return -ENOMEM; |
05f3fb3c | 7435 | } |
6b06314c | 7436 | |
65e19f54 | 7437 | if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) { |
05f3fb3c JA |
7438 | percpu_ref_exit(&ctx->file_data->refs); |
7439 | kfree(ctx->file_data->table); | |
7440 | kfree(ctx->file_data); | |
7441 | ctx->file_data = NULL; | |
65e19f54 JA |
7442 | return -ENOMEM; |
7443 | } | |
7444 | ||
08a45173 | 7445 | for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { |
65e19f54 JA |
7446 | struct fixed_file_table *table; |
7447 | unsigned index; | |
7448 | ||
6b06314c JA |
7449 | ret = -EFAULT; |
7450 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) | |
7451 | break; | |
08a45173 JA |
7452 | /* allow sparse sets */ |
7453 | if (fd == -1) { | |
7454 | ret = 0; | |
7455 | continue; | |
7456 | } | |
6b06314c | 7457 | |
05f3fb3c | 7458 | table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT]; |
65e19f54 | 7459 | index = i & IORING_FILE_TABLE_MASK; |
05f3fb3c | 7460 | file = fget(fd); |
6b06314c JA |
7461 | |
7462 | ret = -EBADF; | |
05f3fb3c | 7463 | if (!file) |
6b06314c | 7464 | break; |
05f3fb3c | 7465 | |
6b06314c JA |
7466 | /* |
7467 | * Don't allow io_uring instances to be registered. If UNIX | |
7468 | * isn't enabled, then this causes a reference cycle and this | |
7469 | * instance can never get freed. If UNIX is enabled we'll | |
7470 | * handle it just fine, but there's still no point in allowing | |
7471 | * a ring fd as it doesn't support regular read/write anyway. | |
7472 | */ | |
05f3fb3c JA |
7473 | if (file->f_op == &io_uring_fops) { |
7474 | fput(file); | |
6b06314c JA |
7475 | break; |
7476 | } | |
6b06314c | 7477 | ret = 0; |
05f3fb3c | 7478 | table->files[index] = file; |
6b06314c JA |
7479 | } |
7480 | ||
7481 | if (ret) { | |
65e19f54 | 7482 | for (i = 0; i < ctx->nr_user_files; i++) { |
65e19f54 JA |
7483 | file = io_file_from_index(ctx, i); |
7484 | if (file) | |
7485 | fput(file); | |
7486 | } | |
7487 | for (i = 0; i < nr_tables; i++) | |
05f3fb3c | 7488 | kfree(ctx->file_data->table[i].files); |
6b06314c | 7489 | |
667e57da | 7490 | percpu_ref_exit(&ctx->file_data->refs); |
05f3fb3c JA |
7491 | kfree(ctx->file_data->table); |
7492 | kfree(ctx->file_data); | |
7493 | ctx->file_data = NULL; | |
6b06314c JA |
7494 | ctx->nr_user_files = 0; |
7495 | return ret; | |
7496 | } | |
7497 | ||
7498 | ret = io_sqe_files_scm(ctx); | |
05589553 | 7499 | if (ret) { |
6b06314c | 7500 | io_sqe_files_unregister(ctx); |
05589553 XW |
7501 | return ret; |
7502 | } | |
6b06314c | 7503 | |
05589553 XW |
7504 | ref_node = alloc_fixed_file_ref_node(ctx); |
7505 | if (IS_ERR(ref_node)) { | |
7506 | io_sqe_files_unregister(ctx); | |
7507 | return PTR_ERR(ref_node); | |
7508 | } | |
7509 | ||
7510 | ctx->file_data->cur_refs = &ref_node->refs; | |
6a4d07cd | 7511 | spin_lock(&ctx->file_data->lock); |
05589553 | 7512 | list_add(&ref_node->node, &ctx->file_data->ref_list); |
6a4d07cd | 7513 | spin_unlock(&ctx->file_data->lock); |
05589553 | 7514 | percpu_ref_get(&ctx->file_data->refs); |
6b06314c JA |
7515 | return ret; |
7516 | } | |
7517 | ||
c3a31e60 JA |
7518 | static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file, |
7519 | int index) | |
7520 | { | |
7521 | #if defined(CONFIG_UNIX) | |
7522 | struct sock *sock = ctx->ring_sock->sk; | |
7523 | struct sk_buff_head *head = &sock->sk_receive_queue; | |
7524 | struct sk_buff *skb; | |
7525 | ||
7526 | /* | |
7527 | * See if we can merge this file into an existing skb SCM_RIGHTS | |
7528 | * file set. If there's no room, fall back to allocating a new skb | |
7529 | * and filling it in. | |
7530 | */ | |
7531 | spin_lock_irq(&head->lock); | |
7532 | skb = skb_peek(head); | |
7533 | if (skb) { | |
7534 | struct scm_fp_list *fpl = UNIXCB(skb).fp; | |
7535 | ||
7536 | if (fpl->count < SCM_MAX_FD) { | |
7537 | __skb_unlink(skb, head); | |
7538 | spin_unlock_irq(&head->lock); | |
7539 | fpl->fp[fpl->count] = get_file(file); | |
7540 | unix_inflight(fpl->user, fpl->fp[fpl->count]); | |
7541 | fpl->count++; | |
7542 | spin_lock_irq(&head->lock); | |
7543 | __skb_queue_head(head, skb); | |
7544 | } else { | |
7545 | skb = NULL; | |
7546 | } | |
7547 | } | |
7548 | spin_unlock_irq(&head->lock); | |
7549 | ||
7550 | if (skb) { | |
7551 | fput(file); | |
7552 | return 0; | |
7553 | } | |
7554 | ||
7555 | return __io_sqe_files_scm(ctx, 1, index); | |
7556 | #else | |
7557 | return 0; | |
7558 | #endif | |
7559 | } | |
7560 | ||
a5318d3c | 7561 | static int io_queue_file_removal(struct fixed_file_data *data, |
05589553 | 7562 | struct file *file) |
05f3fb3c | 7563 | { |
a5318d3c | 7564 | struct io_file_put *pfile; |
05589553 XW |
7565 | struct percpu_ref *refs = data->cur_refs; |
7566 | struct fixed_file_ref_node *ref_node; | |
05f3fb3c | 7567 | |
05f3fb3c | 7568 | pfile = kzalloc(sizeof(*pfile), GFP_KERNEL); |
a5318d3c HD |
7569 | if (!pfile) |
7570 | return -ENOMEM; | |
05f3fb3c | 7571 | |
05589553 | 7572 | ref_node = container_of(refs, struct fixed_file_ref_node, refs); |
05f3fb3c | 7573 | pfile->file = file; |
05589553 XW |
7574 | list_add(&pfile->list, &ref_node->file_list); |
7575 | ||
a5318d3c | 7576 | return 0; |
05f3fb3c JA |
7577 | } |
7578 | ||
7579 | static int __io_sqe_files_update(struct io_ring_ctx *ctx, | |
7580 | struct io_uring_files_update *up, | |
7581 | unsigned nr_args) | |
7582 | { | |
7583 | struct fixed_file_data *data = ctx->file_data; | |
05589553 | 7584 | struct fixed_file_ref_node *ref_node; |
05f3fb3c | 7585 | struct file *file; |
c3a31e60 JA |
7586 | __s32 __user *fds; |
7587 | int fd, i, err; | |
7588 | __u32 done; | |
05589553 | 7589 | bool needs_switch = false; |
c3a31e60 | 7590 | |
05f3fb3c | 7591 | if (check_add_overflow(up->offset, nr_args, &done)) |
c3a31e60 JA |
7592 | return -EOVERFLOW; |
7593 | if (done > ctx->nr_user_files) | |
7594 | return -EINVAL; | |
7595 | ||
05589553 XW |
7596 | ref_node = alloc_fixed_file_ref_node(ctx); |
7597 | if (IS_ERR(ref_node)) | |
7598 | return PTR_ERR(ref_node); | |
7599 | ||
c3a31e60 | 7600 | done = 0; |
05f3fb3c | 7601 | fds = u64_to_user_ptr(up->fds); |
c3a31e60 | 7602 | while (nr_args) { |
65e19f54 JA |
7603 | struct fixed_file_table *table; |
7604 | unsigned index; | |
7605 | ||
c3a31e60 JA |
7606 | err = 0; |
7607 | if (copy_from_user(&fd, &fds[done], sizeof(fd))) { | |
7608 | err = -EFAULT; | |
7609 | break; | |
7610 | } | |
05f3fb3c JA |
7611 | i = array_index_nospec(up->offset, ctx->nr_user_files); |
7612 | table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT]; | |
65e19f54 JA |
7613 | index = i & IORING_FILE_TABLE_MASK; |
7614 | if (table->files[index]) { | |
98dfd502 | 7615 | file = table->files[index]; |
a5318d3c HD |
7616 | err = io_queue_file_removal(data, file); |
7617 | if (err) | |
7618 | break; | |
65e19f54 | 7619 | table->files[index] = NULL; |
05589553 | 7620 | needs_switch = true; |
c3a31e60 JA |
7621 | } |
7622 | if (fd != -1) { | |
c3a31e60 JA |
7623 | file = fget(fd); |
7624 | if (!file) { | |
7625 | err = -EBADF; | |
7626 | break; | |
7627 | } | |
7628 | /* | |
7629 | * Don't allow io_uring instances to be registered. If | |
7630 | * UNIX isn't enabled, then this causes a reference | |
7631 | * cycle and this instance can never get freed. If UNIX | |
7632 | * is enabled we'll handle it just fine, but there's | |
7633 | * still no point in allowing a ring fd as it doesn't | |
7634 | * support regular read/write anyway. | |
7635 | */ | |
7636 | if (file->f_op == &io_uring_fops) { | |
7637 | fput(file); | |
7638 | err = -EBADF; | |
7639 | break; | |
7640 | } | |
65e19f54 | 7641 | table->files[index] = file; |
c3a31e60 | 7642 | err = io_sqe_file_register(ctx, file, i); |
f3bd9dae | 7643 | if (err) { |
95d1c8e5 | 7644 | table->files[index] = NULL; |
f3bd9dae | 7645 | fput(file); |
c3a31e60 | 7646 | break; |
f3bd9dae | 7647 | } |
c3a31e60 JA |
7648 | } |
7649 | nr_args--; | |
7650 | done++; | |
05f3fb3c JA |
7651 | up->offset++; |
7652 | } | |
7653 | ||
05589553 XW |
7654 | if (needs_switch) { |
7655 | percpu_ref_kill(data->cur_refs); | |
6a4d07cd | 7656 | spin_lock(&data->lock); |
05589553 XW |
7657 | list_add(&ref_node->node, &data->ref_list); |
7658 | data->cur_refs = &ref_node->refs; | |
6a4d07cd | 7659 | spin_unlock(&data->lock); |
05589553 XW |
7660 | percpu_ref_get(&ctx->file_data->refs); |
7661 | } else | |
7662 | destroy_fixed_file_ref_node(ref_node); | |
c3a31e60 JA |
7663 | |
7664 | return done ? done : err; | |
7665 | } | |
05589553 | 7666 | |
05f3fb3c JA |
7667 | static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg, |
7668 | unsigned nr_args) | |
7669 | { | |
7670 | struct io_uring_files_update up; | |
7671 | ||
7672 | if (!ctx->file_data) | |
7673 | return -ENXIO; | |
7674 | if (!nr_args) | |
7675 | return -EINVAL; | |
7676 | if (copy_from_user(&up, arg, sizeof(up))) | |
7677 | return -EFAULT; | |
7678 | if (up.resv) | |
7679 | return -EINVAL; | |
7680 | ||
7681 | return __io_sqe_files_update(ctx, &up, nr_args); | |
7682 | } | |
c3a31e60 | 7683 | |
e9fd9396 | 7684 | static void io_free_work(struct io_wq_work *work) |
7d723065 JA |
7685 | { |
7686 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
7687 | ||
e9fd9396 | 7688 | /* Consider that io_steal_work() relies on this ref */ |
7d723065 JA |
7689 | io_put_req(req); |
7690 | } | |
7691 | ||
24369c2e PB |
7692 | static int io_init_wq_offload(struct io_ring_ctx *ctx, |
7693 | struct io_uring_params *p) | |
7694 | { | |
7695 | struct io_wq_data data; | |
7696 | struct fd f; | |
7697 | struct io_ring_ctx *ctx_attach; | |
7698 | unsigned int concurrency; | |
7699 | int ret = 0; | |
7700 | ||
7701 | data.user = ctx->user; | |
e9fd9396 | 7702 | data.free_work = io_free_work; |
f5fa38c5 | 7703 | data.do_work = io_wq_submit_work; |
24369c2e PB |
7704 | |
7705 | if (!(p->flags & IORING_SETUP_ATTACH_WQ)) { | |
7706 | /* Do QD, or 4 * CPUS, whatever is smallest */ | |
7707 | concurrency = min(ctx->sq_entries, 4 * num_online_cpus()); | |
7708 | ||
7709 | ctx->io_wq = io_wq_create(concurrency, &data); | |
7710 | if (IS_ERR(ctx->io_wq)) { | |
7711 | ret = PTR_ERR(ctx->io_wq); | |
7712 | ctx->io_wq = NULL; | |
7713 | } | |
7714 | return ret; | |
7715 | } | |
7716 | ||
7717 | f = fdget(p->wq_fd); | |
7718 | if (!f.file) | |
7719 | return -EBADF; | |
7720 | ||
7721 | if (f.file->f_op != &io_uring_fops) { | |
7722 | ret = -EINVAL; | |
7723 | goto out_fput; | |
7724 | } | |
7725 | ||
7726 | ctx_attach = f.file->private_data; | |
7727 | /* @io_wq is protected by holding the fd */ | |
7728 | if (!io_wq_get(ctx_attach->io_wq, &data)) { | |
7729 | ret = -EINVAL; | |
7730 | goto out_fput; | |
7731 | } | |
7732 | ||
7733 | ctx->io_wq = ctx_attach->io_wq; | |
7734 | out_fput: | |
7735 | fdput(f); | |
7736 | return ret; | |
7737 | } | |
7738 | ||
0f212204 JA |
7739 | static int io_uring_alloc_task_context(struct task_struct *task) |
7740 | { | |
7741 | struct io_uring_task *tctx; | |
7742 | ||
7743 | tctx = kmalloc(sizeof(*tctx), GFP_KERNEL); | |
7744 | if (unlikely(!tctx)) | |
7745 | return -ENOMEM; | |
7746 | ||
7747 | xa_init(&tctx->xa); | |
7748 | init_waitqueue_head(&tctx->wait); | |
7749 | tctx->last = NULL; | |
7750 | tctx->in_idle = 0; | |
7751 | atomic_long_set(&tctx->req_issue, 0); | |
7752 | atomic_long_set(&tctx->req_complete, 0); | |
7753 | task->io_uring = tctx; | |
7754 | return 0; | |
7755 | } | |
7756 | ||
7757 | void __io_uring_free(struct task_struct *tsk) | |
7758 | { | |
7759 | struct io_uring_task *tctx = tsk->io_uring; | |
7760 | ||
7761 | WARN_ON_ONCE(!xa_empty(&tctx->xa)); | |
7762 | xa_destroy(&tctx->xa); | |
7763 | kfree(tctx); | |
7764 | tsk->io_uring = NULL; | |
7765 | } | |
7766 | ||
7e84e1c7 SG |
7767 | static int io_sq_offload_create(struct io_ring_ctx *ctx, |
7768 | struct io_uring_params *p) | |
2b188cc1 JA |
7769 | { |
7770 | int ret; | |
7771 | ||
6c271ce2 | 7772 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
534ca6d6 JA |
7773 | struct io_sq_data *sqd; |
7774 | ||
3ec482d1 JA |
7775 | ret = -EPERM; |
7776 | if (!capable(CAP_SYS_ADMIN)) | |
7777 | goto err; | |
7778 | ||
534ca6d6 JA |
7779 | sqd = io_get_sq_data(p); |
7780 | if (IS_ERR(sqd)) { | |
7781 | ret = PTR_ERR(sqd); | |
7782 | goto err; | |
7783 | } | |
69fb2131 | 7784 | |
534ca6d6 | 7785 | ctx->sq_data = sqd; |
69fb2131 JA |
7786 | io_sq_thread_park(sqd); |
7787 | mutex_lock(&sqd->ctx_lock); | |
7788 | list_add(&ctx->sqd_list, &sqd->ctx_new_list); | |
7789 | mutex_unlock(&sqd->ctx_lock); | |
7790 | io_sq_thread_unpark(sqd); | |
534ca6d6 | 7791 | |
917257da JA |
7792 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
7793 | if (!ctx->sq_thread_idle) | |
7794 | ctx->sq_thread_idle = HZ; | |
7795 | ||
aa06165d JA |
7796 | if (sqd->thread) |
7797 | goto done; | |
7798 | ||
6c271ce2 | 7799 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 7800 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 7801 | |
917257da | 7802 | ret = -EINVAL; |
44a9bd18 JA |
7803 | if (cpu >= nr_cpu_ids) |
7804 | goto err; | |
7889f44d | 7805 | if (!cpu_online(cpu)) |
917257da JA |
7806 | goto err; |
7807 | ||
69fb2131 | 7808 | sqd->thread = kthread_create_on_cpu(io_sq_thread, sqd, |
534ca6d6 | 7809 | cpu, "io_uring-sq"); |
6c271ce2 | 7810 | } else { |
69fb2131 | 7811 | sqd->thread = kthread_create(io_sq_thread, sqd, |
6c271ce2 JA |
7812 | "io_uring-sq"); |
7813 | } | |
534ca6d6 JA |
7814 | if (IS_ERR(sqd->thread)) { |
7815 | ret = PTR_ERR(sqd->thread); | |
7816 | sqd->thread = NULL; | |
6c271ce2 JA |
7817 | goto err; |
7818 | } | |
534ca6d6 | 7819 | ret = io_uring_alloc_task_context(sqd->thread); |
0f212204 JA |
7820 | if (ret) |
7821 | goto err; | |
6c271ce2 JA |
7822 | } else if (p->flags & IORING_SETUP_SQ_AFF) { |
7823 | /* Can't have SQ_AFF without SQPOLL */ | |
7824 | ret = -EINVAL; | |
7825 | goto err; | |
7826 | } | |
7827 | ||
aa06165d | 7828 | done: |
24369c2e PB |
7829 | ret = io_init_wq_offload(ctx, p); |
7830 | if (ret) | |
2b188cc1 | 7831 | goto err; |
2b188cc1 JA |
7832 | |
7833 | return 0; | |
7834 | err: | |
54a91f3b | 7835 | io_finish_async(ctx); |
2b188cc1 JA |
7836 | return ret; |
7837 | } | |
7838 | ||
7e84e1c7 SG |
7839 | static void io_sq_offload_start(struct io_ring_ctx *ctx) |
7840 | { | |
534ca6d6 JA |
7841 | struct io_sq_data *sqd = ctx->sq_data; |
7842 | ||
7843 | if ((ctx->flags & IORING_SETUP_SQPOLL) && sqd->thread) | |
7844 | wake_up_process(sqd->thread); | |
7e84e1c7 SG |
7845 | } |
7846 | ||
a087e2b5 BM |
7847 | static inline void __io_unaccount_mem(struct user_struct *user, |
7848 | unsigned long nr_pages) | |
2b188cc1 JA |
7849 | { |
7850 | atomic_long_sub(nr_pages, &user->locked_vm); | |
7851 | } | |
7852 | ||
a087e2b5 BM |
7853 | static inline int __io_account_mem(struct user_struct *user, |
7854 | unsigned long nr_pages) | |
2b188cc1 JA |
7855 | { |
7856 | unsigned long page_limit, cur_pages, new_pages; | |
7857 | ||
7858 | /* Don't allow more pages than we can safely lock */ | |
7859 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
7860 | ||
7861 | do { | |
7862 | cur_pages = atomic_long_read(&user->locked_vm); | |
7863 | new_pages = cur_pages + nr_pages; | |
7864 | if (new_pages > page_limit) | |
7865 | return -ENOMEM; | |
7866 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
7867 | new_pages) != cur_pages); | |
7868 | ||
7869 | return 0; | |
7870 | } | |
7871 | ||
2e0464d4 BM |
7872 | static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages, |
7873 | enum io_mem_account acct) | |
a087e2b5 | 7874 | { |
aad5d8da | 7875 | if (ctx->limit_mem) |
a087e2b5 | 7876 | __io_unaccount_mem(ctx->user, nr_pages); |
30975825 | 7877 | |
2aede0e4 | 7878 | if (ctx->mm_account) { |
2e0464d4 | 7879 | if (acct == ACCT_LOCKED) |
2aede0e4 | 7880 | ctx->mm_account->locked_vm -= nr_pages; |
2e0464d4 | 7881 | else if (acct == ACCT_PINNED) |
2aede0e4 | 7882 | atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm); |
2e0464d4 | 7883 | } |
a087e2b5 BM |
7884 | } |
7885 | ||
2e0464d4 BM |
7886 | static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages, |
7887 | enum io_mem_account acct) | |
a087e2b5 | 7888 | { |
30975825 BM |
7889 | int ret; |
7890 | ||
7891 | if (ctx->limit_mem) { | |
7892 | ret = __io_account_mem(ctx->user, nr_pages); | |
7893 | if (ret) | |
7894 | return ret; | |
7895 | } | |
7896 | ||
2aede0e4 | 7897 | if (ctx->mm_account) { |
2e0464d4 | 7898 | if (acct == ACCT_LOCKED) |
2aede0e4 | 7899 | ctx->mm_account->locked_vm += nr_pages; |
2e0464d4 | 7900 | else if (acct == ACCT_PINNED) |
2aede0e4 | 7901 | atomic64_add(nr_pages, &ctx->mm_account->pinned_vm); |
2e0464d4 | 7902 | } |
a087e2b5 BM |
7903 | |
7904 | return 0; | |
7905 | } | |
7906 | ||
2b188cc1 JA |
7907 | static void io_mem_free(void *ptr) |
7908 | { | |
52e04ef4 MR |
7909 | struct page *page; |
7910 | ||
7911 | if (!ptr) | |
7912 | return; | |
2b188cc1 | 7913 | |
52e04ef4 | 7914 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
7915 | if (put_page_testzero(page)) |
7916 | free_compound_page(page); | |
7917 | } | |
7918 | ||
7919 | static void *io_mem_alloc(size_t size) | |
7920 | { | |
7921 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
7922 | __GFP_NORETRY; | |
7923 | ||
7924 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
7925 | } | |
7926 | ||
75b28aff HV |
7927 | static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries, |
7928 | size_t *sq_offset) | |
7929 | { | |
7930 | struct io_rings *rings; | |
7931 | size_t off, sq_array_size; | |
7932 | ||
7933 | off = struct_size(rings, cqes, cq_entries); | |
7934 | if (off == SIZE_MAX) | |
7935 | return SIZE_MAX; | |
7936 | ||
7937 | #ifdef CONFIG_SMP | |
7938 | off = ALIGN(off, SMP_CACHE_BYTES); | |
7939 | if (off == 0) | |
7940 | return SIZE_MAX; | |
7941 | #endif | |
7942 | ||
b36200f5 DV |
7943 | if (sq_offset) |
7944 | *sq_offset = off; | |
7945 | ||
75b28aff HV |
7946 | sq_array_size = array_size(sizeof(u32), sq_entries); |
7947 | if (sq_array_size == SIZE_MAX) | |
7948 | return SIZE_MAX; | |
7949 | ||
7950 | if (check_add_overflow(off, sq_array_size, &off)) | |
7951 | return SIZE_MAX; | |
7952 | ||
75b28aff HV |
7953 | return off; |
7954 | } | |
7955 | ||
2b188cc1 JA |
7956 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) |
7957 | { | |
75b28aff | 7958 | size_t pages; |
2b188cc1 | 7959 | |
75b28aff HV |
7960 | pages = (size_t)1 << get_order( |
7961 | rings_size(sq_entries, cq_entries, NULL)); | |
7962 | pages += (size_t)1 << get_order( | |
7963 | array_size(sizeof(struct io_uring_sqe), sq_entries)); | |
2b188cc1 | 7964 | |
75b28aff | 7965 | return pages; |
2b188cc1 JA |
7966 | } |
7967 | ||
edafccee JA |
7968 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
7969 | { | |
7970 | int i, j; | |
7971 | ||
7972 | if (!ctx->user_bufs) | |
7973 | return -ENXIO; | |
7974 | ||
7975 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
7976 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
7977 | ||
7978 | for (j = 0; j < imu->nr_bvecs; j++) | |
f1f6a7dd | 7979 | unpin_user_page(imu->bvec[j].bv_page); |
edafccee | 7980 | |
2e0464d4 | 7981 | io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED); |
d4ef6475 | 7982 | kvfree(imu->bvec); |
edafccee JA |
7983 | imu->nr_bvecs = 0; |
7984 | } | |
7985 | ||
7986 | kfree(ctx->user_bufs); | |
7987 | ctx->user_bufs = NULL; | |
7988 | ctx->nr_user_bufs = 0; | |
7989 | return 0; | |
7990 | } | |
7991 | ||
7992 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
7993 | void __user *arg, unsigned index) | |
7994 | { | |
7995 | struct iovec __user *src; | |
7996 | ||
7997 | #ifdef CONFIG_COMPAT | |
7998 | if (ctx->compat) { | |
7999 | struct compat_iovec __user *ciovs; | |
8000 | struct compat_iovec ciov; | |
8001 | ||
8002 | ciovs = (struct compat_iovec __user *) arg; | |
8003 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
8004 | return -EFAULT; | |
8005 | ||
d55e5f5b | 8006 | dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base); |
edafccee JA |
8007 | dst->iov_len = ciov.iov_len; |
8008 | return 0; | |
8009 | } | |
8010 | #endif | |
8011 | src = (struct iovec __user *) arg; | |
8012 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
8013 | return -EFAULT; | |
8014 | return 0; | |
8015 | } | |
8016 | ||
8017 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, | |
8018 | unsigned nr_args) | |
8019 | { | |
8020 | struct vm_area_struct **vmas = NULL; | |
8021 | struct page **pages = NULL; | |
8022 | int i, j, got_pages = 0; | |
8023 | int ret = -EINVAL; | |
8024 | ||
8025 | if (ctx->user_bufs) | |
8026 | return -EBUSY; | |
8027 | if (!nr_args || nr_args > UIO_MAXIOV) | |
8028 | return -EINVAL; | |
8029 | ||
8030 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
8031 | GFP_KERNEL); | |
8032 | if (!ctx->user_bufs) | |
8033 | return -ENOMEM; | |
8034 | ||
8035 | for (i = 0; i < nr_args; i++) { | |
8036 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
8037 | unsigned long off, start, end, ubuf; | |
8038 | int pret, nr_pages; | |
8039 | struct iovec iov; | |
8040 | size_t size; | |
8041 | ||
8042 | ret = io_copy_iov(ctx, &iov, arg, i); | |
8043 | if (ret) | |
a278682d | 8044 | goto err; |
edafccee JA |
8045 | |
8046 | /* | |
8047 | * Don't impose further limits on the size and buffer | |
8048 | * constraints here, we'll -EINVAL later when IO is | |
8049 | * submitted if they are wrong. | |
8050 | */ | |
8051 | ret = -EFAULT; | |
8052 | if (!iov.iov_base || !iov.iov_len) | |
8053 | goto err; | |
8054 | ||
8055 | /* arbitrary limit, but we need something */ | |
8056 | if (iov.iov_len > SZ_1G) | |
8057 | goto err; | |
8058 | ||
8059 | ubuf = (unsigned long) iov.iov_base; | |
8060 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
8061 | start = ubuf >> PAGE_SHIFT; | |
8062 | nr_pages = end - start; | |
8063 | ||
2e0464d4 | 8064 | ret = io_account_mem(ctx, nr_pages, ACCT_PINNED); |
a087e2b5 BM |
8065 | if (ret) |
8066 | goto err; | |
edafccee JA |
8067 | |
8068 | ret = 0; | |
8069 | if (!pages || nr_pages > got_pages) { | |
a8c73c1a DE |
8070 | kvfree(vmas); |
8071 | kvfree(pages); | |
d4ef6475 | 8072 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), |
edafccee | 8073 | GFP_KERNEL); |
d4ef6475 | 8074 | vmas = kvmalloc_array(nr_pages, |
edafccee JA |
8075 | sizeof(struct vm_area_struct *), |
8076 | GFP_KERNEL); | |
8077 | if (!pages || !vmas) { | |
8078 | ret = -ENOMEM; | |
2e0464d4 | 8079 | io_unaccount_mem(ctx, nr_pages, ACCT_PINNED); |
edafccee JA |
8080 | goto err; |
8081 | } | |
8082 | got_pages = nr_pages; | |
8083 | } | |
8084 | ||
d4ef6475 | 8085 | imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec), |
edafccee JA |
8086 | GFP_KERNEL); |
8087 | ret = -ENOMEM; | |
8088 | if (!imu->bvec) { | |
2e0464d4 | 8089 | io_unaccount_mem(ctx, nr_pages, ACCT_PINNED); |
edafccee JA |
8090 | goto err; |
8091 | } | |
8092 | ||
8093 | ret = 0; | |
d8ed45c5 | 8094 | mmap_read_lock(current->mm); |
2113b05d | 8095 | pret = pin_user_pages(ubuf, nr_pages, |
932f4a63 IW |
8096 | FOLL_WRITE | FOLL_LONGTERM, |
8097 | pages, vmas); | |
edafccee JA |
8098 | if (pret == nr_pages) { |
8099 | /* don't support file backed memory */ | |
8100 | for (j = 0; j < nr_pages; j++) { | |
8101 | struct vm_area_struct *vma = vmas[j]; | |
8102 | ||
8103 | if (vma->vm_file && | |
8104 | !is_file_hugepages(vma->vm_file)) { | |
8105 | ret = -EOPNOTSUPP; | |
8106 | break; | |
8107 | } | |
8108 | } | |
8109 | } else { | |
8110 | ret = pret < 0 ? pret : -EFAULT; | |
8111 | } | |
d8ed45c5 | 8112 | mmap_read_unlock(current->mm); |
edafccee JA |
8113 | if (ret) { |
8114 | /* | |
8115 | * if we did partial map, or found file backed vmas, | |
8116 | * release any pages we did get | |
8117 | */ | |
27c4d3a3 | 8118 | if (pret > 0) |
f1f6a7dd | 8119 | unpin_user_pages(pages, pret); |
2e0464d4 | 8120 | io_unaccount_mem(ctx, nr_pages, ACCT_PINNED); |
d4ef6475 | 8121 | kvfree(imu->bvec); |
edafccee JA |
8122 | goto err; |
8123 | } | |
8124 | ||
8125 | off = ubuf & ~PAGE_MASK; | |
8126 | size = iov.iov_len; | |
8127 | for (j = 0; j < nr_pages; j++) { | |
8128 | size_t vec_len; | |
8129 | ||
8130 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
8131 | imu->bvec[j].bv_page = pages[j]; | |
8132 | imu->bvec[j].bv_len = vec_len; | |
8133 | imu->bvec[j].bv_offset = off; | |
8134 | off = 0; | |
8135 | size -= vec_len; | |
8136 | } | |
8137 | /* store original address for later verification */ | |
8138 | imu->ubuf = ubuf; | |
8139 | imu->len = iov.iov_len; | |
8140 | imu->nr_bvecs = nr_pages; | |
8141 | ||
8142 | ctx->nr_user_bufs++; | |
8143 | } | |
d4ef6475 MR |
8144 | kvfree(pages); |
8145 | kvfree(vmas); | |
edafccee JA |
8146 | return 0; |
8147 | err: | |
d4ef6475 MR |
8148 | kvfree(pages); |
8149 | kvfree(vmas); | |
edafccee JA |
8150 | io_sqe_buffer_unregister(ctx); |
8151 | return ret; | |
8152 | } | |
8153 | ||
9b402849 JA |
8154 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
8155 | { | |
8156 | __s32 __user *fds = arg; | |
8157 | int fd; | |
8158 | ||
8159 | if (ctx->cq_ev_fd) | |
8160 | return -EBUSY; | |
8161 | ||
8162 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
8163 | return -EFAULT; | |
8164 | ||
8165 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
8166 | if (IS_ERR(ctx->cq_ev_fd)) { | |
8167 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
8168 | ctx->cq_ev_fd = NULL; | |
8169 | return ret; | |
8170 | } | |
8171 | ||
8172 | return 0; | |
8173 | } | |
8174 | ||
8175 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
8176 | { | |
8177 | if (ctx->cq_ev_fd) { | |
8178 | eventfd_ctx_put(ctx->cq_ev_fd); | |
8179 | ctx->cq_ev_fd = NULL; | |
8180 | return 0; | |
8181 | } | |
8182 | ||
8183 | return -ENXIO; | |
8184 | } | |
8185 | ||
5a2e745d JA |
8186 | static int __io_destroy_buffers(int id, void *p, void *data) |
8187 | { | |
8188 | struct io_ring_ctx *ctx = data; | |
8189 | struct io_buffer *buf = p; | |
8190 | ||
067524e9 | 8191 | __io_remove_buffers(ctx, buf, id, -1U); |
5a2e745d JA |
8192 | return 0; |
8193 | } | |
8194 | ||
8195 | static void io_destroy_buffers(struct io_ring_ctx *ctx) | |
8196 | { | |
8197 | idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx); | |
8198 | idr_destroy(&ctx->io_buffer_idr); | |
8199 | } | |
8200 | ||
2b188cc1 JA |
8201 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
8202 | { | |
6b06314c | 8203 | io_finish_async(ctx); |
5dbcad51 | 8204 | io_sqe_buffer_unregister(ctx); |
2aede0e4 JA |
8205 | |
8206 | if (ctx->sqo_task) { | |
8207 | put_task_struct(ctx->sqo_task); | |
8208 | ctx->sqo_task = NULL; | |
8209 | mmdrop(ctx->mm_account); | |
8210 | ctx->mm_account = NULL; | |
30975825 | 8211 | } |
def596e9 | 8212 | |
6b06314c | 8213 | io_sqe_files_unregister(ctx); |
9b402849 | 8214 | io_eventfd_unregister(ctx); |
5a2e745d | 8215 | io_destroy_buffers(ctx); |
41726c9a | 8216 | idr_destroy(&ctx->personality_idr); |
def596e9 | 8217 | |
2b188cc1 | 8218 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
8219 | if (ctx->ring_sock) { |
8220 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 8221 | sock_release(ctx->ring_sock); |
355e8d26 | 8222 | } |
2b188cc1 JA |
8223 | #endif |
8224 | ||
75b28aff | 8225 | io_mem_free(ctx->rings); |
2b188cc1 | 8226 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
8227 | |
8228 | percpu_ref_exit(&ctx->refs); | |
2b188cc1 | 8229 | free_uid(ctx->user); |
181e448d | 8230 | put_cred(ctx->creds); |
78076bb6 | 8231 | kfree(ctx->cancel_hash); |
0ddf92e8 | 8232 | kmem_cache_free(req_cachep, ctx->fallback_req); |
2b188cc1 JA |
8233 | kfree(ctx); |
8234 | } | |
8235 | ||
8236 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
8237 | { | |
8238 | struct io_ring_ctx *ctx = file->private_data; | |
8239 | __poll_t mask = 0; | |
8240 | ||
8241 | poll_wait(file, &ctx->cq_wait, wait); | |
4f7067c3 SB |
8242 | /* |
8243 | * synchronizes with barrier from wq_has_sleeper call in | |
8244 | * io_commit_cqring | |
8245 | */ | |
2b188cc1 | 8246 | smp_rmb(); |
90554200 | 8247 | if (!io_sqring_full(ctx)) |
2b188cc1 | 8248 | mask |= EPOLLOUT | EPOLLWRNORM; |
63e5d81f | 8249 | if (io_cqring_events(ctx, false)) |
2b188cc1 JA |
8250 | mask |= EPOLLIN | EPOLLRDNORM; |
8251 | ||
8252 | return mask; | |
8253 | } | |
8254 | ||
8255 | static int io_uring_fasync(int fd, struct file *file, int on) | |
8256 | { | |
8257 | struct io_ring_ctx *ctx = file->private_data; | |
8258 | ||
8259 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
8260 | } | |
8261 | ||
071698e1 JA |
8262 | static int io_remove_personalities(int id, void *p, void *data) |
8263 | { | |
8264 | struct io_ring_ctx *ctx = data; | |
8265 | const struct cred *cred; | |
8266 | ||
8267 | cred = idr_remove(&ctx->personality_idr, id); | |
8268 | if (cred) | |
8269 | put_cred(cred); | |
8270 | return 0; | |
8271 | } | |
8272 | ||
85faa7b8 JA |
8273 | static void io_ring_exit_work(struct work_struct *work) |
8274 | { | |
b2edc0a7 PB |
8275 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, |
8276 | exit_work); | |
85faa7b8 | 8277 | |
56952e91 JA |
8278 | /* |
8279 | * If we're doing polled IO and end up having requests being | |
8280 | * submitted async (out-of-line), then completions can come in while | |
8281 | * we're waiting for refs to drop. We need to reap these manually, | |
8282 | * as nobody else will be looking for them. | |
8283 | */ | |
b2edc0a7 | 8284 | do { |
56952e91 | 8285 | if (ctx->rings) |
e6c8aa9a | 8286 | io_cqring_overflow_flush(ctx, true, NULL, NULL); |
b2edc0a7 PB |
8287 | io_iopoll_try_reap_events(ctx); |
8288 | } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20)); | |
85faa7b8 JA |
8289 | io_ring_ctx_free(ctx); |
8290 | } | |
8291 | ||
2b188cc1 JA |
8292 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
8293 | { | |
8294 | mutex_lock(&ctx->uring_lock); | |
8295 | percpu_ref_kill(&ctx->refs); | |
8296 | mutex_unlock(&ctx->uring_lock); | |
8297 | ||
f3606e3a JA |
8298 | io_kill_timeouts(ctx, NULL); |
8299 | io_poll_remove_all(ctx, NULL); | |
561fb04a JA |
8300 | |
8301 | if (ctx->io_wq) | |
8302 | io_wq_cancel_all(ctx->io_wq); | |
8303 | ||
15dff286 JA |
8304 | /* if we failed setting up the ctx, we might not have any rings */ |
8305 | if (ctx->rings) | |
e6c8aa9a | 8306 | io_cqring_overflow_flush(ctx, true, NULL, NULL); |
b2edc0a7 | 8307 | io_iopoll_try_reap_events(ctx); |
071698e1 | 8308 | idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx); |
309fc03a JA |
8309 | |
8310 | /* | |
8311 | * Do this upfront, so we won't have a grace period where the ring | |
8312 | * is closed but resources aren't reaped yet. This can cause | |
8313 | * spurious failure in setting up a new ring. | |
8314 | */ | |
760618f7 JA |
8315 | io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries), |
8316 | ACCT_LOCKED); | |
309fc03a | 8317 | |
85faa7b8 | 8318 | INIT_WORK(&ctx->exit_work, io_ring_exit_work); |
fc666777 JA |
8319 | /* |
8320 | * Use system_unbound_wq to avoid spawning tons of event kworkers | |
8321 | * if we're exiting a ton of rings at the same time. It just adds | |
8322 | * noise and overhead, there's no discernable change in runtime | |
8323 | * over using system_wq. | |
8324 | */ | |
8325 | queue_work(system_unbound_wq, &ctx->exit_work); | |
2b188cc1 JA |
8326 | } |
8327 | ||
8328 | static int io_uring_release(struct inode *inode, struct file *file) | |
8329 | { | |
8330 | struct io_ring_ctx *ctx = file->private_data; | |
8331 | ||
8332 | file->private_data = NULL; | |
8333 | io_ring_ctx_wait_and_kill(ctx); | |
8334 | return 0; | |
8335 | } | |
8336 | ||
67c4d9e6 PB |
8337 | static bool io_wq_files_match(struct io_wq_work *work, void *data) |
8338 | { | |
8339 | struct files_struct *files = data; | |
8340 | ||
0f212204 | 8341 | return !files || work->files == files; |
67c4d9e6 PB |
8342 | } |
8343 | ||
f254ac04 JA |
8344 | /* |
8345 | * Returns true if 'preq' is the link parent of 'req' | |
8346 | */ | |
8347 | static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req) | |
8348 | { | |
8349 | struct io_kiocb *link; | |
8350 | ||
8351 | if (!(preq->flags & REQ_F_LINK_HEAD)) | |
8352 | return false; | |
8353 | ||
8354 | list_for_each_entry(link, &preq->link_list, link_list) { | |
8355 | if (link == req) | |
8356 | return true; | |
8357 | } | |
8358 | ||
8359 | return false; | |
8360 | } | |
8361 | ||
c127a2a1 PB |
8362 | static bool io_match_link_files(struct io_kiocb *req, |
8363 | struct files_struct *files) | |
8364 | { | |
8365 | struct io_kiocb *link; | |
8366 | ||
8367 | if (io_match_files(req, files)) | |
8368 | return true; | |
8369 | if (req->flags & REQ_F_LINK_HEAD) { | |
8370 | list_for_each_entry(link, &req->link_list, link_list) { | |
8371 | if (io_match_files(link, files)) | |
8372 | return true; | |
8373 | } | |
8374 | } | |
8375 | return false; | |
8376 | } | |
8377 | ||
f254ac04 JA |
8378 | /* |
8379 | * We're looking to cancel 'req' because it's holding on to our files, but | |
8380 | * 'req' could be a link to another request. See if it is, and cancel that | |
8381 | * parent request if so. | |
8382 | */ | |
8383 | static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req) | |
8384 | { | |
8385 | struct hlist_node *tmp; | |
8386 | struct io_kiocb *preq; | |
8387 | bool found = false; | |
8388 | int i; | |
8389 | ||
8390 | spin_lock_irq(&ctx->completion_lock); | |
8391 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { | |
8392 | struct hlist_head *list; | |
8393 | ||
8394 | list = &ctx->cancel_hash[i]; | |
8395 | hlist_for_each_entry_safe(preq, tmp, list, hash_node) { | |
8396 | found = io_match_link(preq, req); | |
8397 | if (found) { | |
8398 | io_poll_remove_one(preq); | |
8399 | break; | |
8400 | } | |
8401 | } | |
8402 | } | |
8403 | spin_unlock_irq(&ctx->completion_lock); | |
8404 | return found; | |
8405 | } | |
8406 | ||
8407 | static bool io_timeout_remove_link(struct io_ring_ctx *ctx, | |
8408 | struct io_kiocb *req) | |
8409 | { | |
8410 | struct io_kiocb *preq; | |
8411 | bool found = false; | |
8412 | ||
8413 | spin_lock_irq(&ctx->completion_lock); | |
8414 | list_for_each_entry(preq, &ctx->timeout_list, timeout.list) { | |
8415 | found = io_match_link(preq, req); | |
8416 | if (found) { | |
8417 | __io_timeout_cancel(preq); | |
8418 | break; | |
8419 | } | |
8420 | } | |
8421 | spin_unlock_irq(&ctx->completion_lock); | |
8422 | return found; | |
8423 | } | |
8424 | ||
b711d4ea JA |
8425 | static bool io_cancel_link_cb(struct io_wq_work *work, void *data) |
8426 | { | |
8427 | return io_match_link(container_of(work, struct io_kiocb, work), data); | |
8428 | } | |
8429 | ||
8430 | static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req) | |
8431 | { | |
8432 | enum io_wq_cancel cret; | |
8433 | ||
8434 | /* cancel this particular work, if it's running */ | |
8435 | cret = io_wq_cancel_work(ctx->io_wq, &req->work); | |
8436 | if (cret != IO_WQ_CANCEL_NOTFOUND) | |
8437 | return; | |
8438 | ||
8439 | /* find links that hold this pending, cancel those */ | |
8440 | cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true); | |
8441 | if (cret != IO_WQ_CANCEL_NOTFOUND) | |
8442 | return; | |
8443 | ||
8444 | /* if we have a poll link holding this pending, cancel that */ | |
8445 | if (io_poll_remove_link(ctx, req)) | |
8446 | return; | |
8447 | ||
8448 | /* final option, timeout link is holding this req pending */ | |
8449 | io_timeout_remove_link(ctx, req); | |
8450 | } | |
8451 | ||
b7ddce3c PB |
8452 | static void io_cancel_defer_files(struct io_ring_ctx *ctx, |
8453 | struct files_struct *files) | |
8454 | { | |
8455 | struct io_defer_entry *de = NULL; | |
8456 | LIST_HEAD(list); | |
8457 | ||
8458 | spin_lock_irq(&ctx->completion_lock); | |
8459 | list_for_each_entry_reverse(de, &ctx->defer_list, list) { | |
c127a2a1 | 8460 | if (io_match_link_files(de->req, files)) { |
b7ddce3c PB |
8461 | list_cut_position(&list, &ctx->defer_list, &de->list); |
8462 | break; | |
8463 | } | |
8464 | } | |
8465 | spin_unlock_irq(&ctx->completion_lock); | |
8466 | ||
8467 | while (!list_empty(&list)) { | |
8468 | de = list_first_entry(&list, struct io_defer_entry, list); | |
8469 | list_del_init(&de->list); | |
8470 | req_set_fail_links(de->req); | |
8471 | io_put_req(de->req); | |
8472 | io_req_complete(de->req, -ECANCELED); | |
8473 | kfree(de); | |
8474 | } | |
8475 | } | |
8476 | ||
76e1b642 JA |
8477 | /* |
8478 | * Returns true if we found and killed one or more files pinning requests | |
8479 | */ | |
8480 | static bool io_uring_cancel_files(struct io_ring_ctx *ctx, | |
fcb323cc JA |
8481 | struct files_struct *files) |
8482 | { | |
67c4d9e6 | 8483 | if (list_empty_careful(&ctx->inflight_list)) |
76e1b642 | 8484 | return false; |
67c4d9e6 | 8485 | |
b7ddce3c | 8486 | io_cancel_defer_files(ctx, files); |
67c4d9e6 PB |
8487 | /* cancel all at once, should be faster than doing it one by one*/ |
8488 | io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true); | |
8489 | ||
fcb323cc | 8490 | while (!list_empty_careful(&ctx->inflight_list)) { |
d8f1b971 XW |
8491 | struct io_kiocb *cancel_req = NULL, *req; |
8492 | DEFINE_WAIT(wait); | |
fcb323cc JA |
8493 | |
8494 | spin_lock_irq(&ctx->inflight_lock); | |
8495 | list_for_each_entry(req, &ctx->inflight_list, inflight_entry) { | |
0f212204 | 8496 | if (files && req->work.files != files) |
768134d4 JA |
8497 | continue; |
8498 | /* req is being completed, ignore */ | |
8499 | if (!refcount_inc_not_zero(&req->refs)) | |
8500 | continue; | |
8501 | cancel_req = req; | |
8502 | break; | |
fcb323cc | 8503 | } |
768134d4 | 8504 | if (cancel_req) |
fcb323cc | 8505 | prepare_to_wait(&ctx->inflight_wait, &wait, |
768134d4 | 8506 | TASK_UNINTERRUPTIBLE); |
fcb323cc JA |
8507 | spin_unlock_irq(&ctx->inflight_lock); |
8508 | ||
768134d4 JA |
8509 | /* We need to keep going until we don't find a matching req */ |
8510 | if (!cancel_req) | |
fcb323cc | 8511 | break; |
bb175342 PB |
8512 | /* cancel this request, or head link requests */ |
8513 | io_attempt_cancel(ctx, cancel_req); | |
8514 | io_put_req(cancel_req); | |
6200b0ae JA |
8515 | /* cancellations _may_ trigger task work */ |
8516 | io_run_task_work(); | |
fcb323cc | 8517 | schedule(); |
d8f1b971 | 8518 | finish_wait(&ctx->inflight_wait, &wait); |
fcb323cc | 8519 | } |
76e1b642 JA |
8520 | |
8521 | return true; | |
fcb323cc JA |
8522 | } |
8523 | ||
801dd57b | 8524 | static bool io_cancel_task_cb(struct io_wq_work *work, void *data) |
44e728b8 | 8525 | { |
801dd57b PB |
8526 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
8527 | struct task_struct *task = data; | |
44e728b8 | 8528 | |
f3606e3a | 8529 | return io_task_match(req, task); |
44e728b8 PB |
8530 | } |
8531 | ||
0f212204 JA |
8532 | static bool __io_uring_cancel_task_requests(struct io_ring_ctx *ctx, |
8533 | struct task_struct *task, | |
8534 | struct files_struct *files) | |
8535 | { | |
8536 | bool ret; | |
8537 | ||
8538 | ret = io_uring_cancel_files(ctx, files); | |
8539 | if (!files) { | |
8540 | enum io_wq_cancel cret; | |
8541 | ||
8542 | cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, task, true); | |
8543 | if (cret != IO_WQ_CANCEL_NOTFOUND) | |
8544 | ret = true; | |
8545 | ||
8546 | /* SQPOLL thread does its own polling */ | |
8547 | if (!(ctx->flags & IORING_SETUP_SQPOLL)) { | |
8548 | while (!list_empty_careful(&ctx->iopoll_list)) { | |
8549 | io_iopoll_try_reap_events(ctx); | |
8550 | ret = true; | |
8551 | } | |
8552 | } | |
8553 | ||
8554 | ret |= io_poll_remove_all(ctx, task); | |
8555 | ret |= io_kill_timeouts(ctx, task); | |
8556 | } | |
8557 | ||
8558 | return ret; | |
8559 | } | |
8560 | ||
8561 | /* | |
8562 | * We need to iteratively cancel requests, in case a request has dependent | |
8563 | * hard links. These persist even for failure of cancelations, hence keep | |
8564 | * looping until none are found. | |
8565 | */ | |
8566 | static void io_uring_cancel_task_requests(struct io_ring_ctx *ctx, | |
8567 | struct files_struct *files) | |
8568 | { | |
8569 | struct task_struct *task = current; | |
8570 | ||
534ca6d6 JA |
8571 | if ((ctx->flags & IORING_SETUP_SQPOLL) && ctx->sq_data) |
8572 | task = ctx->sq_data->thread; | |
0f212204 JA |
8573 | |
8574 | io_cqring_overflow_flush(ctx, true, task, files); | |
8575 | ||
8576 | while (__io_uring_cancel_task_requests(ctx, task, files)) { | |
8577 | io_run_task_work(); | |
8578 | cond_resched(); | |
8579 | } | |
8580 | } | |
8581 | ||
8582 | /* | |
8583 | * Note that this task has used io_uring. We use it for cancelation purposes. | |
8584 | */ | |
8585 | static int io_uring_add_task_file(struct file *file) | |
8586 | { | |
8587 | if (unlikely(!current->io_uring)) { | |
8588 | int ret; | |
8589 | ||
8590 | ret = io_uring_alloc_task_context(current); | |
8591 | if (unlikely(ret)) | |
8592 | return ret; | |
8593 | } | |
8594 | if (current->io_uring->last != file) { | |
8595 | XA_STATE(xas, ¤t->io_uring->xa, (unsigned long) file); | |
8596 | void *old; | |
8597 | ||
8598 | rcu_read_lock(); | |
8599 | old = xas_load(&xas); | |
8600 | if (old != file) { | |
8601 | get_file(file); | |
8602 | xas_lock(&xas); | |
8603 | xas_store(&xas, file); | |
8604 | xas_unlock(&xas); | |
8605 | } | |
8606 | rcu_read_unlock(); | |
8607 | current->io_uring->last = file; | |
8608 | } | |
8609 | ||
8610 | return 0; | |
8611 | } | |
8612 | ||
8613 | /* | |
8614 | * Remove this io_uring_file -> task mapping. | |
8615 | */ | |
8616 | static void io_uring_del_task_file(struct file *file) | |
8617 | { | |
8618 | struct io_uring_task *tctx = current->io_uring; | |
8619 | XA_STATE(xas, &tctx->xa, (unsigned long) file); | |
8620 | ||
8621 | if (tctx->last == file) | |
8622 | tctx->last = NULL; | |
8623 | ||
8624 | xas_lock(&xas); | |
8625 | file = xas_store(&xas, NULL); | |
8626 | xas_unlock(&xas); | |
8627 | ||
8628 | if (file) | |
8629 | fput(file); | |
8630 | } | |
8631 | ||
8632 | static void __io_uring_attempt_task_drop(struct file *file) | |
8633 | { | |
8634 | XA_STATE(xas, ¤t->io_uring->xa, (unsigned long) file); | |
8635 | struct file *old; | |
8636 | ||
8637 | rcu_read_lock(); | |
8638 | old = xas_load(&xas); | |
8639 | rcu_read_unlock(); | |
8640 | ||
8641 | if (old == file) | |
8642 | io_uring_del_task_file(file); | |
8643 | } | |
8644 | ||
8645 | /* | |
8646 | * Drop task note for this file if we're the only ones that hold it after | |
8647 | * pending fput() | |
8648 | */ | |
8649 | static void io_uring_attempt_task_drop(struct file *file, bool exiting) | |
8650 | { | |
8651 | if (!current->io_uring) | |
8652 | return; | |
8653 | /* | |
8654 | * fput() is pending, will be 2 if the only other ref is our potential | |
8655 | * task file note. If the task is exiting, drop regardless of count. | |
8656 | */ | |
8657 | if (!exiting && atomic_long_read(&file->f_count) != 2) | |
8658 | return; | |
8659 | ||
8660 | __io_uring_attempt_task_drop(file); | |
8661 | } | |
8662 | ||
8663 | void __io_uring_files_cancel(struct files_struct *files) | |
8664 | { | |
8665 | struct io_uring_task *tctx = current->io_uring; | |
8666 | XA_STATE(xas, &tctx->xa, 0); | |
8667 | ||
8668 | /* make sure overflow events are dropped */ | |
8669 | tctx->in_idle = true; | |
8670 | ||
8671 | do { | |
8672 | struct io_ring_ctx *ctx; | |
8673 | struct file *file; | |
8674 | ||
8675 | xas_lock(&xas); | |
8676 | file = xas_next_entry(&xas, ULONG_MAX); | |
8677 | xas_unlock(&xas); | |
8678 | ||
8679 | if (!file) | |
8680 | break; | |
8681 | ||
8682 | ctx = file->private_data; | |
8683 | ||
8684 | io_uring_cancel_task_requests(ctx, files); | |
8685 | if (files) | |
8686 | io_uring_del_task_file(file); | |
8687 | } while (1); | |
8688 | } | |
8689 | ||
8690 | static inline bool io_uring_task_idle(struct io_uring_task *tctx) | |
8691 | { | |
8692 | return atomic_long_read(&tctx->req_issue) == | |
8693 | atomic_long_read(&tctx->req_complete); | |
8694 | } | |
8695 | ||
8696 | /* | |
8697 | * Find any io_uring fd that this task has registered or done IO on, and cancel | |
8698 | * requests. | |
8699 | */ | |
8700 | void __io_uring_task_cancel(void) | |
8701 | { | |
8702 | struct io_uring_task *tctx = current->io_uring; | |
8703 | DEFINE_WAIT(wait); | |
8704 | long completions; | |
8705 | ||
8706 | /* make sure overflow events are dropped */ | |
8707 | tctx->in_idle = true; | |
8708 | ||
8709 | while (!io_uring_task_idle(tctx)) { | |
8710 | /* read completions before cancelations */ | |
8711 | completions = atomic_long_read(&tctx->req_complete); | |
8712 | __io_uring_files_cancel(NULL); | |
8713 | ||
8714 | prepare_to_wait(&tctx->wait, &wait, TASK_UNINTERRUPTIBLE); | |
8715 | ||
8716 | /* | |
8717 | * If we've seen completions, retry. This avoids a race where | |
8718 | * a completion comes in before we did prepare_to_wait(). | |
8719 | */ | |
8720 | if (completions != atomic_long_read(&tctx->req_complete)) | |
8721 | continue; | |
8722 | if (io_uring_task_idle(tctx)) | |
8723 | break; | |
8724 | schedule(); | |
8725 | } | |
8726 | ||
8727 | finish_wait(&tctx->wait, &wait); | |
8728 | tctx->in_idle = false; | |
8729 | } | |
8730 | ||
fcb323cc JA |
8731 | static int io_uring_flush(struct file *file, void *data) |
8732 | { | |
8733 | struct io_ring_ctx *ctx = file->private_data; | |
8734 | ||
6ab23144 JA |
8735 | /* |
8736 | * If the task is going away, cancel work it may have pending | |
8737 | */ | |
801dd57b | 8738 | if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) |
0f212204 | 8739 | data = NULL; |
6ab23144 | 8740 | |
0f212204 JA |
8741 | io_uring_cancel_task_requests(ctx, data); |
8742 | io_uring_attempt_task_drop(file, !data); | |
fcb323cc JA |
8743 | return 0; |
8744 | } | |
8745 | ||
6c5c240e RP |
8746 | static void *io_uring_validate_mmap_request(struct file *file, |
8747 | loff_t pgoff, size_t sz) | |
2b188cc1 | 8748 | { |
2b188cc1 | 8749 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 8750 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
8751 | struct page *page; |
8752 | void *ptr; | |
8753 | ||
8754 | switch (offset) { | |
8755 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
8756 | case IORING_OFF_CQ_RING: |
8757 | ptr = ctx->rings; | |
2b188cc1 JA |
8758 | break; |
8759 | case IORING_OFF_SQES: | |
8760 | ptr = ctx->sq_sqes; | |
8761 | break; | |
2b188cc1 | 8762 | default: |
6c5c240e | 8763 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
8764 | } |
8765 | ||
8766 | page = virt_to_head_page(ptr); | |
a50b854e | 8767 | if (sz > page_size(page)) |
6c5c240e RP |
8768 | return ERR_PTR(-EINVAL); |
8769 | ||
8770 | return ptr; | |
8771 | } | |
8772 | ||
8773 | #ifdef CONFIG_MMU | |
8774 | ||
8775 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
8776 | { | |
8777 | size_t sz = vma->vm_end - vma->vm_start; | |
8778 | unsigned long pfn; | |
8779 | void *ptr; | |
8780 | ||
8781 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
8782 | if (IS_ERR(ptr)) | |
8783 | return PTR_ERR(ptr); | |
2b188cc1 JA |
8784 | |
8785 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
8786 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
8787 | } | |
8788 | ||
6c5c240e RP |
8789 | #else /* !CONFIG_MMU */ |
8790 | ||
8791 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
8792 | { | |
8793 | return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL; | |
8794 | } | |
8795 | ||
8796 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
8797 | { | |
8798 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
8799 | } | |
8800 | ||
8801 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
8802 | unsigned long addr, unsigned long len, | |
8803 | unsigned long pgoff, unsigned long flags) | |
8804 | { | |
8805 | void *ptr; | |
8806 | ||
8807 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
8808 | if (IS_ERR(ptr)) | |
8809 | return PTR_ERR(ptr); | |
8810 | ||
8811 | return (unsigned long) ptr; | |
8812 | } | |
8813 | ||
8814 | #endif /* !CONFIG_MMU */ | |
8815 | ||
90554200 JA |
8816 | static void io_sqpoll_wait_sq(struct io_ring_ctx *ctx) |
8817 | { | |
8818 | DEFINE_WAIT(wait); | |
8819 | ||
8820 | do { | |
8821 | if (!io_sqring_full(ctx)) | |
8822 | break; | |
8823 | ||
8824 | prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE); | |
8825 | ||
8826 | if (!io_sqring_full(ctx)) | |
8827 | break; | |
8828 | ||
8829 | schedule(); | |
8830 | } while (!signal_pending(current)); | |
8831 | ||
8832 | finish_wait(&ctx->sqo_sq_wait, &wait); | |
8833 | } | |
8834 | ||
2b188cc1 JA |
8835 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
8836 | u32, min_complete, u32, flags, const sigset_t __user *, sig, | |
8837 | size_t, sigsz) | |
8838 | { | |
8839 | struct io_ring_ctx *ctx; | |
8840 | long ret = -EBADF; | |
8841 | int submitted = 0; | |
8842 | struct fd f; | |
8843 | ||
4c6e277c | 8844 | io_run_task_work(); |
b41e9852 | 8845 | |
90554200 JA |
8846 | if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP | |
8847 | IORING_ENTER_SQ_WAIT)) | |
2b188cc1 JA |
8848 | return -EINVAL; |
8849 | ||
8850 | f = fdget(fd); | |
8851 | if (!f.file) | |
8852 | return -EBADF; | |
8853 | ||
8854 | ret = -EOPNOTSUPP; | |
8855 | if (f.file->f_op != &io_uring_fops) | |
8856 | goto out_fput; | |
8857 | ||
8858 | ret = -ENXIO; | |
8859 | ctx = f.file->private_data; | |
8860 | if (!percpu_ref_tryget(&ctx->refs)) | |
8861 | goto out_fput; | |
8862 | ||
7e84e1c7 SG |
8863 | ret = -EBADFD; |
8864 | if (ctx->flags & IORING_SETUP_R_DISABLED) | |
8865 | goto out; | |
8866 | ||
6c271ce2 JA |
8867 | /* |
8868 | * For SQ polling, the thread will do all submissions and completions. | |
8869 | * Just return the requested submit count, and wake the thread if | |
8870 | * we were asked to. | |
8871 | */ | |
b2a9eada | 8872 | ret = 0; |
6c271ce2 | 8873 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
c1edbf5f | 8874 | if (!list_empty_careful(&ctx->cq_overflow_list)) |
e6c8aa9a | 8875 | io_cqring_overflow_flush(ctx, false, NULL, NULL); |
6c271ce2 | 8876 | if (flags & IORING_ENTER_SQ_WAKEUP) |
534ca6d6 | 8877 | wake_up(&ctx->sq_data->wait); |
90554200 JA |
8878 | if (flags & IORING_ENTER_SQ_WAIT) |
8879 | io_sqpoll_wait_sq(ctx); | |
6c271ce2 | 8880 | submitted = to_submit; |
b2a9eada | 8881 | } else if (to_submit) { |
0f212204 JA |
8882 | ret = io_uring_add_task_file(f.file); |
8883 | if (unlikely(ret)) | |
8884 | goto out; | |
2b188cc1 | 8885 | mutex_lock(&ctx->uring_lock); |
0f212204 | 8886 | submitted = io_submit_sqes(ctx, to_submit); |
2b188cc1 | 8887 | mutex_unlock(&ctx->uring_lock); |
7c504e65 PB |
8888 | |
8889 | if (submitted != to_submit) | |
8890 | goto out; | |
2b188cc1 JA |
8891 | } |
8892 | if (flags & IORING_ENTER_GETEVENTS) { | |
8893 | min_complete = min(min_complete, ctx->cq_entries); | |
8894 | ||
32b2244a XW |
8895 | /* |
8896 | * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user | |
8897 | * space applications don't need to do io completion events | |
8898 | * polling again, they can rely on io_sq_thread to do polling | |
8899 | * work, which can reduce cpu usage and uring_lock contention. | |
8900 | */ | |
8901 | if (ctx->flags & IORING_SETUP_IOPOLL && | |
8902 | !(ctx->flags & IORING_SETUP_SQPOLL)) { | |
7668b92a | 8903 | ret = io_iopoll_check(ctx, min_complete); |
def596e9 JA |
8904 | } else { |
8905 | ret = io_cqring_wait(ctx, min_complete, sig, sigsz); | |
8906 | } | |
2b188cc1 JA |
8907 | } |
8908 | ||
7c504e65 | 8909 | out: |
6805b32e | 8910 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
8911 | out_fput: |
8912 | fdput(f); | |
8913 | return submitted ? submitted : ret; | |
8914 | } | |
8915 | ||
bebdb65e | 8916 | #ifdef CONFIG_PROC_FS |
87ce955b JA |
8917 | static int io_uring_show_cred(int id, void *p, void *data) |
8918 | { | |
8919 | const struct cred *cred = p; | |
8920 | struct seq_file *m = data; | |
8921 | struct user_namespace *uns = seq_user_ns(m); | |
8922 | struct group_info *gi; | |
8923 | kernel_cap_t cap; | |
8924 | unsigned __capi; | |
8925 | int g; | |
8926 | ||
8927 | seq_printf(m, "%5d\n", id); | |
8928 | seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid)); | |
8929 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid)); | |
8930 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid)); | |
8931 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid)); | |
8932 | seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid)); | |
8933 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid)); | |
8934 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid)); | |
8935 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid)); | |
8936 | seq_puts(m, "\n\tGroups:\t"); | |
8937 | gi = cred->group_info; | |
8938 | for (g = 0; g < gi->ngroups; g++) { | |
8939 | seq_put_decimal_ull(m, g ? " " : "", | |
8940 | from_kgid_munged(uns, gi->gid[g])); | |
8941 | } | |
8942 | seq_puts(m, "\n\tCapEff:\t"); | |
8943 | cap = cred->cap_effective; | |
8944 | CAP_FOR_EACH_U32(__capi) | |
8945 | seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8); | |
8946 | seq_putc(m, '\n'); | |
8947 | return 0; | |
8948 | } | |
8949 | ||
8950 | static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m) | |
8951 | { | |
fad8e0de | 8952 | bool has_lock; |
87ce955b JA |
8953 | int i; |
8954 | ||
fad8e0de JA |
8955 | /* |
8956 | * Avoid ABBA deadlock between the seq lock and the io_uring mutex, | |
8957 | * since fdinfo case grabs it in the opposite direction of normal use | |
8958 | * cases. If we fail to get the lock, we just don't iterate any | |
8959 | * structures that could be going away outside the io_uring mutex. | |
8960 | */ | |
8961 | has_lock = mutex_trylock(&ctx->uring_lock); | |
8962 | ||
87ce955b | 8963 | seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files); |
fad8e0de | 8964 | for (i = 0; has_lock && i < ctx->nr_user_files; i++) { |
87ce955b JA |
8965 | struct fixed_file_table *table; |
8966 | struct file *f; | |
8967 | ||
8968 | table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT]; | |
8969 | f = table->files[i & IORING_FILE_TABLE_MASK]; | |
8970 | if (f) | |
8971 | seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname); | |
8972 | else | |
8973 | seq_printf(m, "%5u: <none>\n", i); | |
8974 | } | |
8975 | seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs); | |
fad8e0de | 8976 | for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) { |
87ce955b JA |
8977 | struct io_mapped_ubuf *buf = &ctx->user_bufs[i]; |
8978 | ||
8979 | seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, | |
8980 | (unsigned int) buf->len); | |
8981 | } | |
fad8e0de | 8982 | if (has_lock && !idr_is_empty(&ctx->personality_idr)) { |
87ce955b JA |
8983 | seq_printf(m, "Personalities:\n"); |
8984 | idr_for_each(&ctx->personality_idr, io_uring_show_cred, m); | |
8985 | } | |
d7718a9d JA |
8986 | seq_printf(m, "PollList:\n"); |
8987 | spin_lock_irq(&ctx->completion_lock); | |
8988 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { | |
8989 | struct hlist_head *list = &ctx->cancel_hash[i]; | |
8990 | struct io_kiocb *req; | |
8991 | ||
8992 | hlist_for_each_entry(req, list, hash_node) | |
8993 | seq_printf(m, " op=%d, task_works=%d\n", req->opcode, | |
8994 | req->task->task_works != NULL); | |
8995 | } | |
8996 | spin_unlock_irq(&ctx->completion_lock); | |
fad8e0de JA |
8997 | if (has_lock) |
8998 | mutex_unlock(&ctx->uring_lock); | |
87ce955b JA |
8999 | } |
9000 | ||
9001 | static void io_uring_show_fdinfo(struct seq_file *m, struct file *f) | |
9002 | { | |
9003 | struct io_ring_ctx *ctx = f->private_data; | |
9004 | ||
9005 | if (percpu_ref_tryget(&ctx->refs)) { | |
9006 | __io_uring_show_fdinfo(ctx, m); | |
9007 | percpu_ref_put(&ctx->refs); | |
9008 | } | |
9009 | } | |
bebdb65e | 9010 | #endif |
87ce955b | 9011 | |
2b188cc1 JA |
9012 | static const struct file_operations io_uring_fops = { |
9013 | .release = io_uring_release, | |
fcb323cc | 9014 | .flush = io_uring_flush, |
2b188cc1 | 9015 | .mmap = io_uring_mmap, |
6c5c240e RP |
9016 | #ifndef CONFIG_MMU |
9017 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
9018 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
9019 | #endif | |
2b188cc1 JA |
9020 | .poll = io_uring_poll, |
9021 | .fasync = io_uring_fasync, | |
bebdb65e | 9022 | #ifdef CONFIG_PROC_FS |
87ce955b | 9023 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 9024 | #endif |
2b188cc1 JA |
9025 | }; |
9026 | ||
9027 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
9028 | struct io_uring_params *p) | |
9029 | { | |
75b28aff HV |
9030 | struct io_rings *rings; |
9031 | size_t size, sq_array_offset; | |
2b188cc1 | 9032 | |
bd740481 JA |
9033 | /* make sure these are sane, as we already accounted them */ |
9034 | ctx->sq_entries = p->sq_entries; | |
9035 | ctx->cq_entries = p->cq_entries; | |
9036 | ||
75b28aff HV |
9037 | size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset); |
9038 | if (size == SIZE_MAX) | |
9039 | return -EOVERFLOW; | |
9040 | ||
9041 | rings = io_mem_alloc(size); | |
9042 | if (!rings) | |
2b188cc1 JA |
9043 | return -ENOMEM; |
9044 | ||
75b28aff HV |
9045 | ctx->rings = rings; |
9046 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
9047 | rings->sq_ring_mask = p->sq_entries - 1; | |
9048 | rings->cq_ring_mask = p->cq_entries - 1; | |
9049 | rings->sq_ring_entries = p->sq_entries; | |
9050 | rings->cq_ring_entries = p->cq_entries; | |
9051 | ctx->sq_mask = rings->sq_ring_mask; | |
9052 | ctx->cq_mask = rings->cq_ring_mask; | |
2b188cc1 JA |
9053 | |
9054 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 JA |
9055 | if (size == SIZE_MAX) { |
9056 | io_mem_free(ctx->rings); | |
9057 | ctx->rings = NULL; | |
2b188cc1 | 9058 | return -EOVERFLOW; |
eb065d30 | 9059 | } |
2b188cc1 JA |
9060 | |
9061 | ctx->sq_sqes = io_mem_alloc(size); | |
eb065d30 JA |
9062 | if (!ctx->sq_sqes) { |
9063 | io_mem_free(ctx->rings); | |
9064 | ctx->rings = NULL; | |
2b188cc1 | 9065 | return -ENOMEM; |
eb065d30 | 9066 | } |
2b188cc1 | 9067 | |
2b188cc1 JA |
9068 | return 0; |
9069 | } | |
9070 | ||
9071 | /* | |
9072 | * Allocate an anonymous fd, this is what constitutes the application | |
9073 | * visible backing of an io_uring instance. The application mmaps this | |
9074 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
9075 | * we have to tie this fd to a socket for file garbage collection purposes. | |
9076 | */ | |
9077 | static int io_uring_get_fd(struct io_ring_ctx *ctx) | |
9078 | { | |
9079 | struct file *file; | |
9080 | int ret; | |
9081 | ||
9082 | #if defined(CONFIG_UNIX) | |
9083 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, | |
9084 | &ctx->ring_sock); | |
9085 | if (ret) | |
9086 | return ret; | |
9087 | #endif | |
9088 | ||
9089 | ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
9090 | if (ret < 0) | |
9091 | goto err; | |
9092 | ||
9093 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, | |
9094 | O_RDWR | O_CLOEXEC); | |
9095 | if (IS_ERR(file)) { | |
0f212204 | 9096 | err_fd: |
2b188cc1 JA |
9097 | put_unused_fd(ret); |
9098 | ret = PTR_ERR(file); | |
9099 | goto err; | |
9100 | } | |
9101 | ||
9102 | #if defined(CONFIG_UNIX) | |
9103 | ctx->ring_sock->file = file; | |
9104 | #endif | |
0f212204 JA |
9105 | if (unlikely(io_uring_add_task_file(file))) { |
9106 | file = ERR_PTR(-ENOMEM); | |
9107 | goto err_fd; | |
9108 | } | |
2b188cc1 JA |
9109 | fd_install(ret, file); |
9110 | return ret; | |
9111 | err: | |
9112 | #if defined(CONFIG_UNIX) | |
9113 | sock_release(ctx->ring_sock); | |
9114 | ctx->ring_sock = NULL; | |
9115 | #endif | |
9116 | return ret; | |
9117 | } | |
9118 | ||
7f13657d XW |
9119 | static int io_uring_create(unsigned entries, struct io_uring_params *p, |
9120 | struct io_uring_params __user *params) | |
2b188cc1 JA |
9121 | { |
9122 | struct user_struct *user = NULL; | |
9123 | struct io_ring_ctx *ctx; | |
aad5d8da | 9124 | bool limit_mem; |
2b188cc1 JA |
9125 | int ret; |
9126 | ||
8110c1a6 | 9127 | if (!entries) |
2b188cc1 | 9128 | return -EINVAL; |
8110c1a6 JA |
9129 | if (entries > IORING_MAX_ENTRIES) { |
9130 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
9131 | return -EINVAL; | |
9132 | entries = IORING_MAX_ENTRIES; | |
9133 | } | |
2b188cc1 JA |
9134 | |
9135 | /* | |
9136 | * Use twice as many entries for the CQ ring. It's possible for the | |
9137 | * application to drive a higher depth than the size of the SQ ring, | |
9138 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
9139 | * some flexibility in overcommitting a bit. If the application has |
9140 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
9141 | * of CQ ring entries manually. | |
2b188cc1 JA |
9142 | */ |
9143 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
9144 | if (p->flags & IORING_SETUP_CQSIZE) { |
9145 | /* | |
9146 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
9147 | * to a power-of-two, if it isn't already. We do NOT impose | |
9148 | * any cq vs sq ring sizing. | |
9149 | */ | |
8110c1a6 | 9150 | if (p->cq_entries < p->sq_entries) |
33a107f0 | 9151 | return -EINVAL; |
8110c1a6 JA |
9152 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
9153 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
9154 | return -EINVAL; | |
9155 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
9156 | } | |
33a107f0 JA |
9157 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
9158 | } else { | |
9159 | p->cq_entries = 2 * p->sq_entries; | |
9160 | } | |
2b188cc1 JA |
9161 | |
9162 | user = get_uid(current_user()); | |
aad5d8da | 9163 | limit_mem = !capable(CAP_IPC_LOCK); |
2b188cc1 | 9164 | |
aad5d8da | 9165 | if (limit_mem) { |
a087e2b5 | 9166 | ret = __io_account_mem(user, |
2b188cc1 JA |
9167 | ring_pages(p->sq_entries, p->cq_entries)); |
9168 | if (ret) { | |
9169 | free_uid(user); | |
9170 | return ret; | |
9171 | } | |
9172 | } | |
9173 | ||
9174 | ctx = io_ring_ctx_alloc(p); | |
9175 | if (!ctx) { | |
aad5d8da | 9176 | if (limit_mem) |
a087e2b5 | 9177 | __io_unaccount_mem(user, ring_pages(p->sq_entries, |
2b188cc1 JA |
9178 | p->cq_entries)); |
9179 | free_uid(user); | |
9180 | return -ENOMEM; | |
9181 | } | |
9182 | ctx->compat = in_compat_syscall(); | |
2b188cc1 | 9183 | ctx->user = user; |
0b8c0ec7 | 9184 | ctx->creds = get_current_cred(); |
2b188cc1 | 9185 | |
2aede0e4 JA |
9186 | ctx->sqo_task = get_task_struct(current); |
9187 | ||
9188 | /* | |
9189 | * This is just grabbed for accounting purposes. When a process exits, | |
9190 | * the mm is exited and dropped before the files, hence we need to hang | |
9191 | * on to this mm purely for the purposes of being able to unaccount | |
9192 | * memory (locked/pinned vm). It's not used for anything else. | |
9193 | */ | |
6b7898eb | 9194 | mmgrab(current->mm); |
2aede0e4 | 9195 | ctx->mm_account = current->mm; |
6b7898eb | 9196 | |
f74441e6 JA |
9197 | /* |
9198 | * Account memory _before_ installing the file descriptor. Once | |
9199 | * the descriptor is installed, it can get closed at any time. Also | |
9200 | * do this before hitting the general error path, as ring freeing | |
9201 | * will un-account as well. | |
9202 | */ | |
9203 | io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries), | |
9204 | ACCT_LOCKED); | |
9205 | ctx->limit_mem = limit_mem; | |
9206 | ||
2b188cc1 JA |
9207 | ret = io_allocate_scq_urings(ctx, p); |
9208 | if (ret) | |
9209 | goto err; | |
9210 | ||
7e84e1c7 | 9211 | ret = io_sq_offload_create(ctx, p); |
2b188cc1 JA |
9212 | if (ret) |
9213 | goto err; | |
9214 | ||
7e84e1c7 SG |
9215 | if (!(p->flags & IORING_SETUP_R_DISABLED)) |
9216 | io_sq_offload_start(ctx); | |
9217 | ||
2b188cc1 | 9218 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
9219 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
9220 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
9221 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
9222 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
9223 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
9224 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
9225 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
9226 | |
9227 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
9228 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
9229 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
9230 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
9231 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
9232 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
9233 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
0d9b5b3a | 9234 | p->cq_off.flags = offsetof(struct io_rings, cq_flags); |
ac90f249 | 9235 | |
7f13657d XW |
9236 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
9237 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | | |
5769a351 JX |
9238 | IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | |
9239 | IORING_FEAT_POLL_32BITS; | |
7f13657d XW |
9240 | |
9241 | if (copy_to_user(params, p, sizeof(*p))) { | |
9242 | ret = -EFAULT; | |
9243 | goto err; | |
9244 | } | |
d1719f70 | 9245 | |
044c1ab3 JA |
9246 | /* |
9247 | * Install ring fd as the very last thing, so we don't risk someone | |
9248 | * having closed it before we finish setup | |
9249 | */ | |
9250 | ret = io_uring_get_fd(ctx); | |
9251 | if (ret < 0) | |
9252 | goto err; | |
9253 | ||
c826bd7a | 9254 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
9255 | return ret; |
9256 | err: | |
9257 | io_ring_ctx_wait_and_kill(ctx); | |
9258 | return ret; | |
9259 | } | |
9260 | ||
9261 | /* | |
9262 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
9263 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
9264 | * params structure passed in. | |
9265 | */ | |
9266 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
9267 | { | |
9268 | struct io_uring_params p; | |
2b188cc1 JA |
9269 | int i; |
9270 | ||
9271 | if (copy_from_user(&p, params, sizeof(p))) | |
9272 | return -EFAULT; | |
9273 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
9274 | if (p.resv[i]) | |
9275 | return -EINVAL; | |
9276 | } | |
9277 | ||
6c271ce2 | 9278 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 9279 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
7e84e1c7 SG |
9280 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ | |
9281 | IORING_SETUP_R_DISABLED)) | |
2b188cc1 JA |
9282 | return -EINVAL; |
9283 | ||
7f13657d | 9284 | return io_uring_create(entries, &p, params); |
2b188cc1 JA |
9285 | } |
9286 | ||
9287 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
9288 | struct io_uring_params __user *, params) | |
9289 | { | |
9290 | return io_uring_setup(entries, params); | |
9291 | } | |
9292 | ||
66f4af93 JA |
9293 | static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args) |
9294 | { | |
9295 | struct io_uring_probe *p; | |
9296 | size_t size; | |
9297 | int i, ret; | |
9298 | ||
9299 | size = struct_size(p, ops, nr_args); | |
9300 | if (size == SIZE_MAX) | |
9301 | return -EOVERFLOW; | |
9302 | p = kzalloc(size, GFP_KERNEL); | |
9303 | if (!p) | |
9304 | return -ENOMEM; | |
9305 | ||
9306 | ret = -EFAULT; | |
9307 | if (copy_from_user(p, arg, size)) | |
9308 | goto out; | |
9309 | ret = -EINVAL; | |
9310 | if (memchr_inv(p, 0, size)) | |
9311 | goto out; | |
9312 | ||
9313 | p->last_op = IORING_OP_LAST - 1; | |
9314 | if (nr_args > IORING_OP_LAST) | |
9315 | nr_args = IORING_OP_LAST; | |
9316 | ||
9317 | for (i = 0; i < nr_args; i++) { | |
9318 | p->ops[i].op = i; | |
9319 | if (!io_op_defs[i].not_supported) | |
9320 | p->ops[i].flags = IO_URING_OP_SUPPORTED; | |
9321 | } | |
9322 | p->ops_len = i; | |
9323 | ||
9324 | ret = 0; | |
9325 | if (copy_to_user(arg, p, size)) | |
9326 | ret = -EFAULT; | |
9327 | out: | |
9328 | kfree(p); | |
9329 | return ret; | |
9330 | } | |
9331 | ||
071698e1 JA |
9332 | static int io_register_personality(struct io_ring_ctx *ctx) |
9333 | { | |
9334 | const struct cred *creds = get_current_cred(); | |
9335 | int id; | |
9336 | ||
9337 | id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1, | |
9338 | USHRT_MAX, GFP_KERNEL); | |
9339 | if (id < 0) | |
9340 | put_cred(creds); | |
9341 | return id; | |
9342 | } | |
9343 | ||
9344 | static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) | |
9345 | { | |
9346 | const struct cred *old_creds; | |
9347 | ||
9348 | old_creds = idr_remove(&ctx->personality_idr, id); | |
9349 | if (old_creds) { | |
9350 | put_cred(old_creds); | |
9351 | return 0; | |
9352 | } | |
9353 | ||
9354 | return -EINVAL; | |
9355 | } | |
9356 | ||
21b55dbc SG |
9357 | static int io_register_restrictions(struct io_ring_ctx *ctx, void __user *arg, |
9358 | unsigned int nr_args) | |
9359 | { | |
9360 | struct io_uring_restriction *res; | |
9361 | size_t size; | |
9362 | int i, ret; | |
9363 | ||
7e84e1c7 SG |
9364 | /* Restrictions allowed only if rings started disabled */ |
9365 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
9366 | return -EBADFD; | |
9367 | ||
21b55dbc | 9368 | /* We allow only a single restrictions registration */ |
7e84e1c7 | 9369 | if (ctx->restrictions.registered) |
21b55dbc SG |
9370 | return -EBUSY; |
9371 | ||
9372 | if (!arg || nr_args > IORING_MAX_RESTRICTIONS) | |
9373 | return -EINVAL; | |
9374 | ||
9375 | size = array_size(nr_args, sizeof(*res)); | |
9376 | if (size == SIZE_MAX) | |
9377 | return -EOVERFLOW; | |
9378 | ||
9379 | res = memdup_user(arg, size); | |
9380 | if (IS_ERR(res)) | |
9381 | return PTR_ERR(res); | |
9382 | ||
9383 | ret = 0; | |
9384 | ||
9385 | for (i = 0; i < nr_args; i++) { | |
9386 | switch (res[i].opcode) { | |
9387 | case IORING_RESTRICTION_REGISTER_OP: | |
9388 | if (res[i].register_op >= IORING_REGISTER_LAST) { | |
9389 | ret = -EINVAL; | |
9390 | goto out; | |
9391 | } | |
9392 | ||
9393 | __set_bit(res[i].register_op, | |
9394 | ctx->restrictions.register_op); | |
9395 | break; | |
9396 | case IORING_RESTRICTION_SQE_OP: | |
9397 | if (res[i].sqe_op >= IORING_OP_LAST) { | |
9398 | ret = -EINVAL; | |
9399 | goto out; | |
9400 | } | |
9401 | ||
9402 | __set_bit(res[i].sqe_op, ctx->restrictions.sqe_op); | |
9403 | break; | |
9404 | case IORING_RESTRICTION_SQE_FLAGS_ALLOWED: | |
9405 | ctx->restrictions.sqe_flags_allowed = res[i].sqe_flags; | |
9406 | break; | |
9407 | case IORING_RESTRICTION_SQE_FLAGS_REQUIRED: | |
9408 | ctx->restrictions.sqe_flags_required = res[i].sqe_flags; | |
9409 | break; | |
9410 | default: | |
9411 | ret = -EINVAL; | |
9412 | goto out; | |
9413 | } | |
9414 | } | |
9415 | ||
9416 | out: | |
9417 | /* Reset all restrictions if an error happened */ | |
9418 | if (ret != 0) | |
9419 | memset(&ctx->restrictions, 0, sizeof(ctx->restrictions)); | |
9420 | else | |
7e84e1c7 | 9421 | ctx->restrictions.registered = true; |
21b55dbc SG |
9422 | |
9423 | kfree(res); | |
9424 | return ret; | |
9425 | } | |
9426 | ||
7e84e1c7 SG |
9427 | static int io_register_enable_rings(struct io_ring_ctx *ctx) |
9428 | { | |
9429 | if (!(ctx->flags & IORING_SETUP_R_DISABLED)) | |
9430 | return -EBADFD; | |
9431 | ||
9432 | if (ctx->restrictions.registered) | |
9433 | ctx->restricted = 1; | |
9434 | ||
9435 | ctx->flags &= ~IORING_SETUP_R_DISABLED; | |
9436 | ||
9437 | io_sq_offload_start(ctx); | |
9438 | ||
9439 | return 0; | |
9440 | } | |
9441 | ||
071698e1 JA |
9442 | static bool io_register_op_must_quiesce(int op) |
9443 | { | |
9444 | switch (op) { | |
9445 | case IORING_UNREGISTER_FILES: | |
9446 | case IORING_REGISTER_FILES_UPDATE: | |
9447 | case IORING_REGISTER_PROBE: | |
9448 | case IORING_REGISTER_PERSONALITY: | |
9449 | case IORING_UNREGISTER_PERSONALITY: | |
9450 | return false; | |
9451 | default: | |
9452 | return true; | |
9453 | } | |
9454 | } | |
9455 | ||
edafccee JA |
9456 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
9457 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
9458 | __releases(ctx->uring_lock) |
9459 | __acquires(ctx->uring_lock) | |
edafccee JA |
9460 | { |
9461 | int ret; | |
9462 | ||
35fa71a0 JA |
9463 | /* |
9464 | * We're inside the ring mutex, if the ref is already dying, then | |
9465 | * someone else killed the ctx or is already going through | |
9466 | * io_uring_register(). | |
9467 | */ | |
9468 | if (percpu_ref_is_dying(&ctx->refs)) | |
9469 | return -ENXIO; | |
9470 | ||
071698e1 | 9471 | if (io_register_op_must_quiesce(opcode)) { |
05f3fb3c | 9472 | percpu_ref_kill(&ctx->refs); |
b19062a5 | 9473 | |
05f3fb3c JA |
9474 | /* |
9475 | * Drop uring mutex before waiting for references to exit. If | |
9476 | * another thread is currently inside io_uring_enter() it might | |
9477 | * need to grab the uring_lock to make progress. If we hold it | |
9478 | * here across the drain wait, then we can deadlock. It's safe | |
9479 | * to drop the mutex here, since no new references will come in | |
9480 | * after we've killed the percpu ref. | |
9481 | */ | |
9482 | mutex_unlock(&ctx->uring_lock); | |
0f158b4c | 9483 | ret = wait_for_completion_interruptible(&ctx->ref_comp); |
05f3fb3c | 9484 | mutex_lock(&ctx->uring_lock); |
c150368b JA |
9485 | if (ret) { |
9486 | percpu_ref_resurrect(&ctx->refs); | |
9487 | ret = -EINTR; | |
21b55dbc SG |
9488 | goto out_quiesce; |
9489 | } | |
9490 | } | |
9491 | ||
9492 | if (ctx->restricted) { | |
9493 | if (opcode >= IORING_REGISTER_LAST) { | |
9494 | ret = -EINVAL; | |
9495 | goto out; | |
9496 | } | |
9497 | ||
9498 | if (!test_bit(opcode, ctx->restrictions.register_op)) { | |
9499 | ret = -EACCES; | |
c150368b JA |
9500 | goto out; |
9501 | } | |
05f3fb3c | 9502 | } |
edafccee JA |
9503 | |
9504 | switch (opcode) { | |
9505 | case IORING_REGISTER_BUFFERS: | |
9506 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
9507 | break; | |
9508 | case IORING_UNREGISTER_BUFFERS: | |
9509 | ret = -EINVAL; | |
9510 | if (arg || nr_args) | |
9511 | break; | |
9512 | ret = io_sqe_buffer_unregister(ctx); | |
9513 | break; | |
6b06314c JA |
9514 | case IORING_REGISTER_FILES: |
9515 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
9516 | break; | |
9517 | case IORING_UNREGISTER_FILES: | |
9518 | ret = -EINVAL; | |
9519 | if (arg || nr_args) | |
9520 | break; | |
9521 | ret = io_sqe_files_unregister(ctx); | |
9522 | break; | |
c3a31e60 JA |
9523 | case IORING_REGISTER_FILES_UPDATE: |
9524 | ret = io_sqe_files_update(ctx, arg, nr_args); | |
9525 | break; | |
9b402849 | 9526 | case IORING_REGISTER_EVENTFD: |
f2842ab5 | 9527 | case IORING_REGISTER_EVENTFD_ASYNC: |
9b402849 JA |
9528 | ret = -EINVAL; |
9529 | if (nr_args != 1) | |
9530 | break; | |
9531 | ret = io_eventfd_register(ctx, arg); | |
f2842ab5 JA |
9532 | if (ret) |
9533 | break; | |
9534 | if (opcode == IORING_REGISTER_EVENTFD_ASYNC) | |
9535 | ctx->eventfd_async = 1; | |
9536 | else | |
9537 | ctx->eventfd_async = 0; | |
9b402849 JA |
9538 | break; |
9539 | case IORING_UNREGISTER_EVENTFD: | |
9540 | ret = -EINVAL; | |
9541 | if (arg || nr_args) | |
9542 | break; | |
9543 | ret = io_eventfd_unregister(ctx); | |
9544 | break; | |
66f4af93 JA |
9545 | case IORING_REGISTER_PROBE: |
9546 | ret = -EINVAL; | |
9547 | if (!arg || nr_args > 256) | |
9548 | break; | |
9549 | ret = io_probe(ctx, arg, nr_args); | |
9550 | break; | |
071698e1 JA |
9551 | case IORING_REGISTER_PERSONALITY: |
9552 | ret = -EINVAL; | |
9553 | if (arg || nr_args) | |
9554 | break; | |
9555 | ret = io_register_personality(ctx); | |
9556 | break; | |
9557 | case IORING_UNREGISTER_PERSONALITY: | |
9558 | ret = -EINVAL; | |
9559 | if (arg) | |
9560 | break; | |
9561 | ret = io_unregister_personality(ctx, nr_args); | |
9562 | break; | |
7e84e1c7 SG |
9563 | case IORING_REGISTER_ENABLE_RINGS: |
9564 | ret = -EINVAL; | |
9565 | if (arg || nr_args) | |
9566 | break; | |
9567 | ret = io_register_enable_rings(ctx); | |
9568 | break; | |
21b55dbc SG |
9569 | case IORING_REGISTER_RESTRICTIONS: |
9570 | ret = io_register_restrictions(ctx, arg, nr_args); | |
9571 | break; | |
edafccee JA |
9572 | default: |
9573 | ret = -EINVAL; | |
9574 | break; | |
9575 | } | |
9576 | ||
21b55dbc | 9577 | out: |
071698e1 | 9578 | if (io_register_op_must_quiesce(opcode)) { |
05f3fb3c | 9579 | /* bring the ctx back to life */ |
05f3fb3c | 9580 | percpu_ref_reinit(&ctx->refs); |
21b55dbc | 9581 | out_quiesce: |
0f158b4c | 9582 | reinit_completion(&ctx->ref_comp); |
05f3fb3c | 9583 | } |
edafccee JA |
9584 | return ret; |
9585 | } | |
9586 | ||
9587 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
9588 | void __user *, arg, unsigned int, nr_args) | |
9589 | { | |
9590 | struct io_ring_ctx *ctx; | |
9591 | long ret = -EBADF; | |
9592 | struct fd f; | |
9593 | ||
9594 | f = fdget(fd); | |
9595 | if (!f.file) | |
9596 | return -EBADF; | |
9597 | ||
9598 | ret = -EOPNOTSUPP; | |
9599 | if (f.file->f_op != &io_uring_fops) | |
9600 | goto out_fput; | |
9601 | ||
9602 | ctx = f.file->private_data; | |
9603 | ||
9604 | mutex_lock(&ctx->uring_lock); | |
9605 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
9606 | mutex_unlock(&ctx->uring_lock); | |
c826bd7a DD |
9607 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, |
9608 | ctx->cq_ev_fd != NULL, ret); | |
edafccee JA |
9609 | out_fput: |
9610 | fdput(f); | |
9611 | return ret; | |
9612 | } | |
9613 | ||
2b188cc1 JA |
9614 | static int __init io_uring_init(void) |
9615 | { | |
d7f62e82 SM |
9616 | #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \ |
9617 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ | |
9618 | BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \ | |
9619 | } while (0) | |
9620 | ||
9621 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
9622 | __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename) | |
9623 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); | |
9624 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
9625 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
9626 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
9627 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
9628 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
9629 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
9630 | BUILD_BUG_SQE_ELEM(16, __u64, addr); | |
7d67af2c | 9631 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
9632 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
9633 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
9634 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
9635 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
9636 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
5769a351 JX |
9637 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); |
9638 | BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); | |
d7f62e82 SM |
9639 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); |
9640 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
9641 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
9642 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
9643 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
9644 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
9645 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
9646 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 9647 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
d7f62e82 SM |
9648 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
9649 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
9650 | BUILD_BUG_SQE_ELEM(42, __u16, personality); | |
7d67af2c | 9651 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
d7f62e82 | 9652 | |
d3656344 | 9653 | BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST); |
84557871 | 9654 | BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int)); |
2b188cc1 JA |
9655 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); |
9656 | return 0; | |
9657 | }; | |
9658 | __initcall(io_uring_init); |