]>
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> |
2b188cc1 | 82 | |
c826bd7a DD |
83 | #define CREATE_TRACE_POINTS |
84 | #include <trace/events/io_uring.h> | |
85 | ||
2b188cc1 JA |
86 | #include <uapi/linux/io_uring.h> |
87 | ||
88 | #include "internal.h" | |
561fb04a | 89 | #include "io-wq.h" |
2b188cc1 | 90 | |
5277deaa | 91 | #define IORING_MAX_ENTRIES 32768 |
33a107f0 | 92 | #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) |
65e19f54 JA |
93 | |
94 | /* | |
95 | * Shift of 9 is 512 entries, or exactly one page on 64-bit archs | |
96 | */ | |
97 | #define IORING_FILE_TABLE_SHIFT 9 | |
98 | #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT) | |
99 | #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1) | |
100 | #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE) | |
2b188cc1 JA |
101 | |
102 | struct io_uring { | |
103 | u32 head ____cacheline_aligned_in_smp; | |
104 | u32 tail ____cacheline_aligned_in_smp; | |
105 | }; | |
106 | ||
1e84b97b | 107 | /* |
75b28aff HV |
108 | * This data is shared with the application through the mmap at offsets |
109 | * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING. | |
1e84b97b SB |
110 | * |
111 | * The offsets to the member fields are published through struct | |
112 | * io_sqring_offsets when calling io_uring_setup. | |
113 | */ | |
75b28aff | 114 | struct io_rings { |
1e84b97b SB |
115 | /* |
116 | * Head and tail offsets into the ring; the offsets need to be | |
117 | * masked to get valid indices. | |
118 | * | |
75b28aff HV |
119 | * The kernel controls head of the sq ring and the tail of the cq ring, |
120 | * and the application controls tail of the sq ring and the head of the | |
121 | * cq ring. | |
1e84b97b | 122 | */ |
75b28aff | 123 | struct io_uring sq, cq; |
1e84b97b | 124 | /* |
75b28aff | 125 | * Bitmasks to apply to head and tail offsets (constant, equals |
1e84b97b SB |
126 | * ring_entries - 1) |
127 | */ | |
75b28aff HV |
128 | u32 sq_ring_mask, cq_ring_mask; |
129 | /* Ring sizes (constant, power of 2) */ | |
130 | u32 sq_ring_entries, cq_ring_entries; | |
1e84b97b SB |
131 | /* |
132 | * Number of invalid entries dropped by the kernel due to | |
133 | * invalid index stored in array | |
134 | * | |
135 | * Written by the kernel, shouldn't be modified by the | |
136 | * application (i.e. get number of "new events" by comparing to | |
137 | * cached value). | |
138 | * | |
139 | * After a new SQ head value was read by the application this | |
140 | * counter includes all submissions that were dropped reaching | |
141 | * the new SQ head (and possibly more). | |
142 | */ | |
75b28aff | 143 | u32 sq_dropped; |
1e84b97b | 144 | /* |
0d9b5b3a | 145 | * Runtime SQ flags |
1e84b97b SB |
146 | * |
147 | * Written by the kernel, shouldn't be modified by the | |
148 | * application. | |
149 | * | |
150 | * The application needs a full memory barrier before checking | |
151 | * for IORING_SQ_NEED_WAKEUP after updating the sq tail. | |
152 | */ | |
75b28aff | 153 | u32 sq_flags; |
0d9b5b3a SG |
154 | /* |
155 | * Runtime CQ flags | |
156 | * | |
157 | * Written by the application, shouldn't be modified by the | |
158 | * kernel. | |
159 | */ | |
160 | u32 cq_flags; | |
1e84b97b SB |
161 | /* |
162 | * Number of completion events lost because the queue was full; | |
163 | * this should be avoided by the application by making sure | |
0b4295b5 | 164 | * there are not more requests pending than there is space in |
1e84b97b SB |
165 | * the completion queue. |
166 | * | |
167 | * Written by the kernel, shouldn't be modified by the | |
168 | * application (i.e. get number of "new events" by comparing to | |
169 | * cached value). | |
170 | * | |
171 | * As completion events come in out of order this counter is not | |
172 | * ordered with any other data. | |
173 | */ | |
75b28aff | 174 | u32 cq_overflow; |
1e84b97b SB |
175 | /* |
176 | * Ring buffer of completion events. | |
177 | * | |
178 | * The kernel writes completion events fresh every time they are | |
179 | * produced, so the application is allowed to modify pending | |
180 | * entries. | |
181 | */ | |
75b28aff | 182 | struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp; |
2b188cc1 JA |
183 | }; |
184 | ||
edafccee JA |
185 | struct io_mapped_ubuf { |
186 | u64 ubuf; | |
187 | size_t len; | |
188 | struct bio_vec *bvec; | |
189 | unsigned int nr_bvecs; | |
190 | }; | |
191 | ||
65e19f54 JA |
192 | struct fixed_file_table { |
193 | struct file **files; | |
31b51510 JA |
194 | }; |
195 | ||
05589553 XW |
196 | struct fixed_file_ref_node { |
197 | struct percpu_ref refs; | |
198 | struct list_head node; | |
199 | struct list_head file_list; | |
200 | struct fixed_file_data *file_data; | |
4a38aed2 | 201 | struct llist_node llist; |
05589553 XW |
202 | }; |
203 | ||
05f3fb3c JA |
204 | struct fixed_file_data { |
205 | struct fixed_file_table *table; | |
206 | struct io_ring_ctx *ctx; | |
207 | ||
05589553 | 208 | struct percpu_ref *cur_refs; |
05f3fb3c | 209 | struct percpu_ref refs; |
05f3fb3c | 210 | struct completion done; |
05589553 XW |
211 | struct list_head ref_list; |
212 | spinlock_t lock; | |
05f3fb3c JA |
213 | }; |
214 | ||
5a2e745d JA |
215 | struct io_buffer { |
216 | struct list_head list; | |
217 | __u64 addr; | |
218 | __s32 len; | |
219 | __u16 bid; | |
220 | }; | |
221 | ||
2b188cc1 JA |
222 | struct io_ring_ctx { |
223 | struct { | |
224 | struct percpu_ref refs; | |
225 | } ____cacheline_aligned_in_smp; | |
226 | ||
227 | struct { | |
228 | unsigned int flags; | |
e1d85334 | 229 | unsigned int compat: 1; |
aad5d8da | 230 | unsigned int limit_mem: 1; |
e1d85334 RD |
231 | unsigned int cq_overflow_flushed: 1; |
232 | unsigned int drain_next: 1; | |
233 | unsigned int eventfd_async: 1; | |
2b188cc1 | 234 | |
75b28aff HV |
235 | /* |
236 | * Ring buffer of indices into array of io_uring_sqe, which is | |
237 | * mmapped by the application using the IORING_OFF_SQES offset. | |
238 | * | |
239 | * This indirection could e.g. be used to assign fixed | |
240 | * io_uring_sqe entries to operations and only submit them to | |
241 | * the queue when needed. | |
242 | * | |
243 | * The kernel modifies neither the indices array nor the entries | |
244 | * array. | |
245 | */ | |
246 | u32 *sq_array; | |
2b188cc1 JA |
247 | unsigned cached_sq_head; |
248 | unsigned sq_entries; | |
249 | unsigned sq_mask; | |
6c271ce2 | 250 | unsigned sq_thread_idle; |
498ccd9e | 251 | unsigned cached_sq_dropped; |
206aefde | 252 | atomic_t cached_cq_overflow; |
ad3eb2c8 | 253 | unsigned long sq_check_overflow; |
de0617e4 JA |
254 | |
255 | struct list_head defer_list; | |
5262f567 | 256 | struct list_head timeout_list; |
1d7bb1d5 | 257 | struct list_head cq_overflow_list; |
fcb323cc JA |
258 | |
259 | wait_queue_head_t inflight_wait; | |
ad3eb2c8 | 260 | struct io_uring_sqe *sq_sqes; |
2b188cc1 JA |
261 | } ____cacheline_aligned_in_smp; |
262 | ||
206aefde JA |
263 | struct io_rings *rings; |
264 | ||
2b188cc1 | 265 | /* IO offload */ |
561fb04a | 266 | struct io_wq *io_wq; |
6c271ce2 | 267 | struct task_struct *sqo_thread; /* if using sq thread polling */ |
2aede0e4 JA |
268 | |
269 | /* | |
270 | * For SQPOLL usage - we hold a reference to the parent task, so we | |
271 | * have access to the ->files | |
272 | */ | |
273 | struct task_struct *sqo_task; | |
274 | ||
275 | /* Only used for accounting purposes */ | |
276 | struct mm_struct *mm_account; | |
277 | ||
6c271ce2 | 278 | wait_queue_head_t sqo_wait; |
75b28aff | 279 | |
6b06314c JA |
280 | /* |
281 | * If used, fixed file set. Writers must ensure that ->refs is dead, | |
282 | * readers must ensure that ->refs is alive as long as the file* is | |
283 | * used. Only updated through io_uring_register(2). | |
284 | */ | |
05f3fb3c | 285 | struct fixed_file_data *file_data; |
6b06314c | 286 | unsigned nr_user_files; |
b14cca0c PB |
287 | int ring_fd; |
288 | struct file *ring_file; | |
6b06314c | 289 | |
edafccee JA |
290 | /* if used, fixed mapped user buffers */ |
291 | unsigned nr_user_bufs; | |
292 | struct io_mapped_ubuf *user_bufs; | |
293 | ||
2b188cc1 JA |
294 | struct user_struct *user; |
295 | ||
0b8c0ec7 | 296 | const struct cred *creds; |
181e448d | 297 | |
0f158b4c JA |
298 | struct completion ref_comp; |
299 | struct completion sq_thread_comp; | |
206aefde | 300 | |
0ddf92e8 JA |
301 | /* if all else fails... */ |
302 | struct io_kiocb *fallback_req; | |
303 | ||
206aefde JA |
304 | #if defined(CONFIG_UNIX) |
305 | struct socket *ring_sock; | |
306 | #endif | |
307 | ||
5a2e745d JA |
308 | struct idr io_buffer_idr; |
309 | ||
071698e1 JA |
310 | struct idr personality_idr; |
311 | ||
206aefde JA |
312 | struct { |
313 | unsigned cached_cq_tail; | |
314 | unsigned cq_entries; | |
315 | unsigned cq_mask; | |
316 | atomic_t cq_timeouts; | |
ad3eb2c8 | 317 | unsigned long cq_check_overflow; |
206aefde JA |
318 | struct wait_queue_head cq_wait; |
319 | struct fasync_struct *cq_fasync; | |
320 | struct eventfd_ctx *cq_ev_fd; | |
321 | } ____cacheline_aligned_in_smp; | |
2b188cc1 JA |
322 | |
323 | struct { | |
324 | struct mutex uring_lock; | |
325 | wait_queue_head_t wait; | |
326 | } ____cacheline_aligned_in_smp; | |
327 | ||
328 | struct { | |
329 | spinlock_t completion_lock; | |
e94f141b | 330 | |
def596e9 | 331 | /* |
540e32a0 | 332 | * ->iopoll_list is protected by the ctx->uring_lock for |
def596e9 JA |
333 | * io_uring instances that don't use IORING_SETUP_SQPOLL. |
334 | * For SQPOLL, only the single threaded io_sq_thread() will | |
335 | * manipulate the list, hence no extra locking is needed there. | |
336 | */ | |
540e32a0 | 337 | struct list_head iopoll_list; |
78076bb6 JA |
338 | struct hlist_head *cancel_hash; |
339 | unsigned cancel_hash_bits; | |
e94f141b | 340 | bool poll_multi_file; |
31b51510 | 341 | |
fcb323cc JA |
342 | spinlock_t inflight_lock; |
343 | struct list_head inflight_list; | |
2b188cc1 | 344 | } ____cacheline_aligned_in_smp; |
85faa7b8 | 345 | |
4a38aed2 JA |
346 | struct delayed_work file_put_work; |
347 | struct llist_head file_put_llist; | |
348 | ||
85faa7b8 | 349 | struct work_struct exit_work; |
2b188cc1 JA |
350 | }; |
351 | ||
09bb8394 JA |
352 | /* |
353 | * First field must be the file pointer in all the | |
354 | * iocb unions! See also 'struct kiocb' in <linux/fs.h> | |
355 | */ | |
221c5eb2 JA |
356 | struct io_poll_iocb { |
357 | struct file *file; | |
0969e783 JA |
358 | union { |
359 | struct wait_queue_head *head; | |
360 | u64 addr; | |
361 | }; | |
221c5eb2 | 362 | __poll_t events; |
8c838788 | 363 | bool done; |
221c5eb2 | 364 | bool canceled; |
392edb45 | 365 | struct wait_queue_entry wait; |
221c5eb2 JA |
366 | }; |
367 | ||
b5dba59e JA |
368 | struct io_close { |
369 | struct file *file; | |
370 | struct file *put_file; | |
371 | int fd; | |
372 | }; | |
373 | ||
ad8a48ac JA |
374 | struct io_timeout_data { |
375 | struct io_kiocb *req; | |
376 | struct hrtimer timer; | |
377 | struct timespec64 ts; | |
378 | enum hrtimer_mode mode; | |
379 | }; | |
380 | ||
8ed8d3c3 JA |
381 | struct io_accept { |
382 | struct file *file; | |
383 | struct sockaddr __user *addr; | |
384 | int __user *addr_len; | |
385 | int flags; | |
09952e3e | 386 | unsigned long nofile; |
8ed8d3c3 JA |
387 | }; |
388 | ||
389 | struct io_sync { | |
390 | struct file *file; | |
391 | loff_t len; | |
392 | loff_t off; | |
393 | int flags; | |
d63d1b5e | 394 | int mode; |
8ed8d3c3 JA |
395 | }; |
396 | ||
fbf23849 JA |
397 | struct io_cancel { |
398 | struct file *file; | |
399 | u64 addr; | |
400 | }; | |
401 | ||
b29472ee JA |
402 | struct io_timeout { |
403 | struct file *file; | |
404 | u64 addr; | |
405 | int flags; | |
bfe68a22 PB |
406 | u32 off; |
407 | u32 target_seq; | |
135fcde8 | 408 | struct list_head list; |
b29472ee JA |
409 | }; |
410 | ||
9adbd45d JA |
411 | struct io_rw { |
412 | /* NOTE: kiocb has the file as the first member, so don't do it here */ | |
413 | struct kiocb kiocb; | |
414 | u64 addr; | |
415 | u64 len; | |
416 | }; | |
417 | ||
3fbb51c1 JA |
418 | struct io_connect { |
419 | struct file *file; | |
420 | struct sockaddr __user *addr; | |
421 | int addr_len; | |
422 | }; | |
423 | ||
e47293fd JA |
424 | struct io_sr_msg { |
425 | struct file *file; | |
fddaface | 426 | union { |
270a5940 | 427 | struct user_msghdr __user *umsg; |
fddaface JA |
428 | void __user *buf; |
429 | }; | |
e47293fd | 430 | int msg_flags; |
bcda7baa | 431 | int bgid; |
fddaface | 432 | size_t len; |
bcda7baa | 433 | struct io_buffer *kbuf; |
e47293fd JA |
434 | }; |
435 | ||
15b71abe JA |
436 | struct io_open { |
437 | struct file *file; | |
438 | int dfd; | |
15b71abe | 439 | struct filename *filename; |
c12cedf2 | 440 | struct open_how how; |
4022e7af | 441 | unsigned long nofile; |
15b71abe JA |
442 | }; |
443 | ||
05f3fb3c JA |
444 | struct io_files_update { |
445 | struct file *file; | |
446 | u64 arg; | |
447 | u32 nr_args; | |
448 | u32 offset; | |
449 | }; | |
450 | ||
4840e418 JA |
451 | struct io_fadvise { |
452 | struct file *file; | |
453 | u64 offset; | |
454 | u32 len; | |
455 | u32 advice; | |
456 | }; | |
457 | ||
c1ca757b JA |
458 | struct io_madvise { |
459 | struct file *file; | |
460 | u64 addr; | |
461 | u32 len; | |
462 | u32 advice; | |
463 | }; | |
464 | ||
3e4827b0 JA |
465 | struct io_epoll { |
466 | struct file *file; | |
467 | int epfd; | |
468 | int op; | |
469 | int fd; | |
470 | struct epoll_event event; | |
e47293fd JA |
471 | }; |
472 | ||
7d67af2c PB |
473 | struct io_splice { |
474 | struct file *file_out; | |
475 | struct file *file_in; | |
476 | loff_t off_out; | |
477 | loff_t off_in; | |
478 | u64 len; | |
479 | unsigned int flags; | |
480 | }; | |
481 | ||
ddf0322d JA |
482 | struct io_provide_buf { |
483 | struct file *file; | |
484 | __u64 addr; | |
485 | __s32 len; | |
486 | __u32 bgid; | |
487 | __u16 nbufs; | |
488 | __u16 bid; | |
489 | }; | |
490 | ||
1d9e1288 BM |
491 | struct io_statx { |
492 | struct file *file; | |
493 | int dfd; | |
494 | unsigned int mask; | |
495 | unsigned int flags; | |
e62753e4 | 496 | const char __user *filename; |
1d9e1288 BM |
497 | struct statx __user *buffer; |
498 | }; | |
499 | ||
3ca405eb PB |
500 | struct io_completion { |
501 | struct file *file; | |
502 | struct list_head list; | |
0f7e466b | 503 | int cflags; |
3ca405eb PB |
504 | }; |
505 | ||
f499a021 JA |
506 | struct io_async_connect { |
507 | struct sockaddr_storage address; | |
508 | }; | |
509 | ||
03b1230c JA |
510 | struct io_async_msghdr { |
511 | struct iovec fast_iov[UIO_FASTIOV]; | |
512 | struct iovec *iov; | |
513 | struct sockaddr __user *uaddr; | |
514 | struct msghdr msg; | |
b537916c | 515 | struct sockaddr_storage addr; |
03b1230c JA |
516 | }; |
517 | ||
f67676d1 JA |
518 | struct io_async_rw { |
519 | struct iovec fast_iov[UIO_FASTIOV]; | |
ff6165b2 JA |
520 | const struct iovec *free_iovec; |
521 | struct iov_iter iter; | |
227c0c96 | 522 | size_t bytes_done; |
bcf5a063 | 523 | struct wait_page_queue wpq; |
f67676d1 JA |
524 | }; |
525 | ||
1a6b74fc | 526 | struct io_async_ctx { |
f67676d1 JA |
527 | union { |
528 | struct io_async_rw rw; | |
03b1230c | 529 | struct io_async_msghdr msg; |
f499a021 | 530 | struct io_async_connect connect; |
2d28390a | 531 | struct io_timeout_data timeout; |
f67676d1 | 532 | }; |
1a6b74fc JA |
533 | }; |
534 | ||
6b47ee6e PB |
535 | enum { |
536 | REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT, | |
537 | REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT, | |
538 | REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT, | |
539 | REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT, | |
540 | REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT, | |
bcda7baa | 541 | REQ_F_BUFFER_SELECT_BIT = IOSQE_BUFFER_SELECT_BIT, |
6b47ee6e | 542 | |
dea3b49c | 543 | REQ_F_LINK_HEAD_BIT, |
6b47ee6e PB |
544 | REQ_F_FAIL_LINK_BIT, |
545 | REQ_F_INFLIGHT_BIT, | |
546 | REQ_F_CUR_POS_BIT, | |
547 | REQ_F_NOWAIT_BIT, | |
6b47ee6e | 548 | REQ_F_LINK_TIMEOUT_BIT, |
6b47ee6e | 549 | REQ_F_ISREG_BIT, |
6b47ee6e | 550 | REQ_F_COMP_LOCKED_BIT, |
99bc4c38 | 551 | REQ_F_NEED_CLEANUP_BIT, |
d7718a9d | 552 | REQ_F_POLLED_BIT, |
bcda7baa | 553 | REQ_F_BUFFER_SELECTED_BIT, |
5b0bbee4 | 554 | REQ_F_NO_FILE_TABLE_BIT, |
7cdaf587 | 555 | REQ_F_WORK_INITIALIZED_BIT, |
4dd2824d | 556 | REQ_F_TASK_PINNED_BIT, |
84557871 JA |
557 | |
558 | /* not a real bit, just to check we're not overflowing the space */ | |
559 | __REQ_F_LAST_BIT, | |
6b47ee6e PB |
560 | }; |
561 | ||
562 | enum { | |
563 | /* ctx owns file */ | |
564 | REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT), | |
565 | /* drain existing IO first */ | |
566 | REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT), | |
567 | /* linked sqes */ | |
568 | REQ_F_LINK = BIT(REQ_F_LINK_BIT), | |
569 | /* doesn't sever on completion < 0 */ | |
570 | REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT), | |
571 | /* IOSQE_ASYNC */ | |
572 | REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT), | |
bcda7baa JA |
573 | /* IOSQE_BUFFER_SELECT */ |
574 | REQ_F_BUFFER_SELECT = BIT(REQ_F_BUFFER_SELECT_BIT), | |
6b47ee6e | 575 | |
dea3b49c PB |
576 | /* head of a link */ |
577 | REQ_F_LINK_HEAD = BIT(REQ_F_LINK_HEAD_BIT), | |
6b47ee6e PB |
578 | /* fail rest of links */ |
579 | REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT), | |
580 | /* on inflight list */ | |
581 | REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT), | |
582 | /* read/write uses file position */ | |
583 | REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT), | |
584 | /* must not punt to workers */ | |
585 | REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT), | |
6b47ee6e PB |
586 | /* has linked timeout */ |
587 | REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT), | |
6b47ee6e PB |
588 | /* regular file */ |
589 | REQ_F_ISREG = BIT(REQ_F_ISREG_BIT), | |
6b47ee6e PB |
590 | /* completion under lock */ |
591 | REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT), | |
99bc4c38 PB |
592 | /* needs cleanup */ |
593 | REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT), | |
d7718a9d JA |
594 | /* already went through poll handler */ |
595 | REQ_F_POLLED = BIT(REQ_F_POLLED_BIT), | |
bcda7baa JA |
596 | /* buffer already selected */ |
597 | REQ_F_BUFFER_SELECTED = BIT(REQ_F_BUFFER_SELECTED_BIT), | |
5b0bbee4 JA |
598 | /* doesn't need file table for this request */ |
599 | REQ_F_NO_FILE_TABLE = BIT(REQ_F_NO_FILE_TABLE_BIT), | |
7cdaf587 XW |
600 | /* io_wq_work is initialized */ |
601 | REQ_F_WORK_INITIALIZED = BIT(REQ_F_WORK_INITIALIZED_BIT), | |
4dd2824d PB |
602 | /* req->task is refcounted */ |
603 | REQ_F_TASK_PINNED = BIT(REQ_F_TASK_PINNED_BIT), | |
d7718a9d JA |
604 | }; |
605 | ||
606 | struct async_poll { | |
607 | struct io_poll_iocb poll; | |
807abcb0 | 608 | struct io_poll_iocb *double_poll; |
6b47ee6e PB |
609 | }; |
610 | ||
09bb8394 JA |
611 | /* |
612 | * NOTE! Each of the iocb union members has the file pointer | |
613 | * as the first entry in their struct definition. So you can | |
614 | * access the file pointer through any of the sub-structs, | |
615 | * or directly as just 'ki_filp' in this struct. | |
616 | */ | |
2b188cc1 | 617 | struct io_kiocb { |
221c5eb2 | 618 | union { |
09bb8394 | 619 | struct file *file; |
9adbd45d | 620 | struct io_rw rw; |
221c5eb2 | 621 | struct io_poll_iocb poll; |
8ed8d3c3 JA |
622 | struct io_accept accept; |
623 | struct io_sync sync; | |
fbf23849 | 624 | struct io_cancel cancel; |
b29472ee | 625 | struct io_timeout timeout; |
3fbb51c1 | 626 | struct io_connect connect; |
e47293fd | 627 | struct io_sr_msg sr_msg; |
15b71abe | 628 | struct io_open open; |
b5dba59e | 629 | struct io_close close; |
05f3fb3c | 630 | struct io_files_update files_update; |
4840e418 | 631 | struct io_fadvise fadvise; |
c1ca757b | 632 | struct io_madvise madvise; |
3e4827b0 | 633 | struct io_epoll epoll; |
7d67af2c | 634 | struct io_splice splice; |
ddf0322d | 635 | struct io_provide_buf pbuf; |
1d9e1288 | 636 | struct io_statx statx; |
3ca405eb PB |
637 | /* use only after cleaning per-op data, see io_clean_op() */ |
638 | struct io_completion compl; | |
221c5eb2 | 639 | }; |
2b188cc1 | 640 | |
1a6b74fc | 641 | struct io_async_ctx *io; |
d625c6ee | 642 | u8 opcode; |
65a6543d XW |
643 | /* polled IO has completed */ |
644 | u8 iopoll_completed; | |
2b188cc1 | 645 | |
4f4eeba8 | 646 | u16 buf_index; |
9cf7c104 | 647 | u32 result; |
4f4eeba8 | 648 | |
010e8e6b PB |
649 | struct io_ring_ctx *ctx; |
650 | unsigned int flags; | |
651 | refcount_t refs; | |
652 | struct task_struct *task; | |
653 | u64 user_data; | |
d7718a9d | 654 | |
010e8e6b | 655 | struct list_head link_list; |
fcb323cc | 656 | |
d21ffe7e PB |
657 | /* |
658 | * 1. used with ctx->iopoll_list with reads/writes | |
659 | * 2. to track reqs with ->files (see io_op_def::file_table) | |
660 | */ | |
010e8e6b PB |
661 | struct list_head inflight_entry; |
662 | ||
663 | struct percpu_ref *fixed_file_refs; | |
664 | struct callback_head task_work; | |
665 | /* for polled requests, i.e. IORING_OP_POLL_ADD and async armed poll */ | |
666 | struct hlist_node hash_node; | |
667 | struct async_poll *apoll; | |
668 | struct io_wq_work work; | |
2b188cc1 | 669 | }; |
05589553 | 670 | |
27dc8338 PB |
671 | struct io_defer_entry { |
672 | struct list_head list; | |
673 | struct io_kiocb *req; | |
9cf7c104 | 674 | u32 seq; |
2b188cc1 JA |
675 | }; |
676 | ||
def596e9 | 677 | #define IO_IOPOLL_BATCH 8 |
2b188cc1 | 678 | |
013538bd JA |
679 | struct io_comp_state { |
680 | unsigned int nr; | |
681 | struct list_head list; | |
682 | struct io_ring_ctx *ctx; | |
683 | }; | |
684 | ||
9a56a232 JA |
685 | struct io_submit_state { |
686 | struct blk_plug plug; | |
687 | ||
2579f913 JA |
688 | /* |
689 | * io_kiocb alloc cache | |
690 | */ | |
691 | void *reqs[IO_IOPOLL_BATCH]; | |
6c8a3134 | 692 | unsigned int free_reqs; |
2579f913 | 693 | |
013538bd JA |
694 | /* |
695 | * Batch completion logic | |
696 | */ | |
697 | struct io_comp_state comp; | |
698 | ||
9a56a232 JA |
699 | /* |
700 | * File reference cache | |
701 | */ | |
702 | struct file *file; | |
703 | unsigned int fd; | |
704 | unsigned int has_refs; | |
9a56a232 JA |
705 | unsigned int ios_left; |
706 | }; | |
707 | ||
d3656344 JA |
708 | struct io_op_def { |
709 | /* needs req->io allocated for deferral/async */ | |
710 | unsigned async_ctx : 1; | |
711 | /* needs current->mm setup, does mm access */ | |
712 | unsigned needs_mm : 1; | |
713 | /* needs req->file assigned */ | |
714 | unsigned needs_file : 1; | |
fd2206e4 JA |
715 | /* don't fail if file grab fails */ |
716 | unsigned needs_file_no_error : 1; | |
d3656344 JA |
717 | /* hash wq insertion if file is a regular file */ |
718 | unsigned hash_reg_file : 1; | |
719 | /* unbound wq insertion if file is a non-regular file */ | |
720 | unsigned unbound_nonreg_file : 1; | |
66f4af93 JA |
721 | /* opcode is not supported by this kernel */ |
722 | unsigned not_supported : 1; | |
f86cd20c JA |
723 | /* needs file table */ |
724 | unsigned file_table : 1; | |
ff002b30 JA |
725 | /* needs ->fs */ |
726 | unsigned needs_fs : 1; | |
8a72758c JA |
727 | /* set if opcode supports polled "wait" */ |
728 | unsigned pollin : 1; | |
729 | unsigned pollout : 1; | |
bcda7baa JA |
730 | /* op supports buffer selection */ |
731 | unsigned buffer_select : 1; | |
57f1a649 | 732 | unsigned needs_fsize : 1; |
d3656344 JA |
733 | }; |
734 | ||
735 | static const struct io_op_def io_op_defs[] = { | |
0463b6c5 PB |
736 | [IORING_OP_NOP] = {}, |
737 | [IORING_OP_READV] = { | |
d3656344 JA |
738 | .async_ctx = 1, |
739 | .needs_mm = 1, | |
740 | .needs_file = 1, | |
741 | .unbound_nonreg_file = 1, | |
8a72758c | 742 | .pollin = 1, |
4d954c25 | 743 | .buffer_select = 1, |
d3656344 | 744 | }, |
0463b6c5 | 745 | [IORING_OP_WRITEV] = { |
d3656344 JA |
746 | .async_ctx = 1, |
747 | .needs_mm = 1, | |
748 | .needs_file = 1, | |
749 | .hash_reg_file = 1, | |
750 | .unbound_nonreg_file = 1, | |
8a72758c | 751 | .pollout = 1, |
57f1a649 | 752 | .needs_fsize = 1, |
d3656344 | 753 | }, |
0463b6c5 | 754 | [IORING_OP_FSYNC] = { |
d3656344 JA |
755 | .needs_file = 1, |
756 | }, | |
0463b6c5 | 757 | [IORING_OP_READ_FIXED] = { |
d3656344 JA |
758 | .needs_file = 1, |
759 | .unbound_nonreg_file = 1, | |
8a72758c | 760 | .pollin = 1, |
d3656344 | 761 | }, |
0463b6c5 | 762 | [IORING_OP_WRITE_FIXED] = { |
d3656344 JA |
763 | .needs_file = 1, |
764 | .hash_reg_file = 1, | |
765 | .unbound_nonreg_file = 1, | |
8a72758c | 766 | .pollout = 1, |
57f1a649 | 767 | .needs_fsize = 1, |
d3656344 | 768 | }, |
0463b6c5 | 769 | [IORING_OP_POLL_ADD] = { |
d3656344 JA |
770 | .needs_file = 1, |
771 | .unbound_nonreg_file = 1, | |
772 | }, | |
0463b6c5 PB |
773 | [IORING_OP_POLL_REMOVE] = {}, |
774 | [IORING_OP_SYNC_FILE_RANGE] = { | |
d3656344 JA |
775 | .needs_file = 1, |
776 | }, | |
0463b6c5 | 777 | [IORING_OP_SENDMSG] = { |
d3656344 JA |
778 | .async_ctx = 1, |
779 | .needs_mm = 1, | |
780 | .needs_file = 1, | |
781 | .unbound_nonreg_file = 1, | |
ff002b30 | 782 | .needs_fs = 1, |
8a72758c | 783 | .pollout = 1, |
d3656344 | 784 | }, |
0463b6c5 | 785 | [IORING_OP_RECVMSG] = { |
d3656344 JA |
786 | .async_ctx = 1, |
787 | .needs_mm = 1, | |
788 | .needs_file = 1, | |
789 | .unbound_nonreg_file = 1, | |
ff002b30 | 790 | .needs_fs = 1, |
8a72758c | 791 | .pollin = 1, |
52de1fe1 | 792 | .buffer_select = 1, |
d3656344 | 793 | }, |
0463b6c5 | 794 | [IORING_OP_TIMEOUT] = { |
d3656344 JA |
795 | .async_ctx = 1, |
796 | .needs_mm = 1, | |
797 | }, | |
0463b6c5 PB |
798 | [IORING_OP_TIMEOUT_REMOVE] = {}, |
799 | [IORING_OP_ACCEPT] = { | |
d3656344 JA |
800 | .needs_mm = 1, |
801 | .needs_file = 1, | |
802 | .unbound_nonreg_file = 1, | |
f86cd20c | 803 | .file_table = 1, |
8a72758c | 804 | .pollin = 1, |
d3656344 | 805 | }, |
0463b6c5 PB |
806 | [IORING_OP_ASYNC_CANCEL] = {}, |
807 | [IORING_OP_LINK_TIMEOUT] = { | |
d3656344 JA |
808 | .async_ctx = 1, |
809 | .needs_mm = 1, | |
810 | }, | |
0463b6c5 | 811 | [IORING_OP_CONNECT] = { |
d3656344 JA |
812 | .async_ctx = 1, |
813 | .needs_mm = 1, | |
814 | .needs_file = 1, | |
815 | .unbound_nonreg_file = 1, | |
8a72758c | 816 | .pollout = 1, |
d3656344 | 817 | }, |
0463b6c5 | 818 | [IORING_OP_FALLOCATE] = { |
d3656344 | 819 | .needs_file = 1, |
57f1a649 | 820 | .needs_fsize = 1, |
d3656344 | 821 | }, |
0463b6c5 | 822 | [IORING_OP_OPENAT] = { |
f86cd20c | 823 | .file_table = 1, |
ff002b30 | 824 | .needs_fs = 1, |
d3656344 | 825 | }, |
0463b6c5 | 826 | [IORING_OP_CLOSE] = { |
fd2206e4 JA |
827 | .needs_file = 1, |
828 | .needs_file_no_error = 1, | |
f86cd20c | 829 | .file_table = 1, |
d3656344 | 830 | }, |
0463b6c5 | 831 | [IORING_OP_FILES_UPDATE] = { |
d3656344 | 832 | .needs_mm = 1, |
f86cd20c | 833 | .file_table = 1, |
d3656344 | 834 | }, |
0463b6c5 | 835 | [IORING_OP_STATX] = { |
d3656344 | 836 | .needs_mm = 1, |
ff002b30 | 837 | .needs_fs = 1, |
5b0bbee4 | 838 | .file_table = 1, |
d3656344 | 839 | }, |
0463b6c5 | 840 | [IORING_OP_READ] = { |
3a6820f2 JA |
841 | .needs_mm = 1, |
842 | .needs_file = 1, | |
843 | .unbound_nonreg_file = 1, | |
8a72758c | 844 | .pollin = 1, |
bcda7baa | 845 | .buffer_select = 1, |
3a6820f2 | 846 | }, |
0463b6c5 | 847 | [IORING_OP_WRITE] = { |
3a6820f2 JA |
848 | .needs_mm = 1, |
849 | .needs_file = 1, | |
850 | .unbound_nonreg_file = 1, | |
8a72758c | 851 | .pollout = 1, |
57f1a649 | 852 | .needs_fsize = 1, |
3a6820f2 | 853 | }, |
0463b6c5 | 854 | [IORING_OP_FADVISE] = { |
4840e418 JA |
855 | .needs_file = 1, |
856 | }, | |
0463b6c5 | 857 | [IORING_OP_MADVISE] = { |
c1ca757b JA |
858 | .needs_mm = 1, |
859 | }, | |
0463b6c5 | 860 | [IORING_OP_SEND] = { |
fddaface JA |
861 | .needs_mm = 1, |
862 | .needs_file = 1, | |
863 | .unbound_nonreg_file = 1, | |
8a72758c | 864 | .pollout = 1, |
fddaface | 865 | }, |
0463b6c5 | 866 | [IORING_OP_RECV] = { |
fddaface JA |
867 | .needs_mm = 1, |
868 | .needs_file = 1, | |
869 | .unbound_nonreg_file = 1, | |
8a72758c | 870 | .pollin = 1, |
bcda7baa | 871 | .buffer_select = 1, |
fddaface | 872 | }, |
0463b6c5 | 873 | [IORING_OP_OPENAT2] = { |
f86cd20c | 874 | .file_table = 1, |
ff002b30 | 875 | .needs_fs = 1, |
cebdb986 | 876 | }, |
3e4827b0 JA |
877 | [IORING_OP_EPOLL_CTL] = { |
878 | .unbound_nonreg_file = 1, | |
879 | .file_table = 1, | |
880 | }, | |
7d67af2c PB |
881 | [IORING_OP_SPLICE] = { |
882 | .needs_file = 1, | |
883 | .hash_reg_file = 1, | |
884 | .unbound_nonreg_file = 1, | |
ddf0322d JA |
885 | }, |
886 | [IORING_OP_PROVIDE_BUFFERS] = {}, | |
067524e9 | 887 | [IORING_OP_REMOVE_BUFFERS] = {}, |
f2a8d5c7 PB |
888 | [IORING_OP_TEE] = { |
889 | .needs_file = 1, | |
890 | .hash_reg_file = 1, | |
891 | .unbound_nonreg_file = 1, | |
892 | }, | |
d3656344 JA |
893 | }; |
894 | ||
2e0464d4 BM |
895 | enum io_mem_account { |
896 | ACCT_LOCKED, | |
897 | ACCT_PINNED, | |
898 | }; | |
899 | ||
81b68a5c PB |
900 | static void __io_complete_rw(struct io_kiocb *req, long res, long res2, |
901 | struct io_comp_state *cs); | |
78e19bbe | 902 | static void io_cqring_fill_event(struct io_kiocb *req, long res); |
ec9c02ad | 903 | static void io_put_req(struct io_kiocb *req); |
c40f6379 | 904 | static void io_double_put_req(struct io_kiocb *req); |
978db57e | 905 | static void __io_double_put_req(struct io_kiocb *req); |
94ae5e77 | 906 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req); |
7271ef3a | 907 | static void __io_queue_linked_timeout(struct io_kiocb *req); |
94ae5e77 | 908 | static void io_queue_linked_timeout(struct io_kiocb *req); |
05f3fb3c JA |
909 | static int __io_sqe_files_update(struct io_ring_ctx *ctx, |
910 | struct io_uring_files_update *ip, | |
911 | unsigned nr_args); | |
f56040b8 | 912 | static int io_prep_work_files(struct io_kiocb *req); |
3ca405eb | 913 | static void __io_clean_op(struct io_kiocb *req); |
b41e9852 JA |
914 | static int io_file_get(struct io_submit_state *state, struct io_kiocb *req, |
915 | int fd, struct file **out_file, bool fixed); | |
916 | static void __io_queue_sqe(struct io_kiocb *req, | |
f13fad7b JA |
917 | const struct io_uring_sqe *sqe, |
918 | struct io_comp_state *cs); | |
4349f30e | 919 | static void io_file_put_work(struct work_struct *work); |
de0617e4 | 920 | |
b63534c4 JA |
921 | static ssize_t io_import_iovec(int rw, struct io_kiocb *req, |
922 | struct iovec **iovec, struct iov_iter *iter, | |
923 | bool needs_lock); | |
ff6165b2 JA |
924 | static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec, |
925 | const struct iovec *fast_iov, | |
227c0c96 | 926 | struct iov_iter *iter, bool force); |
de0617e4 | 927 | |
2b188cc1 JA |
928 | static struct kmem_cache *req_cachep; |
929 | ||
930 | static const struct file_operations io_uring_fops; | |
931 | ||
932 | struct sock *io_uring_get_socket(struct file *file) | |
933 | { | |
934 | #if defined(CONFIG_UNIX) | |
935 | if (file->f_op == &io_uring_fops) { | |
936 | struct io_ring_ctx *ctx = file->private_data; | |
937 | ||
938 | return ctx->ring_sock->sk; | |
939 | } | |
940 | #endif | |
941 | return NULL; | |
942 | } | |
943 | EXPORT_SYMBOL(io_uring_get_socket); | |
944 | ||
4dd2824d PB |
945 | static void io_get_req_task(struct io_kiocb *req) |
946 | { | |
947 | if (req->flags & REQ_F_TASK_PINNED) | |
948 | return; | |
949 | get_task_struct(req->task); | |
950 | req->flags |= REQ_F_TASK_PINNED; | |
951 | } | |
952 | ||
3ca405eb PB |
953 | static inline void io_clean_op(struct io_kiocb *req) |
954 | { | |
bb175342 PB |
955 | if (req->flags & (REQ_F_NEED_CLEANUP | REQ_F_BUFFER_SELECTED | |
956 | REQ_F_INFLIGHT)) | |
3ca405eb PB |
957 | __io_clean_op(req); |
958 | } | |
959 | ||
4dd2824d PB |
960 | /* not idempotent -- it doesn't clear REQ_F_TASK_PINNED */ |
961 | static void __io_put_req_task(struct io_kiocb *req) | |
962 | { | |
963 | if (req->flags & REQ_F_TASK_PINNED) | |
964 | put_task_struct(req->task); | |
965 | } | |
966 | ||
4349f30e | 967 | static void io_sq_thread_drop_mm(void) |
c40f6379 JA |
968 | { |
969 | struct mm_struct *mm = current->mm; | |
970 | ||
971 | if (mm) { | |
972 | kthread_unuse_mm(mm); | |
973 | mmput(mm); | |
974 | } | |
975 | } | |
976 | ||
977 | static int __io_sq_thread_acquire_mm(struct io_ring_ctx *ctx) | |
978 | { | |
979 | if (!current->mm) { | |
cbcf7214 | 980 | if (unlikely(!(ctx->flags & IORING_SETUP_SQPOLL) || |
2aede0e4 JA |
981 | !ctx->sqo_task->mm || |
982 | !mmget_not_zero(ctx->sqo_task->mm))) | |
c40f6379 | 983 | return -EFAULT; |
2aede0e4 | 984 | kthread_use_mm(ctx->sqo_task->mm); |
c40f6379 JA |
985 | } |
986 | ||
987 | return 0; | |
988 | } | |
989 | ||
990 | static int io_sq_thread_acquire_mm(struct io_ring_ctx *ctx, | |
991 | struct io_kiocb *req) | |
992 | { | |
993 | if (!io_op_defs[req->opcode].needs_mm) | |
994 | return 0; | |
995 | return __io_sq_thread_acquire_mm(ctx); | |
996 | } | |
997 | ||
998 | static inline void req_set_fail_links(struct io_kiocb *req) | |
999 | { | |
1000 | if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK) | |
1001 | req->flags |= REQ_F_FAIL_LINK; | |
1002 | } | |
4a38aed2 | 1003 | |
7cdaf587 XW |
1004 | /* |
1005 | * Note: must call io_req_init_async() for the first time you | |
1006 | * touch any members of io_wq_work. | |
1007 | */ | |
1008 | static inline void io_req_init_async(struct io_kiocb *req) | |
1009 | { | |
1010 | if (req->flags & REQ_F_WORK_INITIALIZED) | |
1011 | return; | |
1012 | ||
1013 | memset(&req->work, 0, sizeof(req->work)); | |
1014 | req->flags |= REQ_F_WORK_INITIALIZED; | |
1015 | } | |
1016 | ||
0cdaf760 PB |
1017 | static inline bool io_async_submit(struct io_ring_ctx *ctx) |
1018 | { | |
1019 | return ctx->flags & IORING_SETUP_SQPOLL; | |
1020 | } | |
1021 | ||
2b188cc1 JA |
1022 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) |
1023 | { | |
1024 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
1025 | ||
0f158b4c | 1026 | complete(&ctx->ref_comp); |
2b188cc1 JA |
1027 | } |
1028 | ||
8eb7e2d0 PB |
1029 | static inline bool io_is_timeout_noseq(struct io_kiocb *req) |
1030 | { | |
1031 | return !req->timeout.off; | |
1032 | } | |
1033 | ||
2b188cc1 JA |
1034 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) |
1035 | { | |
1036 | struct io_ring_ctx *ctx; | |
78076bb6 | 1037 | int hash_bits; |
2b188cc1 JA |
1038 | |
1039 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
1040 | if (!ctx) | |
1041 | return NULL; | |
1042 | ||
0ddf92e8 JA |
1043 | ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL); |
1044 | if (!ctx->fallback_req) | |
1045 | goto err; | |
1046 | ||
78076bb6 JA |
1047 | /* |
1048 | * Use 5 bits less than the max cq entries, that should give us around | |
1049 | * 32 entries per hash list if totally full and uniformly spread. | |
1050 | */ | |
1051 | hash_bits = ilog2(p->cq_entries); | |
1052 | hash_bits -= 5; | |
1053 | if (hash_bits <= 0) | |
1054 | hash_bits = 1; | |
1055 | ctx->cancel_hash_bits = hash_bits; | |
1056 | ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head), | |
1057 | GFP_KERNEL); | |
1058 | if (!ctx->cancel_hash) | |
1059 | goto err; | |
1060 | __hash_init(ctx->cancel_hash, 1U << hash_bits); | |
1061 | ||
21482896 | 1062 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
206aefde JA |
1063 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) |
1064 | goto err; | |
2b188cc1 JA |
1065 | |
1066 | ctx->flags = p->flags; | |
583863ed | 1067 | init_waitqueue_head(&ctx->sqo_wait); |
2b188cc1 | 1068 | init_waitqueue_head(&ctx->cq_wait); |
1d7bb1d5 | 1069 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
0f158b4c JA |
1070 | init_completion(&ctx->ref_comp); |
1071 | init_completion(&ctx->sq_thread_comp); | |
5a2e745d | 1072 | idr_init(&ctx->io_buffer_idr); |
071698e1 | 1073 | idr_init(&ctx->personality_idr); |
2b188cc1 JA |
1074 | mutex_init(&ctx->uring_lock); |
1075 | init_waitqueue_head(&ctx->wait); | |
1076 | spin_lock_init(&ctx->completion_lock); | |
540e32a0 | 1077 | INIT_LIST_HEAD(&ctx->iopoll_list); |
de0617e4 | 1078 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 1079 | INIT_LIST_HEAD(&ctx->timeout_list); |
fcb323cc JA |
1080 | init_waitqueue_head(&ctx->inflight_wait); |
1081 | spin_lock_init(&ctx->inflight_lock); | |
1082 | INIT_LIST_HEAD(&ctx->inflight_list); | |
4a38aed2 JA |
1083 | INIT_DELAYED_WORK(&ctx->file_put_work, io_file_put_work); |
1084 | init_llist_head(&ctx->file_put_llist); | |
2b188cc1 | 1085 | return ctx; |
206aefde | 1086 | err: |
0ddf92e8 JA |
1087 | if (ctx->fallback_req) |
1088 | kmem_cache_free(req_cachep, ctx->fallback_req); | |
78076bb6 | 1089 | kfree(ctx->cancel_hash); |
206aefde JA |
1090 | kfree(ctx); |
1091 | return NULL; | |
2b188cc1 JA |
1092 | } |
1093 | ||
9cf7c104 | 1094 | static bool req_need_defer(struct io_kiocb *req, u32 seq) |
7adf4eaf | 1095 | { |
2bc9930e JA |
1096 | if (unlikely(req->flags & REQ_F_IO_DRAIN)) { |
1097 | struct io_ring_ctx *ctx = req->ctx; | |
a197f664 | 1098 | |
9cf7c104 | 1099 | return seq != ctx->cached_cq_tail |
31af27c7 | 1100 | + atomic_read(&ctx->cached_cq_overflow); |
2bc9930e | 1101 | } |
de0617e4 | 1102 | |
9d858b21 | 1103 | return false; |
de0617e4 JA |
1104 | } |
1105 | ||
de0617e4 | 1106 | static void __io_commit_cqring(struct io_ring_ctx *ctx) |
2b188cc1 | 1107 | { |
75b28aff | 1108 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 1109 | |
07910158 PB |
1110 | /* order cqe stores with ring update */ |
1111 | smp_store_release(&rings->cq.tail, ctx->cached_cq_tail); | |
2b188cc1 | 1112 | |
07910158 PB |
1113 | if (wq_has_sleeper(&ctx->cq_wait)) { |
1114 | wake_up_interruptible(&ctx->cq_wait); | |
1115 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
2b188cc1 JA |
1116 | } |
1117 | } | |
1118 | ||
51a4cc11 JA |
1119 | /* |
1120 | * Returns true if we need to defer file table putting. This can only happen | |
1121 | * from the error path with REQ_F_COMP_LOCKED set. | |
1122 | */ | |
1123 | static bool io_req_clean_work(struct io_kiocb *req) | |
18d9be1a | 1124 | { |
7cdaf587 | 1125 | if (!(req->flags & REQ_F_WORK_INITIALIZED)) |
51a4cc11 JA |
1126 | return false; |
1127 | ||
1128 | req->flags &= ~REQ_F_WORK_INITIALIZED; | |
7cdaf587 | 1129 | |
cccf0ee8 JA |
1130 | if (req->work.mm) { |
1131 | mmdrop(req->work.mm); | |
1132 | req->work.mm = NULL; | |
1133 | } | |
1134 | if (req->work.creds) { | |
1135 | put_cred(req->work.creds); | |
1136 | req->work.creds = NULL; | |
1137 | } | |
ff002b30 JA |
1138 | if (req->work.fs) { |
1139 | struct fs_struct *fs = req->work.fs; | |
1140 | ||
51a4cc11 JA |
1141 | if (req->flags & REQ_F_COMP_LOCKED) |
1142 | return true; | |
1143 | ||
ff002b30 JA |
1144 | spin_lock(&req->work.fs->lock); |
1145 | if (--fs->users) | |
1146 | fs = NULL; | |
1147 | spin_unlock(&req->work.fs->lock); | |
1148 | if (fs) | |
1149 | free_fs_struct(fs); | |
b65e0dd6 | 1150 | req->work.fs = NULL; |
ff002b30 | 1151 | } |
51a4cc11 JA |
1152 | |
1153 | return false; | |
561fb04a JA |
1154 | } |
1155 | ||
cbdcb435 | 1156 | static void io_prep_async_work(struct io_kiocb *req) |
18d9be1a | 1157 | { |
d3656344 | 1158 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
54a91f3b | 1159 | |
16d59803 PB |
1160 | io_req_init_async(req); |
1161 | ||
d3656344 | 1162 | if (req->flags & REQ_F_ISREG) { |
eefdf30f | 1163 | if (def->hash_reg_file || (req->ctx->flags & IORING_SETUP_IOPOLL)) |
8766dd51 | 1164 | io_wq_hash_work(&req->work, file_inode(req->file)); |
d3656344 JA |
1165 | } else { |
1166 | if (def->unbound_nonreg_file) | |
3529d8c2 | 1167 | req->work.flags |= IO_WQ_WORK_UNBOUND; |
54a91f3b | 1168 | } |
dca9cf8b PB |
1169 | if (!req->work.mm && def->needs_mm) { |
1170 | mmgrab(current->mm); | |
1171 | req->work.mm = current->mm; | |
1172 | } | |
1173 | if (!req->work.creds) | |
1174 | req->work.creds = get_current_cred(); | |
1175 | if (!req->work.fs && def->needs_fs) { | |
1176 | spin_lock(¤t->fs->lock); | |
1177 | if (!current->fs->in_exec) { | |
1178 | req->work.fs = current->fs; | |
1179 | req->work.fs->users++; | |
1180 | } else { | |
1181 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
1182 | } | |
1183 | spin_unlock(¤t->fs->lock); | |
1184 | } | |
57f1a649 PB |
1185 | if (def->needs_fsize) |
1186 | req->work.fsize = rlimit(RLIMIT_FSIZE); | |
1187 | else | |
1188 | req->work.fsize = RLIM_INFINITY; | |
561fb04a | 1189 | } |
cccf0ee8 | 1190 | |
cbdcb435 | 1191 | static void io_prep_async_link(struct io_kiocb *req) |
561fb04a | 1192 | { |
cbdcb435 | 1193 | struct io_kiocb *cur; |
54a91f3b | 1194 | |
cbdcb435 PB |
1195 | io_prep_async_work(req); |
1196 | if (req->flags & REQ_F_LINK_HEAD) | |
1197 | list_for_each_entry(cur, &req->link_list, link_list) | |
1198 | io_prep_async_work(cur); | |
561fb04a JA |
1199 | } |
1200 | ||
7271ef3a | 1201 | static struct io_kiocb *__io_queue_async_work(struct io_kiocb *req) |
561fb04a | 1202 | { |
a197f664 | 1203 | struct io_ring_ctx *ctx = req->ctx; |
cbdcb435 | 1204 | struct io_kiocb *link = io_prep_linked_timeout(req); |
561fb04a | 1205 | |
8766dd51 PB |
1206 | trace_io_uring_queue_async_work(ctx, io_wq_is_hashed(&req->work), req, |
1207 | &req->work, req->flags); | |
1208 | io_wq_enqueue(ctx->io_wq, &req->work); | |
7271ef3a | 1209 | return link; |
18d9be1a JA |
1210 | } |
1211 | ||
cbdcb435 PB |
1212 | static void io_queue_async_work(struct io_kiocb *req) |
1213 | { | |
7271ef3a JA |
1214 | struct io_kiocb *link; |
1215 | ||
cbdcb435 PB |
1216 | /* init ->work of the whole link before punting */ |
1217 | io_prep_async_link(req); | |
7271ef3a JA |
1218 | link = __io_queue_async_work(req); |
1219 | ||
1220 | if (link) | |
1221 | io_queue_linked_timeout(link); | |
cbdcb435 PB |
1222 | } |
1223 | ||
5262f567 JA |
1224 | static void io_kill_timeout(struct io_kiocb *req) |
1225 | { | |
1226 | int ret; | |
1227 | ||
2d28390a | 1228 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
5262f567 | 1229 | if (ret != -1) { |
01cec8c1 PB |
1230 | atomic_set(&req->ctx->cq_timeouts, |
1231 | atomic_read(&req->ctx->cq_timeouts) + 1); | |
135fcde8 | 1232 | list_del_init(&req->timeout.list); |
f0e20b89 | 1233 | req->flags |= REQ_F_COMP_LOCKED; |
78e19bbe | 1234 | io_cqring_fill_event(req, 0); |
ec9c02ad | 1235 | io_put_req(req); |
5262f567 JA |
1236 | } |
1237 | } | |
1238 | ||
f3606e3a JA |
1239 | static bool io_task_match(struct io_kiocb *req, struct task_struct *tsk) |
1240 | { | |
1241 | struct io_ring_ctx *ctx = req->ctx; | |
1242 | ||
1243 | if (!tsk || req->task == tsk) | |
1244 | return true; | |
1245 | if ((ctx->flags & IORING_SETUP_SQPOLL) && req->task == ctx->sqo_thread) | |
1246 | return true; | |
1247 | return false; | |
1248 | } | |
1249 | ||
1250 | static void io_kill_timeouts(struct io_ring_ctx *ctx, struct task_struct *tsk) | |
5262f567 JA |
1251 | { |
1252 | struct io_kiocb *req, *tmp; | |
1253 | ||
1254 | spin_lock_irq(&ctx->completion_lock); | |
f3606e3a JA |
1255 | list_for_each_entry_safe(req, tmp, &ctx->timeout_list, timeout.list) { |
1256 | if (io_task_match(req, tsk)) | |
1257 | io_kill_timeout(req); | |
1258 | } | |
5262f567 JA |
1259 | spin_unlock_irq(&ctx->completion_lock); |
1260 | } | |
1261 | ||
04518945 | 1262 | static void __io_queue_deferred(struct io_ring_ctx *ctx) |
de0617e4 | 1263 | { |
04518945 | 1264 | do { |
27dc8338 PB |
1265 | struct io_defer_entry *de = list_first_entry(&ctx->defer_list, |
1266 | struct io_defer_entry, list); | |
7271ef3a | 1267 | struct io_kiocb *link; |
de0617e4 | 1268 | |
9cf7c104 | 1269 | if (req_need_defer(de->req, de->seq)) |
04518945 | 1270 | break; |
27dc8338 | 1271 | list_del_init(&de->list); |
cbdcb435 | 1272 | /* punt-init is done before queueing for defer */ |
7271ef3a JA |
1273 | link = __io_queue_async_work(de->req); |
1274 | if (link) { | |
1275 | __io_queue_linked_timeout(link); | |
1276 | /* drop submission reference */ | |
1277 | link->flags |= REQ_F_COMP_LOCKED; | |
1278 | io_put_req(link); | |
1279 | } | |
27dc8338 | 1280 | kfree(de); |
04518945 PB |
1281 | } while (!list_empty(&ctx->defer_list)); |
1282 | } | |
1283 | ||
360428f8 | 1284 | static void io_flush_timeouts(struct io_ring_ctx *ctx) |
de0617e4 | 1285 | { |
360428f8 PB |
1286 | while (!list_empty(&ctx->timeout_list)) { |
1287 | struct io_kiocb *req = list_first_entry(&ctx->timeout_list, | |
135fcde8 | 1288 | struct io_kiocb, timeout.list); |
de0617e4 | 1289 | |
8eb7e2d0 | 1290 | if (io_is_timeout_noseq(req)) |
360428f8 | 1291 | break; |
bfe68a22 PB |
1292 | if (req->timeout.target_seq != ctx->cached_cq_tail |
1293 | - atomic_read(&ctx->cq_timeouts)) | |
360428f8 | 1294 | break; |
bfe68a22 | 1295 | |
135fcde8 | 1296 | list_del_init(&req->timeout.list); |
5262f567 | 1297 | io_kill_timeout(req); |
360428f8 PB |
1298 | } |
1299 | } | |
5262f567 | 1300 | |
360428f8 PB |
1301 | static void io_commit_cqring(struct io_ring_ctx *ctx) |
1302 | { | |
1303 | io_flush_timeouts(ctx); | |
de0617e4 JA |
1304 | __io_commit_cqring(ctx); |
1305 | ||
04518945 PB |
1306 | if (unlikely(!list_empty(&ctx->defer_list))) |
1307 | __io_queue_deferred(ctx); | |
de0617e4 JA |
1308 | } |
1309 | ||
2b188cc1 JA |
1310 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) |
1311 | { | |
75b28aff | 1312 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
1313 | unsigned tail; |
1314 | ||
1315 | tail = ctx->cached_cq_tail; | |
115e12e5 SB |
1316 | /* |
1317 | * writes to the cq entry need to come after reading head; the | |
1318 | * control dependency is enough as we're using WRITE_ONCE to | |
1319 | * fill the cq entry | |
1320 | */ | |
75b28aff | 1321 | if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries) |
2b188cc1 JA |
1322 | return NULL; |
1323 | ||
1324 | ctx->cached_cq_tail++; | |
75b28aff | 1325 | return &rings->cqes[tail & ctx->cq_mask]; |
2b188cc1 JA |
1326 | } |
1327 | ||
f2842ab5 JA |
1328 | static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx) |
1329 | { | |
f0b493e6 JA |
1330 | if (!ctx->cq_ev_fd) |
1331 | return false; | |
7e55a19c SG |
1332 | if (READ_ONCE(ctx->rings->cq_flags) & IORING_CQ_EVENTFD_DISABLED) |
1333 | return false; | |
f2842ab5 JA |
1334 | if (!ctx->eventfd_async) |
1335 | return true; | |
b41e9852 | 1336 | return io_wq_current_is_worker(); |
f2842ab5 JA |
1337 | } |
1338 | ||
b41e9852 | 1339 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
1d7bb1d5 JA |
1340 | { |
1341 | if (waitqueue_active(&ctx->wait)) | |
1342 | wake_up(&ctx->wait); | |
1343 | if (waitqueue_active(&ctx->sqo_wait)) | |
1344 | wake_up(&ctx->sqo_wait); | |
b41e9852 | 1345 | if (io_should_trigger_evfd(ctx)) |
1d7bb1d5 JA |
1346 | eventfd_signal(ctx->cq_ev_fd, 1); |
1347 | } | |
1348 | ||
46930143 PB |
1349 | static void io_cqring_mark_overflow(struct io_ring_ctx *ctx) |
1350 | { | |
1351 | if (list_empty(&ctx->cq_overflow_list)) { | |
1352 | clear_bit(0, &ctx->sq_check_overflow); | |
1353 | clear_bit(0, &ctx->cq_check_overflow); | |
1354 | ctx->rings->sq_flags &= ~IORING_SQ_CQ_OVERFLOW; | |
1355 | } | |
1356 | } | |
1357 | ||
c4a2ed72 JA |
1358 | /* Returns true if there are no backlogged entries after the flush */ |
1359 | static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force) | |
1d7bb1d5 JA |
1360 | { |
1361 | struct io_rings *rings = ctx->rings; | |
1362 | struct io_uring_cqe *cqe; | |
1363 | struct io_kiocb *req; | |
1364 | unsigned long flags; | |
1365 | LIST_HEAD(list); | |
1366 | ||
1367 | if (!force) { | |
1368 | if (list_empty_careful(&ctx->cq_overflow_list)) | |
c4a2ed72 | 1369 | return true; |
1d7bb1d5 JA |
1370 | if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) == |
1371 | rings->cq_ring_entries)) | |
c4a2ed72 | 1372 | return false; |
1d7bb1d5 JA |
1373 | } |
1374 | ||
1375 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
1376 | ||
1377 | /* if force is set, the ring is going away. always drop after that */ | |
1378 | if (force) | |
69b3e546 | 1379 | ctx->cq_overflow_flushed = 1; |
1d7bb1d5 | 1380 | |
c4a2ed72 | 1381 | cqe = NULL; |
1d7bb1d5 JA |
1382 | while (!list_empty(&ctx->cq_overflow_list)) { |
1383 | cqe = io_get_cqring(ctx); | |
1384 | if (!cqe && !force) | |
1385 | break; | |
1386 | ||
1387 | req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb, | |
40d8ddd4 PB |
1388 | compl.list); |
1389 | list_move(&req->compl.list, &list); | |
1d7bb1d5 JA |
1390 | if (cqe) { |
1391 | WRITE_ONCE(cqe->user_data, req->user_data); | |
1392 | WRITE_ONCE(cqe->res, req->result); | |
0f7e466b | 1393 | WRITE_ONCE(cqe->flags, req->compl.cflags); |
1d7bb1d5 JA |
1394 | } else { |
1395 | WRITE_ONCE(ctx->rings->cq_overflow, | |
1396 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
1397 | } | |
1398 | } | |
1399 | ||
1400 | io_commit_cqring(ctx); | |
46930143 PB |
1401 | io_cqring_mark_overflow(ctx); |
1402 | ||
1d7bb1d5 JA |
1403 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
1404 | io_cqring_ev_posted(ctx); | |
1405 | ||
1406 | while (!list_empty(&list)) { | |
40d8ddd4 PB |
1407 | req = list_first_entry(&list, struct io_kiocb, compl.list); |
1408 | list_del(&req->compl.list); | |
ec9c02ad | 1409 | io_put_req(req); |
1d7bb1d5 | 1410 | } |
c4a2ed72 JA |
1411 | |
1412 | return cqe != NULL; | |
1d7bb1d5 JA |
1413 | } |
1414 | ||
bcda7baa | 1415 | static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags) |
2b188cc1 | 1416 | { |
78e19bbe | 1417 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
1418 | struct io_uring_cqe *cqe; |
1419 | ||
78e19bbe | 1420 | trace_io_uring_complete(ctx, req->user_data, res); |
51c3ff62 | 1421 | |
2b188cc1 JA |
1422 | /* |
1423 | * If we can't get a cq entry, userspace overflowed the | |
1424 | * submission (by quite a lot). Increment the overflow count in | |
1425 | * the ring. | |
1426 | */ | |
1427 | cqe = io_get_cqring(ctx); | |
1d7bb1d5 | 1428 | if (likely(cqe)) { |
78e19bbe | 1429 | WRITE_ONCE(cqe->user_data, req->user_data); |
2b188cc1 | 1430 | WRITE_ONCE(cqe->res, res); |
bcda7baa | 1431 | WRITE_ONCE(cqe->flags, cflags); |
1d7bb1d5 | 1432 | } else if (ctx->cq_overflow_flushed) { |
498ccd9e JA |
1433 | WRITE_ONCE(ctx->rings->cq_overflow, |
1434 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
1d7bb1d5 | 1435 | } else { |
ad3eb2c8 JA |
1436 | if (list_empty(&ctx->cq_overflow_list)) { |
1437 | set_bit(0, &ctx->sq_check_overflow); | |
1438 | set_bit(0, &ctx->cq_check_overflow); | |
6d5f9049 | 1439 | ctx->rings->sq_flags |= IORING_SQ_CQ_OVERFLOW; |
ad3eb2c8 | 1440 | } |
40d8ddd4 | 1441 | io_clean_op(req); |
1d7bb1d5 | 1442 | req->result = res; |
0f7e466b | 1443 | req->compl.cflags = cflags; |
40d8ddd4 PB |
1444 | refcount_inc(&req->refs); |
1445 | list_add_tail(&req->compl.list, &ctx->cq_overflow_list); | |
2b188cc1 JA |
1446 | } |
1447 | } | |
1448 | ||
bcda7baa JA |
1449 | static void io_cqring_fill_event(struct io_kiocb *req, long res) |
1450 | { | |
1451 | __io_cqring_fill_event(req, res, 0); | |
1452 | } | |
1453 | ||
e1e16097 | 1454 | static void io_cqring_add_event(struct io_kiocb *req, long res, long cflags) |
2b188cc1 | 1455 | { |
78e19bbe | 1456 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
1457 | unsigned long flags; |
1458 | ||
1459 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
bcda7baa | 1460 | __io_cqring_fill_event(req, res, cflags); |
2b188cc1 JA |
1461 | io_commit_cqring(ctx); |
1462 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1463 | ||
8c838788 | 1464 | io_cqring_ev_posted(ctx); |
2b188cc1 JA |
1465 | } |
1466 | ||
229a7b63 | 1467 | static void io_submit_flush_completions(struct io_comp_state *cs) |
bcda7baa | 1468 | { |
229a7b63 JA |
1469 | struct io_ring_ctx *ctx = cs->ctx; |
1470 | ||
1471 | spin_lock_irq(&ctx->completion_lock); | |
1472 | while (!list_empty(&cs->list)) { | |
1473 | struct io_kiocb *req; | |
1474 | ||
3ca405eb PB |
1475 | req = list_first_entry(&cs->list, struct io_kiocb, compl.list); |
1476 | list_del(&req->compl.list); | |
0f7e466b | 1477 | __io_cqring_fill_event(req, req->result, req->compl.cflags); |
229a7b63 JA |
1478 | if (!(req->flags & REQ_F_LINK_HEAD)) { |
1479 | req->flags |= REQ_F_COMP_LOCKED; | |
1480 | io_put_req(req); | |
1481 | } else { | |
1482 | spin_unlock_irq(&ctx->completion_lock); | |
1483 | io_put_req(req); | |
1484 | spin_lock_irq(&ctx->completion_lock); | |
1485 | } | |
1486 | } | |
1487 | io_commit_cqring(ctx); | |
1488 | spin_unlock_irq(&ctx->completion_lock); | |
1489 | ||
1490 | io_cqring_ev_posted(ctx); | |
1491 | cs->nr = 0; | |
1492 | } | |
1493 | ||
1494 | static void __io_req_complete(struct io_kiocb *req, long res, unsigned cflags, | |
1495 | struct io_comp_state *cs) | |
1496 | { | |
1497 | if (!cs) { | |
1498 | io_cqring_add_event(req, res, cflags); | |
1499 | io_put_req(req); | |
1500 | } else { | |
3ca405eb | 1501 | io_clean_op(req); |
229a7b63 | 1502 | req->result = res; |
0f7e466b | 1503 | req->compl.cflags = cflags; |
3ca405eb | 1504 | list_add_tail(&req->compl.list, &cs->list); |
229a7b63 JA |
1505 | if (++cs->nr >= 32) |
1506 | io_submit_flush_completions(cs); | |
1507 | } | |
e1e16097 JA |
1508 | } |
1509 | ||
1510 | static void io_req_complete(struct io_kiocb *req, long res) | |
bcda7baa | 1511 | { |
229a7b63 | 1512 | __io_req_complete(req, res, 0, NULL); |
bcda7baa JA |
1513 | } |
1514 | ||
0ddf92e8 JA |
1515 | static inline bool io_is_fallback_req(struct io_kiocb *req) |
1516 | { | |
1517 | return req == (struct io_kiocb *) | |
1518 | ((unsigned long) req->ctx->fallback_req & ~1UL); | |
1519 | } | |
1520 | ||
1521 | static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx) | |
1522 | { | |
1523 | struct io_kiocb *req; | |
1524 | ||
1525 | req = ctx->fallback_req; | |
dd461af6 | 1526 | if (!test_and_set_bit_lock(0, (unsigned long *) &ctx->fallback_req)) |
0ddf92e8 JA |
1527 | return req; |
1528 | ||
1529 | return NULL; | |
1530 | } | |
1531 | ||
0553b8bd PB |
1532 | static struct io_kiocb *io_alloc_req(struct io_ring_ctx *ctx, |
1533 | struct io_submit_state *state) | |
2b188cc1 | 1534 | { |
fd6fab2c | 1535 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
2b188cc1 JA |
1536 | struct io_kiocb *req; |
1537 | ||
f6b6c7d6 | 1538 | if (!state->free_reqs) { |
2579f913 JA |
1539 | size_t sz; |
1540 | int ret; | |
1541 | ||
1542 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
fd6fab2c JA |
1543 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs); |
1544 | ||
1545 | /* | |
1546 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
1547 | * retry single alloc to be on the safe side. | |
1548 | */ | |
1549 | if (unlikely(ret <= 0)) { | |
1550 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
1551 | if (!state->reqs[0]) | |
0ddf92e8 | 1552 | goto fallback; |
fd6fab2c JA |
1553 | ret = 1; |
1554 | } | |
2579f913 | 1555 | state->free_reqs = ret - 1; |
6c8a3134 | 1556 | req = state->reqs[ret - 1]; |
2579f913 | 1557 | } else { |
2579f913 | 1558 | state->free_reqs--; |
6c8a3134 | 1559 | req = state->reqs[state->free_reqs]; |
2b188cc1 JA |
1560 | } |
1561 | ||
2579f913 | 1562 | return req; |
0ddf92e8 | 1563 | fallback: |
0553b8bd | 1564 | return io_get_fallback_req(ctx); |
2b188cc1 JA |
1565 | } |
1566 | ||
8da11c19 PB |
1567 | static inline void io_put_file(struct io_kiocb *req, struct file *file, |
1568 | bool fixed) | |
1569 | { | |
1570 | if (fixed) | |
05589553 | 1571 | percpu_ref_put(req->fixed_file_refs); |
8da11c19 PB |
1572 | else |
1573 | fput(file); | |
1574 | } | |
1575 | ||
51a4cc11 | 1576 | static bool io_dismantle_req(struct io_kiocb *req) |
2b188cc1 | 1577 | { |
3ca405eb | 1578 | io_clean_op(req); |
929a3af9 | 1579 | |
5acbbc8e JA |
1580 | if (req->io) |
1581 | kfree(req->io); | |
8da11c19 PB |
1582 | if (req->file) |
1583 | io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE)); | |
fcb323cc | 1584 | |
51a4cc11 | 1585 | return io_req_clean_work(req); |
e65ef56d JA |
1586 | } |
1587 | ||
51a4cc11 | 1588 | static void __io_free_req_finish(struct io_kiocb *req) |
c6ca97b3 | 1589 | { |
51a4cc11 | 1590 | struct io_ring_ctx *ctx = req->ctx; |
c6ca97b3 | 1591 | |
5af1d13e | 1592 | __io_put_req_task(req); |
b1e50e54 PB |
1593 | if (likely(!io_is_fallback_req(req))) |
1594 | kmem_cache_free(req_cachep, req); | |
1595 | else | |
ecfc5177 PB |
1596 | clear_bit_unlock(0, (unsigned long *) &ctx->fallback_req); |
1597 | percpu_ref_put(&ctx->refs); | |
e65ef56d JA |
1598 | } |
1599 | ||
51a4cc11 JA |
1600 | static void io_req_task_file_table_put(struct callback_head *cb) |
1601 | { | |
1602 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); | |
1603 | struct fs_struct *fs = req->work.fs; | |
1604 | ||
1605 | spin_lock(&req->work.fs->lock); | |
1606 | if (--fs->users) | |
1607 | fs = NULL; | |
1608 | spin_unlock(&req->work.fs->lock); | |
1609 | if (fs) | |
1610 | free_fs_struct(fs); | |
1611 | req->work.fs = NULL; | |
1612 | __io_free_req_finish(req); | |
1613 | } | |
1614 | ||
1615 | static void __io_free_req(struct io_kiocb *req) | |
1616 | { | |
1617 | if (!io_dismantle_req(req)) { | |
1618 | __io_free_req_finish(req); | |
1619 | } else { | |
1620 | int ret; | |
1621 | ||
1622 | init_task_work(&req->task_work, io_req_task_file_table_put); | |
1623 | ret = task_work_add(req->task, &req->task_work, TWA_RESUME); | |
1624 | if (unlikely(ret)) { | |
1625 | struct task_struct *tsk; | |
1626 | ||
1627 | tsk = io_wq_get_task(req->ctx->io_wq); | |
1628 | task_work_add(tsk, &req->task_work, 0); | |
1629 | } | |
1630 | } | |
1631 | } | |
1632 | ||
a197f664 | 1633 | static bool io_link_cancel_timeout(struct io_kiocb *req) |
2665abfd | 1634 | { |
a197f664 | 1635 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd JA |
1636 | int ret; |
1637 | ||
2d28390a | 1638 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
2665abfd | 1639 | if (ret != -1) { |
78e19bbe | 1640 | io_cqring_fill_event(req, -ECANCELED); |
2665abfd | 1641 | io_commit_cqring(ctx); |
dea3b49c | 1642 | req->flags &= ~REQ_F_LINK_HEAD; |
ec9c02ad | 1643 | io_put_req(req); |
2665abfd JA |
1644 | return true; |
1645 | } | |
1646 | ||
1647 | return false; | |
e65ef56d JA |
1648 | } |
1649 | ||
ab0b6451 | 1650 | static bool __io_kill_linked_timeout(struct io_kiocb *req) |
9e645e11 | 1651 | { |
7c86ffee | 1652 | struct io_kiocb *link; |
ab0b6451 | 1653 | bool wake_ev; |
7c86ffee PB |
1654 | |
1655 | if (list_empty(&req->link_list)) | |
ab0b6451 | 1656 | return false; |
7c86ffee PB |
1657 | link = list_first_entry(&req->link_list, struct io_kiocb, link_list); |
1658 | if (link->opcode != IORING_OP_LINK_TIMEOUT) | |
ab0b6451 | 1659 | return false; |
7c86ffee PB |
1660 | |
1661 | list_del_init(&link->link_list); | |
9b7adba9 | 1662 | link->flags |= REQ_F_COMP_LOCKED; |
7c86ffee PB |
1663 | wake_ev = io_link_cancel_timeout(link); |
1664 | req->flags &= ~REQ_F_LINK_TIMEOUT; | |
ab0b6451 JA |
1665 | return wake_ev; |
1666 | } | |
1667 | ||
1668 | static void io_kill_linked_timeout(struct io_kiocb *req) | |
9e645e11 | 1669 | { |
2665abfd | 1670 | struct io_ring_ctx *ctx = req->ctx; |
ab0b6451 | 1671 | bool wake_ev; |
9e645e11 | 1672 | |
ab0b6451 JA |
1673 | if (!(req->flags & REQ_F_COMP_LOCKED)) { |
1674 | unsigned long flags; | |
1675 | ||
1676 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
1677 | wake_ev = __io_kill_linked_timeout(req); | |
7c86ffee | 1678 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
ab0b6451 JA |
1679 | } else { |
1680 | wake_ev = __io_kill_linked_timeout(req); | |
1681 | } | |
1682 | ||
7c86ffee PB |
1683 | if (wake_ev) |
1684 | io_cqring_ev_posted(ctx); | |
1685 | } | |
1686 | ||
9b5f7bd9 | 1687 | static struct io_kiocb *io_req_link_next(struct io_kiocb *req) |
7c86ffee PB |
1688 | { |
1689 | struct io_kiocb *nxt; | |
4d7dd462 | 1690 | |
9e645e11 JA |
1691 | /* |
1692 | * The list should never be empty when we are called here. But could | |
1693 | * potentially happen if the chain is messed up, check to be on the | |
1694 | * safe side. | |
1695 | */ | |
7c86ffee | 1696 | if (unlikely(list_empty(&req->link_list))) |
9b5f7bd9 | 1697 | return NULL; |
2665abfd | 1698 | |
7c86ffee PB |
1699 | nxt = list_first_entry(&req->link_list, struct io_kiocb, link_list); |
1700 | list_del_init(&req->link_list); | |
1701 | if (!list_empty(&nxt->link_list)) | |
1702 | nxt->flags |= REQ_F_LINK_HEAD; | |
9b5f7bd9 | 1703 | return nxt; |
9e645e11 JA |
1704 | } |
1705 | ||
1706 | /* | |
dea3b49c | 1707 | * Called if REQ_F_LINK_HEAD is set, and we fail the head request |
9e645e11 | 1708 | */ |
7c86ffee | 1709 | static void __io_fail_links(struct io_kiocb *req) |
9e645e11 | 1710 | { |
2665abfd | 1711 | struct io_ring_ctx *ctx = req->ctx; |
9e645e11 JA |
1712 | |
1713 | while (!list_empty(&req->link_list)) { | |
4493233e PB |
1714 | struct io_kiocb *link = list_first_entry(&req->link_list, |
1715 | struct io_kiocb, link_list); | |
9e645e11 | 1716 | |
4493233e | 1717 | list_del_init(&link->link_list); |
c826bd7a | 1718 | trace_io_uring_fail_link(req, link); |
2665abfd | 1719 | |
7c86ffee | 1720 | io_cqring_fill_event(link, -ECANCELED); |
9b7adba9 | 1721 | link->flags |= REQ_F_COMP_LOCKED; |
7c86ffee | 1722 | __io_double_put_req(link); |
5d960724 | 1723 | req->flags &= ~REQ_F_LINK_TIMEOUT; |
9e645e11 | 1724 | } |
2665abfd JA |
1725 | |
1726 | io_commit_cqring(ctx); | |
2665abfd | 1727 | io_cqring_ev_posted(ctx); |
9e645e11 JA |
1728 | } |
1729 | ||
7c86ffee | 1730 | static void io_fail_links(struct io_kiocb *req) |
9e645e11 | 1731 | { |
7c86ffee | 1732 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd | 1733 | |
7c86ffee | 1734 | if (!(req->flags & REQ_F_COMP_LOCKED)) { |
2665abfd JA |
1735 | unsigned long flags; |
1736 | ||
2665abfd | 1737 | spin_lock_irqsave(&ctx->completion_lock, flags); |
7c86ffee | 1738 | __io_fail_links(req); |
2665abfd JA |
1739 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
1740 | } else { | |
7c86ffee | 1741 | __io_fail_links(req); |
9e645e11 JA |
1742 | } |
1743 | ||
2665abfd | 1744 | io_cqring_ev_posted(ctx); |
9e645e11 JA |
1745 | } |
1746 | ||
3fa5e0f3 | 1747 | static struct io_kiocb *__io_req_find_next(struct io_kiocb *req) |
c69f8dbe | 1748 | { |
9b0d911a | 1749 | req->flags &= ~REQ_F_LINK_HEAD; |
7c86ffee PB |
1750 | if (req->flags & REQ_F_LINK_TIMEOUT) |
1751 | io_kill_linked_timeout(req); | |
944e58bf | 1752 | |
9e645e11 JA |
1753 | /* |
1754 | * If LINK is set, we have dependent requests in this chain. If we | |
1755 | * didn't fail this request, queue the first one up, moving any other | |
1756 | * dependencies to the next request. In case of failure, fail the rest | |
1757 | * of the chain. | |
1758 | */ | |
9b5f7bd9 PB |
1759 | if (likely(!(req->flags & REQ_F_FAIL_LINK))) |
1760 | return io_req_link_next(req); | |
1761 | io_fail_links(req); | |
1762 | return NULL; | |
4d7dd462 | 1763 | } |
9e645e11 | 1764 | |
3fa5e0f3 PB |
1765 | static struct io_kiocb *io_req_find_next(struct io_kiocb *req) |
1766 | { | |
1767 | if (likely(!(req->flags & REQ_F_LINK_HEAD))) | |
1768 | return NULL; | |
1769 | return __io_req_find_next(req); | |
1770 | } | |
1771 | ||
fd7d6de2 JA |
1772 | static int io_req_task_work_add(struct io_kiocb *req, struct callback_head *cb, |
1773 | bool twa_signal_ok) | |
c2c4c83c JA |
1774 | { |
1775 | struct task_struct *tsk = req->task; | |
1776 | struct io_ring_ctx *ctx = req->ctx; | |
0ba9c9ed | 1777 | int ret, notify; |
c2c4c83c | 1778 | |
6200b0ae JA |
1779 | if (tsk->flags & PF_EXITING) |
1780 | return -ESRCH; | |
1781 | ||
c2c4c83c | 1782 | /* |
0ba9c9ed JA |
1783 | * SQPOLL kernel thread doesn't need notification, just a wakeup. For |
1784 | * all other cases, use TWA_SIGNAL unconditionally to ensure we're | |
1785 | * processing task_work. There's no reliable way to tell if TWA_RESUME | |
1786 | * will do the job. | |
c2c4c83c | 1787 | */ |
0ba9c9ed | 1788 | notify = 0; |
fd7d6de2 | 1789 | if (!(ctx->flags & IORING_SETUP_SQPOLL) && twa_signal_ok) |
c2c4c83c JA |
1790 | notify = TWA_SIGNAL; |
1791 | ||
1792 | ret = task_work_add(tsk, cb, notify); | |
1793 | if (!ret) | |
1794 | wake_up_process(tsk); | |
0ba9c9ed | 1795 | |
c2c4c83c JA |
1796 | return ret; |
1797 | } | |
1798 | ||
c40f6379 JA |
1799 | static void __io_req_task_cancel(struct io_kiocb *req, int error) |
1800 | { | |
1801 | struct io_ring_ctx *ctx = req->ctx; | |
1802 | ||
1803 | spin_lock_irq(&ctx->completion_lock); | |
1804 | io_cqring_fill_event(req, error); | |
1805 | io_commit_cqring(ctx); | |
1806 | spin_unlock_irq(&ctx->completion_lock); | |
1807 | ||
1808 | io_cqring_ev_posted(ctx); | |
1809 | req_set_fail_links(req); | |
1810 | io_double_put_req(req); | |
1811 | } | |
1812 | ||
1813 | static void io_req_task_cancel(struct callback_head *cb) | |
1814 | { | |
1815 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); | |
87ceb6a6 | 1816 | struct io_ring_ctx *ctx = req->ctx; |
c40f6379 JA |
1817 | |
1818 | __io_req_task_cancel(req, -ECANCELED); | |
87ceb6a6 | 1819 | percpu_ref_put(&ctx->refs); |
c40f6379 JA |
1820 | } |
1821 | ||
1822 | static void __io_req_task_submit(struct io_kiocb *req) | |
1823 | { | |
1824 | struct io_ring_ctx *ctx = req->ctx; | |
1825 | ||
c40f6379 JA |
1826 | if (!__io_sq_thread_acquire_mm(ctx)) { |
1827 | mutex_lock(&ctx->uring_lock); | |
1828 | __io_queue_sqe(req, NULL, NULL); | |
1829 | mutex_unlock(&ctx->uring_lock); | |
1830 | } else { | |
1831 | __io_req_task_cancel(req, -EFAULT); | |
1832 | } | |
1833 | } | |
1834 | ||
1835 | static void io_req_task_submit(struct callback_head *cb) | |
1836 | { | |
1837 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); | |
6d816e08 | 1838 | struct io_ring_ctx *ctx = req->ctx; |
c40f6379 JA |
1839 | |
1840 | __io_req_task_submit(req); | |
6d816e08 | 1841 | percpu_ref_put(&ctx->refs); |
c40f6379 JA |
1842 | } |
1843 | ||
1844 | static void io_req_task_queue(struct io_kiocb *req) | |
1845 | { | |
c40f6379 JA |
1846 | int ret; |
1847 | ||
1848 | init_task_work(&req->task_work, io_req_task_submit); | |
6d816e08 | 1849 | percpu_ref_get(&req->ctx->refs); |
c40f6379 | 1850 | |
fd7d6de2 | 1851 | ret = io_req_task_work_add(req, &req->task_work, true); |
c40f6379 | 1852 | if (unlikely(ret)) { |
c2c4c83c JA |
1853 | struct task_struct *tsk; |
1854 | ||
c40f6379 JA |
1855 | init_task_work(&req->task_work, io_req_task_cancel); |
1856 | tsk = io_wq_get_task(req->ctx->io_wq); | |
c2c4c83c JA |
1857 | task_work_add(tsk, &req->task_work, 0); |
1858 | wake_up_process(tsk); | |
c40f6379 | 1859 | } |
c40f6379 JA |
1860 | } |
1861 | ||
c3524383 | 1862 | static void io_queue_next(struct io_kiocb *req) |
c69f8dbe | 1863 | { |
9b5f7bd9 | 1864 | struct io_kiocb *nxt = io_req_find_next(req); |
944e58bf PB |
1865 | |
1866 | if (nxt) | |
906a8c3f | 1867 | io_req_task_queue(nxt); |
c69f8dbe JL |
1868 | } |
1869 | ||
c3524383 | 1870 | static void io_free_req(struct io_kiocb *req) |
7a743e22 | 1871 | { |
c3524383 PB |
1872 | io_queue_next(req); |
1873 | __io_free_req(req); | |
1874 | } | |
8766dd51 | 1875 | |
2d6500d4 PB |
1876 | struct req_batch { |
1877 | void *reqs[IO_IOPOLL_BATCH]; | |
1878 | int to_free; | |
7a743e22 | 1879 | |
5af1d13e PB |
1880 | struct task_struct *task; |
1881 | int task_refs; | |
2d6500d4 PB |
1882 | }; |
1883 | ||
5af1d13e PB |
1884 | static inline void io_init_req_batch(struct req_batch *rb) |
1885 | { | |
1886 | rb->to_free = 0; | |
1887 | rb->task_refs = 0; | |
1888 | rb->task = NULL; | |
1889 | } | |
1890 | ||
2d6500d4 PB |
1891 | static void __io_req_free_batch_flush(struct io_ring_ctx *ctx, |
1892 | struct req_batch *rb) | |
1893 | { | |
1894 | kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs); | |
1895 | percpu_ref_put_many(&ctx->refs, rb->to_free); | |
1896 | rb->to_free = 0; | |
1897 | } | |
1898 | ||
1899 | static void io_req_free_batch_finish(struct io_ring_ctx *ctx, | |
1900 | struct req_batch *rb) | |
1901 | { | |
1902 | if (rb->to_free) | |
1903 | __io_req_free_batch_flush(ctx, rb); | |
5af1d13e PB |
1904 | if (rb->task) { |
1905 | put_task_struct_many(rb->task, rb->task_refs); | |
1906 | rb->task = NULL; | |
1907 | } | |
2d6500d4 PB |
1908 | } |
1909 | ||
1910 | static void io_req_free_batch(struct req_batch *rb, struct io_kiocb *req) | |
1911 | { | |
1912 | if (unlikely(io_is_fallback_req(req))) { | |
1913 | io_free_req(req); | |
1914 | return; | |
1915 | } | |
1916 | if (req->flags & REQ_F_LINK_HEAD) | |
1917 | io_queue_next(req); | |
1918 | ||
5af1d13e PB |
1919 | if (req->flags & REQ_F_TASK_PINNED) { |
1920 | if (req->task != rb->task) { | |
1921 | if (rb->task) | |
1922 | put_task_struct_many(rb->task, rb->task_refs); | |
1923 | rb->task = req->task; | |
1924 | rb->task_refs = 0; | |
1925 | } | |
1926 | rb->task_refs++; | |
1927 | req->flags &= ~REQ_F_TASK_PINNED; | |
1928 | } | |
1929 | ||
51a4cc11 | 1930 | WARN_ON_ONCE(io_dismantle_req(req)); |
2d6500d4 PB |
1931 | rb->reqs[rb->to_free++] = req; |
1932 | if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs))) | |
1933 | __io_req_free_batch_flush(req->ctx, rb); | |
7a743e22 PB |
1934 | } |
1935 | ||
ba816ad6 JA |
1936 | /* |
1937 | * Drop reference to request, return next in chain (if there is one) if this | |
1938 | * was the last reference to this request. | |
1939 | */ | |
9b5f7bd9 | 1940 | static struct io_kiocb *io_put_req_find_next(struct io_kiocb *req) |
e65ef56d | 1941 | { |
9b5f7bd9 PB |
1942 | struct io_kiocb *nxt = NULL; |
1943 | ||
2a44f467 | 1944 | if (refcount_dec_and_test(&req->refs)) { |
9b5f7bd9 | 1945 | nxt = io_req_find_next(req); |
4d7dd462 | 1946 | __io_free_req(req); |
2a44f467 | 1947 | } |
9b5f7bd9 | 1948 | return nxt; |
2b188cc1 JA |
1949 | } |
1950 | ||
e65ef56d JA |
1951 | static void io_put_req(struct io_kiocb *req) |
1952 | { | |
1953 | if (refcount_dec_and_test(&req->refs)) | |
1954 | io_free_req(req); | |
2b188cc1 JA |
1955 | } |
1956 | ||
f4db7182 | 1957 | static struct io_wq_work *io_steal_work(struct io_kiocb *req) |
7a743e22 | 1958 | { |
6df1db6b | 1959 | struct io_kiocb *nxt; |
f4db7182 | 1960 | |
7a743e22 | 1961 | /* |
f4db7182 PB |
1962 | * A ref is owned by io-wq in which context we're. So, if that's the |
1963 | * last one, it's safe to steal next work. False negatives are Ok, | |
1964 | * it just will be re-punted async in io_put_work() | |
7a743e22 | 1965 | */ |
f4db7182 PB |
1966 | if (refcount_read(&req->refs) != 1) |
1967 | return NULL; | |
7a743e22 | 1968 | |
9b5f7bd9 | 1969 | nxt = io_req_find_next(req); |
6df1db6b | 1970 | return nxt ? &nxt->work : NULL; |
7a743e22 PB |
1971 | } |
1972 | ||
978db57e JA |
1973 | /* |
1974 | * Must only be used if we don't need to care about links, usually from | |
1975 | * within the completion handling itself. | |
1976 | */ | |
1977 | static void __io_double_put_req(struct io_kiocb *req) | |
78e19bbe JA |
1978 | { |
1979 | /* drop both submit and complete references */ | |
1980 | if (refcount_sub_and_test(2, &req->refs)) | |
1981 | __io_free_req(req); | |
1982 | } | |
1983 | ||
978db57e JA |
1984 | static void io_double_put_req(struct io_kiocb *req) |
1985 | { | |
1986 | /* drop both submit and complete references */ | |
1987 | if (refcount_sub_and_test(2, &req->refs)) | |
1988 | io_free_req(req); | |
1989 | } | |
1990 | ||
1d7bb1d5 | 1991 | static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush) |
a3a0e43f | 1992 | { |
84f97dc2 JA |
1993 | struct io_rings *rings = ctx->rings; |
1994 | ||
ad3eb2c8 JA |
1995 | if (test_bit(0, &ctx->cq_check_overflow)) { |
1996 | /* | |
1997 | * noflush == true is from the waitqueue handler, just ensure | |
1998 | * we wake up the task, and the next invocation will flush the | |
1999 | * entries. We cannot safely to it from here. | |
2000 | */ | |
2001 | if (noflush && !list_empty(&ctx->cq_overflow_list)) | |
2002 | return -1U; | |
1d7bb1d5 | 2003 | |
ad3eb2c8 JA |
2004 | io_cqring_overflow_flush(ctx, false); |
2005 | } | |
1d7bb1d5 | 2006 | |
a3a0e43f JA |
2007 | /* See comment at the top of this file */ |
2008 | smp_rmb(); | |
ad3eb2c8 | 2009 | return ctx->cached_cq_tail - READ_ONCE(rings->cq.head); |
a3a0e43f JA |
2010 | } |
2011 | ||
fb5ccc98 PB |
2012 | static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) |
2013 | { | |
2014 | struct io_rings *rings = ctx->rings; | |
2015 | ||
2016 | /* make sure SQ entry isn't read before tail */ | |
2017 | return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; | |
2018 | } | |
2019 | ||
8ff069bf | 2020 | static unsigned int io_put_kbuf(struct io_kiocb *req, struct io_buffer *kbuf) |
e94f141b | 2021 | { |
8ff069bf | 2022 | unsigned int cflags; |
e94f141b | 2023 | |
bcda7baa JA |
2024 | cflags = kbuf->bid << IORING_CQE_BUFFER_SHIFT; |
2025 | cflags |= IORING_CQE_F_BUFFER; | |
0e1b6fe3 | 2026 | req->flags &= ~REQ_F_BUFFER_SELECTED; |
bcda7baa JA |
2027 | kfree(kbuf); |
2028 | return cflags; | |
e94f141b JA |
2029 | } |
2030 | ||
8ff069bf | 2031 | static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req) |
bcda7baa | 2032 | { |
4d954c25 | 2033 | struct io_buffer *kbuf; |
bcda7baa | 2034 | |
4d954c25 | 2035 | kbuf = (struct io_buffer *) (unsigned long) req->rw.addr; |
8ff069bf PB |
2036 | return io_put_kbuf(req, kbuf); |
2037 | } | |
2038 | ||
4c6e277c JA |
2039 | static inline bool io_run_task_work(void) |
2040 | { | |
6200b0ae JA |
2041 | /* |
2042 | * Not safe to run on exiting task, and the task_work handling will | |
2043 | * not add work to such a task. | |
2044 | */ | |
2045 | if (unlikely(current->flags & PF_EXITING)) | |
2046 | return false; | |
4c6e277c JA |
2047 | if (current->task_works) { |
2048 | __set_current_state(TASK_RUNNING); | |
2049 | task_work_run(); | |
2050 | return true; | |
2051 | } | |
2052 | ||
2053 | return false; | |
bcda7baa JA |
2054 | } |
2055 | ||
bbde017a XW |
2056 | static void io_iopoll_queue(struct list_head *again) |
2057 | { | |
2058 | struct io_kiocb *req; | |
2059 | ||
2060 | do { | |
d21ffe7e PB |
2061 | req = list_first_entry(again, struct io_kiocb, inflight_entry); |
2062 | list_del(&req->inflight_entry); | |
81b68a5c | 2063 | __io_complete_rw(req, -EAGAIN, 0, NULL); |
bbde017a XW |
2064 | } while (!list_empty(again)); |
2065 | } | |
2066 | ||
def596e9 JA |
2067 | /* |
2068 | * Find and free completed poll iocbs | |
2069 | */ | |
2070 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
2071 | struct list_head *done) | |
2072 | { | |
8237e045 | 2073 | struct req_batch rb; |
def596e9 | 2074 | struct io_kiocb *req; |
bbde017a XW |
2075 | LIST_HEAD(again); |
2076 | ||
2077 | /* order with ->result store in io_complete_rw_iopoll() */ | |
2078 | smp_rmb(); | |
def596e9 | 2079 | |
5af1d13e | 2080 | io_init_req_batch(&rb); |
def596e9 | 2081 | while (!list_empty(done)) { |
bcda7baa JA |
2082 | int cflags = 0; |
2083 | ||
d21ffe7e | 2084 | req = list_first_entry(done, struct io_kiocb, inflight_entry); |
bbde017a | 2085 | if (READ_ONCE(req->result) == -EAGAIN) { |
56450c20 | 2086 | req->result = 0; |
bbde017a | 2087 | req->iopoll_completed = 0; |
d21ffe7e | 2088 | list_move_tail(&req->inflight_entry, &again); |
bbde017a XW |
2089 | continue; |
2090 | } | |
d21ffe7e | 2091 | list_del(&req->inflight_entry); |
def596e9 | 2092 | |
bcda7baa | 2093 | if (req->flags & REQ_F_BUFFER_SELECTED) |
8ff069bf | 2094 | cflags = io_put_rw_kbuf(req); |
bcda7baa JA |
2095 | |
2096 | __io_cqring_fill_event(req, req->result, cflags); | |
def596e9 JA |
2097 | (*nr_events)++; |
2098 | ||
c3524383 | 2099 | if (refcount_dec_and_test(&req->refs)) |
2d6500d4 | 2100 | io_req_free_batch(&rb, req); |
def596e9 | 2101 | } |
def596e9 | 2102 | |
09bb8394 | 2103 | io_commit_cqring(ctx); |
32b2244a XW |
2104 | if (ctx->flags & IORING_SETUP_SQPOLL) |
2105 | io_cqring_ev_posted(ctx); | |
2d6500d4 | 2106 | io_req_free_batch_finish(ctx, &rb); |
581f9810 | 2107 | |
bbde017a XW |
2108 | if (!list_empty(&again)) |
2109 | io_iopoll_queue(&again); | |
581f9810 BM |
2110 | } |
2111 | ||
def596e9 JA |
2112 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, |
2113 | long min) | |
2114 | { | |
2115 | struct io_kiocb *req, *tmp; | |
2116 | LIST_HEAD(done); | |
2117 | bool spin; | |
2118 | int ret; | |
2119 | ||
2120 | /* | |
2121 | * Only spin for completions if we don't have multiple devices hanging | |
2122 | * off our complete list, and we're under the requested amount. | |
2123 | */ | |
2124 | spin = !ctx->poll_multi_file && *nr_events < min; | |
2125 | ||
2126 | ret = 0; | |
d21ffe7e | 2127 | list_for_each_entry_safe(req, tmp, &ctx->iopoll_list, inflight_entry) { |
9adbd45d | 2128 | struct kiocb *kiocb = &req->rw.kiocb; |
def596e9 JA |
2129 | |
2130 | /* | |
581f9810 BM |
2131 | * Move completed and retryable entries to our local lists. |
2132 | * If we find a request that requires polling, break out | |
2133 | * and complete those lists first, if we have entries there. | |
def596e9 | 2134 | */ |
65a6543d | 2135 | if (READ_ONCE(req->iopoll_completed)) { |
d21ffe7e | 2136 | list_move_tail(&req->inflight_entry, &done); |
def596e9 JA |
2137 | continue; |
2138 | } | |
2139 | if (!list_empty(&done)) | |
2140 | break; | |
2141 | ||
2142 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
2143 | if (ret < 0) | |
2144 | break; | |
2145 | ||
3aadc23e PB |
2146 | /* iopoll may have completed current req */ |
2147 | if (READ_ONCE(req->iopoll_completed)) | |
d21ffe7e | 2148 | list_move_tail(&req->inflight_entry, &done); |
3aadc23e | 2149 | |
def596e9 JA |
2150 | if (ret && spin) |
2151 | spin = false; | |
2152 | ret = 0; | |
2153 | } | |
2154 | ||
2155 | if (!list_empty(&done)) | |
2156 | io_iopoll_complete(ctx, nr_events, &done); | |
2157 | ||
2158 | return ret; | |
2159 | } | |
2160 | ||
2161 | /* | |
d195a66e | 2162 | * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a |
def596e9 JA |
2163 | * non-spinning poll check - we'll still enter the driver poll loop, but only |
2164 | * as a non-spinning completion check. | |
2165 | */ | |
2166 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
2167 | long min) | |
2168 | { | |
540e32a0 | 2169 | while (!list_empty(&ctx->iopoll_list) && !need_resched()) { |
def596e9 JA |
2170 | int ret; |
2171 | ||
2172 | ret = io_do_iopoll(ctx, nr_events, min); | |
2173 | if (ret < 0) | |
2174 | return ret; | |
eba0a4dd | 2175 | if (*nr_events >= min) |
def596e9 JA |
2176 | return 0; |
2177 | } | |
2178 | ||
2179 | return 1; | |
2180 | } | |
2181 | ||
2182 | /* | |
2183 | * We can't just wait for polled events to come to us, we have to actively | |
2184 | * find and complete them. | |
2185 | */ | |
b2edc0a7 | 2186 | static void io_iopoll_try_reap_events(struct io_ring_ctx *ctx) |
def596e9 JA |
2187 | { |
2188 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
2189 | return; | |
2190 | ||
2191 | mutex_lock(&ctx->uring_lock); | |
540e32a0 | 2192 | while (!list_empty(&ctx->iopoll_list)) { |
def596e9 JA |
2193 | unsigned int nr_events = 0; |
2194 | ||
b2edc0a7 | 2195 | io_do_iopoll(ctx, &nr_events, 0); |
08f5439f | 2196 | |
b2edc0a7 PB |
2197 | /* let it sleep and repeat later if can't complete a request */ |
2198 | if (nr_events == 0) | |
2199 | break; | |
08f5439f JA |
2200 | /* |
2201 | * Ensure we allow local-to-the-cpu processing to take place, | |
2202 | * in this case we need to ensure that we reap all events. | |
3fcee5a6 | 2203 | * Also let task_work, etc. to progress by releasing the mutex |
08f5439f | 2204 | */ |
3fcee5a6 PB |
2205 | if (need_resched()) { |
2206 | mutex_unlock(&ctx->uring_lock); | |
2207 | cond_resched(); | |
2208 | mutex_lock(&ctx->uring_lock); | |
2209 | } | |
def596e9 JA |
2210 | } |
2211 | mutex_unlock(&ctx->uring_lock); | |
2212 | } | |
2213 | ||
7668b92a | 2214 | static int io_iopoll_check(struct io_ring_ctx *ctx, long min) |
def596e9 | 2215 | { |
7668b92a | 2216 | unsigned int nr_events = 0; |
2b2ed975 | 2217 | int iters = 0, ret = 0; |
500f9fba | 2218 | |
c7849be9 XW |
2219 | /* |
2220 | * We disallow the app entering submit/complete with polling, but we | |
2221 | * still need to lock the ring to prevent racing with polled issue | |
2222 | * that got punted to a workqueue. | |
2223 | */ | |
2224 | mutex_lock(&ctx->uring_lock); | |
def596e9 | 2225 | do { |
a3a0e43f JA |
2226 | /* |
2227 | * Don't enter poll loop if we already have events pending. | |
2228 | * If we do, we can potentially be spinning for commands that | |
2229 | * already triggered a CQE (eg in error). | |
2230 | */ | |
1d7bb1d5 | 2231 | if (io_cqring_events(ctx, false)) |
a3a0e43f JA |
2232 | break; |
2233 | ||
500f9fba JA |
2234 | /* |
2235 | * If a submit got punted to a workqueue, we can have the | |
2236 | * application entering polling for a command before it gets | |
2237 | * issued. That app will hold the uring_lock for the duration | |
2238 | * of the poll right here, so we need to take a breather every | |
2239 | * now and then to ensure that the issue has a chance to add | |
2240 | * the poll to the issued list. Otherwise we can spin here | |
2241 | * forever, while the workqueue is stuck trying to acquire the | |
2242 | * very same mutex. | |
2243 | */ | |
2244 | if (!(++iters & 7)) { | |
2245 | mutex_unlock(&ctx->uring_lock); | |
4c6e277c | 2246 | io_run_task_work(); |
500f9fba JA |
2247 | mutex_lock(&ctx->uring_lock); |
2248 | } | |
2249 | ||
7668b92a | 2250 | ret = io_iopoll_getevents(ctx, &nr_events, min); |
def596e9 JA |
2251 | if (ret <= 0) |
2252 | break; | |
2253 | ret = 0; | |
7668b92a | 2254 | } while (min && !nr_events && !need_resched()); |
def596e9 | 2255 | |
500f9fba | 2256 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
2257 | return ret; |
2258 | } | |
2259 | ||
491381ce | 2260 | static void kiocb_end_write(struct io_kiocb *req) |
2b188cc1 | 2261 | { |
491381ce JA |
2262 | /* |
2263 | * Tell lockdep we inherited freeze protection from submission | |
2264 | * thread. | |
2265 | */ | |
2266 | if (req->flags & REQ_F_ISREG) { | |
2267 | struct inode *inode = file_inode(req->file); | |
2b188cc1 | 2268 | |
491381ce | 2269 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); |
2b188cc1 | 2270 | } |
491381ce | 2271 | file_end_write(req->file); |
2b188cc1 JA |
2272 | } |
2273 | ||
a1d7c393 JA |
2274 | static void io_complete_rw_common(struct kiocb *kiocb, long res, |
2275 | struct io_comp_state *cs) | |
2b188cc1 | 2276 | { |
9adbd45d | 2277 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
bcda7baa | 2278 | int cflags = 0; |
2b188cc1 | 2279 | |
491381ce JA |
2280 | if (kiocb->ki_flags & IOCB_WRITE) |
2281 | kiocb_end_write(req); | |
2b188cc1 | 2282 | |
4e88d6e7 JA |
2283 | if (res != req->result) |
2284 | req_set_fail_links(req); | |
bcda7baa | 2285 | if (req->flags & REQ_F_BUFFER_SELECTED) |
8ff069bf | 2286 | cflags = io_put_rw_kbuf(req); |
a1d7c393 | 2287 | __io_req_complete(req, res, cflags, cs); |
ba816ad6 JA |
2288 | } |
2289 | ||
b63534c4 JA |
2290 | #ifdef CONFIG_BLOCK |
2291 | static bool io_resubmit_prep(struct io_kiocb *req, int error) | |
2292 | { | |
2293 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
2294 | ssize_t ret = -ECANCELED; | |
2295 | struct iov_iter iter; | |
2296 | int rw; | |
2297 | ||
2298 | if (error) { | |
2299 | ret = error; | |
2300 | goto end_req; | |
2301 | } | |
2302 | ||
2303 | switch (req->opcode) { | |
2304 | case IORING_OP_READV: | |
2305 | case IORING_OP_READ_FIXED: | |
2306 | case IORING_OP_READ: | |
2307 | rw = READ; | |
2308 | break; | |
2309 | case IORING_OP_WRITEV: | |
2310 | case IORING_OP_WRITE_FIXED: | |
2311 | case IORING_OP_WRITE: | |
2312 | rw = WRITE; | |
2313 | break; | |
2314 | default: | |
2315 | printk_once(KERN_WARNING "io_uring: bad opcode in resubmit %d\n", | |
2316 | req->opcode); | |
2317 | goto end_req; | |
2318 | } | |
2319 | ||
8f3d7496 JA |
2320 | if (!req->io) { |
2321 | ret = io_import_iovec(rw, req, &iovec, &iter, false); | |
2322 | if (ret < 0) | |
2323 | goto end_req; | |
2324 | ret = io_setup_async_rw(req, iovec, inline_vecs, &iter, false); | |
2325 | if (!ret) | |
2326 | return true; | |
2327 | kfree(iovec); | |
2328 | } else { | |
b63534c4 | 2329 | return true; |
8f3d7496 | 2330 | } |
b63534c4 | 2331 | end_req: |
b63534c4 | 2332 | req_set_fail_links(req); |
e1e16097 | 2333 | io_req_complete(req, ret); |
b63534c4 JA |
2334 | return false; |
2335 | } | |
b63534c4 JA |
2336 | #endif |
2337 | ||
2338 | static bool io_rw_reissue(struct io_kiocb *req, long res) | |
2339 | { | |
2340 | #ifdef CONFIG_BLOCK | |
355afaeb | 2341 | umode_t mode = file_inode(req->file)->i_mode; |
b63534c4 JA |
2342 | int ret; |
2343 | ||
355afaeb JA |
2344 | if (!S_ISBLK(mode) && !S_ISREG(mode)) |
2345 | return false; | |
b63534c4 JA |
2346 | if ((res != -EAGAIN && res != -EOPNOTSUPP) || io_wq_current_is_worker()) |
2347 | return false; | |
2348 | ||
fdee946d | 2349 | ret = io_sq_thread_acquire_mm(req->ctx, req); |
6d816e08 | 2350 | |
fdee946d JA |
2351 | if (io_resubmit_prep(req, ret)) { |
2352 | refcount_inc(&req->refs); | |
2353 | io_queue_async_work(req); | |
b63534c4 | 2354 | return true; |
fdee946d JA |
2355 | } |
2356 | ||
b63534c4 JA |
2357 | #endif |
2358 | return false; | |
2359 | } | |
2360 | ||
a1d7c393 JA |
2361 | static void __io_complete_rw(struct io_kiocb *req, long res, long res2, |
2362 | struct io_comp_state *cs) | |
2363 | { | |
2364 | if (!io_rw_reissue(req, res)) | |
2365 | io_complete_rw_common(&req->rw.kiocb, res, cs); | |
ba816ad6 JA |
2366 | } |
2367 | ||
2368 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
2369 | { | |
9adbd45d | 2370 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
ba816ad6 | 2371 | |
a1d7c393 | 2372 | __io_complete_rw(req, res, res2, NULL); |
2b188cc1 JA |
2373 | } |
2374 | ||
def596e9 JA |
2375 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
2376 | { | |
9adbd45d | 2377 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
def596e9 | 2378 | |
491381ce JA |
2379 | if (kiocb->ki_flags & IOCB_WRITE) |
2380 | kiocb_end_write(req); | |
def596e9 | 2381 | |
2d7d6792 | 2382 | if (res != -EAGAIN && res != req->result) |
4e88d6e7 | 2383 | req_set_fail_links(req); |
bbde017a XW |
2384 | |
2385 | WRITE_ONCE(req->result, res); | |
2386 | /* order with io_poll_complete() checking ->result */ | |
cd664b0e PB |
2387 | smp_wmb(); |
2388 | WRITE_ONCE(req->iopoll_completed, 1); | |
def596e9 JA |
2389 | } |
2390 | ||
2391 | /* | |
2392 | * After the iocb has been issued, it's safe to be found on the poll list. | |
2393 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
2394 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
2395 | * accessing the kiocb cookie. | |
2396 | */ | |
2397 | static void io_iopoll_req_issued(struct io_kiocb *req) | |
2398 | { | |
2399 | struct io_ring_ctx *ctx = req->ctx; | |
2400 | ||
2401 | /* | |
2402 | * Track whether we have multiple files in our lists. This will impact | |
2403 | * how we do polling eventually, not spinning if we're on potentially | |
2404 | * different devices. | |
2405 | */ | |
540e32a0 | 2406 | if (list_empty(&ctx->iopoll_list)) { |
def596e9 JA |
2407 | ctx->poll_multi_file = false; |
2408 | } else if (!ctx->poll_multi_file) { | |
2409 | struct io_kiocb *list_req; | |
2410 | ||
540e32a0 | 2411 | list_req = list_first_entry(&ctx->iopoll_list, struct io_kiocb, |
d21ffe7e | 2412 | inflight_entry); |
9adbd45d | 2413 | if (list_req->file != req->file) |
def596e9 JA |
2414 | ctx->poll_multi_file = true; |
2415 | } | |
2416 | ||
2417 | /* | |
2418 | * For fast devices, IO may have already completed. If it has, add | |
2419 | * it to the front so we find it first. | |
2420 | */ | |
65a6543d | 2421 | if (READ_ONCE(req->iopoll_completed)) |
d21ffe7e | 2422 | list_add(&req->inflight_entry, &ctx->iopoll_list); |
def596e9 | 2423 | else |
d21ffe7e | 2424 | list_add_tail(&req->inflight_entry, &ctx->iopoll_list); |
bdcd3eab XW |
2425 | |
2426 | if ((ctx->flags & IORING_SETUP_SQPOLL) && | |
2427 | wq_has_sleeper(&ctx->sqo_wait)) | |
2428 | wake_up(&ctx->sqo_wait); | |
def596e9 JA |
2429 | } |
2430 | ||
9f13c35b | 2431 | static void __io_state_file_put(struct io_submit_state *state) |
9a56a232 | 2432 | { |
06ef3608 PB |
2433 | if (state->has_refs) |
2434 | fput_many(state->file, state->has_refs); | |
9f13c35b PB |
2435 | state->file = NULL; |
2436 | } | |
2437 | ||
2438 | static inline void io_state_file_put(struct io_submit_state *state) | |
2439 | { | |
2440 | if (state->file) | |
2441 | __io_state_file_put(state); | |
9a56a232 JA |
2442 | } |
2443 | ||
2444 | /* | |
2445 | * Get as many references to a file as we have IOs left in this submission, | |
2446 | * assuming most submissions are for one file, or at least that each file | |
2447 | * has more than one submission. | |
2448 | */ | |
8da11c19 | 2449 | static struct file *__io_file_get(struct io_submit_state *state, int fd) |
9a56a232 JA |
2450 | { |
2451 | if (!state) | |
2452 | return fget(fd); | |
2453 | ||
2454 | if (state->file) { | |
2455 | if (state->fd == fd) { | |
06ef3608 | 2456 | state->has_refs--; |
9a56a232 JA |
2457 | state->ios_left--; |
2458 | return state->file; | |
2459 | } | |
9f13c35b | 2460 | __io_state_file_put(state); |
9a56a232 JA |
2461 | } |
2462 | state->file = fget_many(fd, state->ios_left); | |
2463 | if (!state->file) | |
2464 | return NULL; | |
2465 | ||
2466 | state->fd = fd; | |
9a56a232 | 2467 | state->ios_left--; |
06ef3608 | 2468 | state->has_refs = state->ios_left; |
9a56a232 JA |
2469 | return state->file; |
2470 | } | |
2471 | ||
4503b767 JA |
2472 | static bool io_bdev_nowait(struct block_device *bdev) |
2473 | { | |
2474 | #ifdef CONFIG_BLOCK | |
2475 | return !bdev || queue_is_mq(bdev_get_queue(bdev)); | |
2476 | #else | |
2477 | return true; | |
2478 | #endif | |
2479 | } | |
2480 | ||
2b188cc1 JA |
2481 | /* |
2482 | * If we tracked the file through the SCM inflight mechanism, we could support | |
2483 | * any file. For now, just ensure that anything potentially problematic is done | |
2484 | * inline. | |
2485 | */ | |
af197f50 | 2486 | static bool io_file_supports_async(struct file *file, int rw) |
2b188cc1 JA |
2487 | { |
2488 | umode_t mode = file_inode(file)->i_mode; | |
2489 | ||
4503b767 JA |
2490 | if (S_ISBLK(mode)) { |
2491 | if (io_bdev_nowait(file->f_inode->i_bdev)) | |
2492 | return true; | |
2493 | return false; | |
2494 | } | |
2495 | if (S_ISCHR(mode) || S_ISSOCK(mode)) | |
2b188cc1 | 2496 | return true; |
4503b767 JA |
2497 | if (S_ISREG(mode)) { |
2498 | if (io_bdev_nowait(file->f_inode->i_sb->s_bdev) && | |
2499 | file->f_op != &io_uring_fops) | |
2500 | return true; | |
2501 | return false; | |
2502 | } | |
2b188cc1 | 2503 | |
c5b85625 JA |
2504 | /* any ->read/write should understand O_NONBLOCK */ |
2505 | if (file->f_flags & O_NONBLOCK) | |
2506 | return true; | |
2507 | ||
af197f50 JA |
2508 | if (!(file->f_mode & FMODE_NOWAIT)) |
2509 | return false; | |
2510 | ||
2511 | if (rw == READ) | |
2512 | return file->f_op->read_iter != NULL; | |
2513 | ||
2514 | return file->f_op->write_iter != NULL; | |
2b188cc1 JA |
2515 | } |
2516 | ||
3529d8c2 JA |
2517 | static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2518 | bool force_nonblock) | |
2b188cc1 | 2519 | { |
def596e9 | 2520 | struct io_ring_ctx *ctx = req->ctx; |
9adbd45d | 2521 | struct kiocb *kiocb = &req->rw.kiocb; |
09bb8394 JA |
2522 | unsigned ioprio; |
2523 | int ret; | |
2b188cc1 | 2524 | |
491381ce JA |
2525 | if (S_ISREG(file_inode(req->file)->i_mode)) |
2526 | req->flags |= REQ_F_ISREG; | |
2527 | ||
2b188cc1 | 2528 | kiocb->ki_pos = READ_ONCE(sqe->off); |
ba04291e JA |
2529 | if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) { |
2530 | req->flags |= REQ_F_CUR_POS; | |
2531 | kiocb->ki_pos = req->file->f_pos; | |
2532 | } | |
2b188cc1 | 2533 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); |
3e577dcd PB |
2534 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); |
2535 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
2536 | if (unlikely(ret)) | |
2537 | return ret; | |
2b188cc1 JA |
2538 | |
2539 | ioprio = READ_ONCE(sqe->ioprio); | |
2540 | if (ioprio) { | |
2541 | ret = ioprio_check_cap(ioprio); | |
2542 | if (ret) | |
09bb8394 | 2543 | return ret; |
2b188cc1 JA |
2544 | |
2545 | kiocb->ki_ioprio = ioprio; | |
2546 | } else | |
2547 | kiocb->ki_ioprio = get_current_ioprio(); | |
2548 | ||
8449eeda | 2549 | /* don't allow async punt if RWF_NOWAIT was requested */ |
c5b85625 | 2550 | if (kiocb->ki_flags & IOCB_NOWAIT) |
8449eeda SB |
2551 | req->flags |= REQ_F_NOWAIT; |
2552 | ||
b63534c4 JA |
2553 | if (kiocb->ki_flags & IOCB_DIRECT) |
2554 | io_get_req_task(req); | |
2555 | ||
8449eeda | 2556 | if (force_nonblock) |
2b188cc1 | 2557 | kiocb->ki_flags |= IOCB_NOWAIT; |
8449eeda | 2558 | |
def596e9 | 2559 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
2560 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
2561 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 2562 | return -EOPNOTSUPP; |
2b188cc1 | 2563 | |
def596e9 JA |
2564 | kiocb->ki_flags |= IOCB_HIPRI; |
2565 | kiocb->ki_complete = io_complete_rw_iopoll; | |
65a6543d | 2566 | req->iopoll_completed = 0; |
f3a6fa22 | 2567 | io_get_req_task(req); |
def596e9 | 2568 | } else { |
09bb8394 JA |
2569 | if (kiocb->ki_flags & IOCB_HIPRI) |
2570 | return -EINVAL; | |
def596e9 JA |
2571 | kiocb->ki_complete = io_complete_rw; |
2572 | } | |
9adbd45d | 2573 | |
3529d8c2 JA |
2574 | req->rw.addr = READ_ONCE(sqe->addr); |
2575 | req->rw.len = READ_ONCE(sqe->len); | |
4f4eeba8 | 2576 | req->buf_index = READ_ONCE(sqe->buf_index); |
2b188cc1 | 2577 | return 0; |
2b188cc1 JA |
2578 | } |
2579 | ||
2580 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
2581 | { | |
2582 | switch (ret) { | |
2583 | case -EIOCBQUEUED: | |
2584 | break; | |
2585 | case -ERESTARTSYS: | |
2586 | case -ERESTARTNOINTR: | |
2587 | case -ERESTARTNOHAND: | |
2588 | case -ERESTART_RESTARTBLOCK: | |
2589 | /* | |
2590 | * We can't just restart the syscall, since previously | |
2591 | * submitted sqes may already be in progress. Just fail this | |
2592 | * IO with EINTR. | |
2593 | */ | |
2594 | ret = -EINTR; | |
df561f66 | 2595 | fallthrough; |
2b188cc1 JA |
2596 | default: |
2597 | kiocb->ki_complete(kiocb, ret, 0); | |
2598 | } | |
2599 | } | |
2600 | ||
a1d7c393 JA |
2601 | static void kiocb_done(struct kiocb *kiocb, ssize_t ret, |
2602 | struct io_comp_state *cs) | |
ba816ad6 | 2603 | { |
ba04291e JA |
2604 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
2605 | ||
227c0c96 JA |
2606 | /* add previously done IO, if any */ |
2607 | if (req->io && req->io->rw.bytes_done > 0) { | |
2608 | if (ret < 0) | |
2609 | ret = req->io->rw.bytes_done; | |
2610 | else | |
2611 | ret += req->io->rw.bytes_done; | |
2612 | } | |
2613 | ||
ba04291e JA |
2614 | if (req->flags & REQ_F_CUR_POS) |
2615 | req->file->f_pos = kiocb->ki_pos; | |
bcaec089 | 2616 | if (ret >= 0 && kiocb->ki_complete == io_complete_rw) |
a1d7c393 | 2617 | __io_complete_rw(req, ret, 0, cs); |
ba816ad6 JA |
2618 | else |
2619 | io_rw_done(kiocb, ret); | |
2620 | } | |
2621 | ||
9adbd45d | 2622 | static ssize_t io_import_fixed(struct io_kiocb *req, int rw, |
7d009165 | 2623 | struct iov_iter *iter) |
edafccee | 2624 | { |
9adbd45d JA |
2625 | struct io_ring_ctx *ctx = req->ctx; |
2626 | size_t len = req->rw.len; | |
edafccee | 2627 | struct io_mapped_ubuf *imu; |
4f4eeba8 | 2628 | u16 index, buf_index; |
edafccee JA |
2629 | size_t offset; |
2630 | u64 buf_addr; | |
2631 | ||
2632 | /* attempt to use fixed buffers without having provided iovecs */ | |
2633 | if (unlikely(!ctx->user_bufs)) | |
2634 | return -EFAULT; | |
2635 | ||
4f4eeba8 | 2636 | buf_index = req->buf_index; |
edafccee JA |
2637 | if (unlikely(buf_index >= ctx->nr_user_bufs)) |
2638 | return -EFAULT; | |
2639 | ||
2640 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
2641 | imu = &ctx->user_bufs[index]; | |
9adbd45d | 2642 | buf_addr = req->rw.addr; |
edafccee JA |
2643 | |
2644 | /* overflow */ | |
2645 | if (buf_addr + len < buf_addr) | |
2646 | return -EFAULT; | |
2647 | /* not inside the mapped region */ | |
2648 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
2649 | return -EFAULT; | |
2650 | ||
2651 | /* | |
2652 | * May not be a start of buffer, set size appropriately | |
2653 | * and advance us to the beginning. | |
2654 | */ | |
2655 | offset = buf_addr - imu->ubuf; | |
2656 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
2657 | |
2658 | if (offset) { | |
2659 | /* | |
2660 | * Don't use iov_iter_advance() here, as it's really slow for | |
2661 | * using the latter parts of a big fixed buffer - it iterates | |
2662 | * over each segment manually. We can cheat a bit here, because | |
2663 | * we know that: | |
2664 | * | |
2665 | * 1) it's a BVEC iter, we set it up | |
2666 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
2667 | * first and last bvec | |
2668 | * | |
2669 | * So just find our index, and adjust the iterator afterwards. | |
2670 | * If the offset is within the first bvec (or the whole first | |
2671 | * bvec, just use iov_iter_advance(). This makes it easier | |
2672 | * since we can just skip the first segment, which may not | |
2673 | * be PAGE_SIZE aligned. | |
2674 | */ | |
2675 | const struct bio_vec *bvec = imu->bvec; | |
2676 | ||
2677 | if (offset <= bvec->bv_len) { | |
2678 | iov_iter_advance(iter, offset); | |
2679 | } else { | |
2680 | unsigned long seg_skip; | |
2681 | ||
2682 | /* skip first vec */ | |
2683 | offset -= bvec->bv_len; | |
2684 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
2685 | ||
2686 | iter->bvec = bvec + seg_skip; | |
2687 | iter->nr_segs -= seg_skip; | |
99c79f66 | 2688 | iter->count -= bvec->bv_len + offset; |
bd11b3a3 | 2689 | iter->iov_offset = offset & ~PAGE_MASK; |
bd11b3a3 JA |
2690 | } |
2691 | } | |
2692 | ||
5e559561 | 2693 | return len; |
edafccee JA |
2694 | } |
2695 | ||
bcda7baa JA |
2696 | static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock) |
2697 | { | |
2698 | if (needs_lock) | |
2699 | mutex_unlock(&ctx->uring_lock); | |
2700 | } | |
2701 | ||
2702 | static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock) | |
2703 | { | |
2704 | /* | |
2705 | * "Normal" inline submissions always hold the uring_lock, since we | |
2706 | * grab it from the system call. Same is true for the SQPOLL offload. | |
2707 | * The only exception is when we've detached the request and issue it | |
2708 | * from an async worker thread, grab the lock for that case. | |
2709 | */ | |
2710 | if (needs_lock) | |
2711 | mutex_lock(&ctx->uring_lock); | |
2712 | } | |
2713 | ||
2714 | static struct io_buffer *io_buffer_select(struct io_kiocb *req, size_t *len, | |
2715 | int bgid, struct io_buffer *kbuf, | |
2716 | bool needs_lock) | |
2717 | { | |
2718 | struct io_buffer *head; | |
2719 | ||
2720 | if (req->flags & REQ_F_BUFFER_SELECTED) | |
2721 | return kbuf; | |
2722 | ||
2723 | io_ring_submit_lock(req->ctx, needs_lock); | |
2724 | ||
2725 | lockdep_assert_held(&req->ctx->uring_lock); | |
2726 | ||
2727 | head = idr_find(&req->ctx->io_buffer_idr, bgid); | |
2728 | if (head) { | |
2729 | if (!list_empty(&head->list)) { | |
2730 | kbuf = list_last_entry(&head->list, struct io_buffer, | |
2731 | list); | |
2732 | list_del(&kbuf->list); | |
2733 | } else { | |
2734 | kbuf = head; | |
2735 | idr_remove(&req->ctx->io_buffer_idr, bgid); | |
2736 | } | |
2737 | if (*len > kbuf->len) | |
2738 | *len = kbuf->len; | |
2739 | } else { | |
2740 | kbuf = ERR_PTR(-ENOBUFS); | |
2741 | } | |
2742 | ||
2743 | io_ring_submit_unlock(req->ctx, needs_lock); | |
2744 | ||
2745 | return kbuf; | |
2746 | } | |
2747 | ||
4d954c25 JA |
2748 | static void __user *io_rw_buffer_select(struct io_kiocb *req, size_t *len, |
2749 | bool needs_lock) | |
2750 | { | |
2751 | struct io_buffer *kbuf; | |
4f4eeba8 | 2752 | u16 bgid; |
4d954c25 JA |
2753 | |
2754 | kbuf = (struct io_buffer *) (unsigned long) req->rw.addr; | |
4f4eeba8 | 2755 | bgid = req->buf_index; |
4d954c25 JA |
2756 | kbuf = io_buffer_select(req, len, bgid, kbuf, needs_lock); |
2757 | if (IS_ERR(kbuf)) | |
2758 | return kbuf; | |
2759 | req->rw.addr = (u64) (unsigned long) kbuf; | |
2760 | req->flags |= REQ_F_BUFFER_SELECTED; | |
2761 | return u64_to_user_ptr(kbuf->addr); | |
2762 | } | |
2763 | ||
2764 | #ifdef CONFIG_COMPAT | |
2765 | static ssize_t io_compat_import(struct io_kiocb *req, struct iovec *iov, | |
2766 | bool needs_lock) | |
2767 | { | |
2768 | struct compat_iovec __user *uiov; | |
2769 | compat_ssize_t clen; | |
2770 | void __user *buf; | |
2771 | ssize_t len; | |
2772 | ||
2773 | uiov = u64_to_user_ptr(req->rw.addr); | |
2774 | if (!access_ok(uiov, sizeof(*uiov))) | |
2775 | return -EFAULT; | |
2776 | if (__get_user(clen, &uiov->iov_len)) | |
2777 | return -EFAULT; | |
2778 | if (clen < 0) | |
2779 | return -EINVAL; | |
2780 | ||
2781 | len = clen; | |
2782 | buf = io_rw_buffer_select(req, &len, needs_lock); | |
2783 | if (IS_ERR(buf)) | |
2784 | return PTR_ERR(buf); | |
2785 | iov[0].iov_base = buf; | |
2786 | iov[0].iov_len = (compat_size_t) len; | |
2787 | return 0; | |
2788 | } | |
2789 | #endif | |
2790 | ||
2791 | static ssize_t __io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov, | |
2792 | bool needs_lock) | |
2793 | { | |
2794 | struct iovec __user *uiov = u64_to_user_ptr(req->rw.addr); | |
2795 | void __user *buf; | |
2796 | ssize_t len; | |
2797 | ||
2798 | if (copy_from_user(iov, uiov, sizeof(*uiov))) | |
2799 | return -EFAULT; | |
2800 | ||
2801 | len = iov[0].iov_len; | |
2802 | if (len < 0) | |
2803 | return -EINVAL; | |
2804 | buf = io_rw_buffer_select(req, &len, needs_lock); | |
2805 | if (IS_ERR(buf)) | |
2806 | return PTR_ERR(buf); | |
2807 | iov[0].iov_base = buf; | |
2808 | iov[0].iov_len = len; | |
2809 | return 0; | |
2810 | } | |
2811 | ||
2812 | static ssize_t io_iov_buffer_select(struct io_kiocb *req, struct iovec *iov, | |
2813 | bool needs_lock) | |
2814 | { | |
dddb3e26 JA |
2815 | if (req->flags & REQ_F_BUFFER_SELECTED) { |
2816 | struct io_buffer *kbuf; | |
2817 | ||
2818 | kbuf = (struct io_buffer *) (unsigned long) req->rw.addr; | |
2819 | iov[0].iov_base = u64_to_user_ptr(kbuf->addr); | |
2820 | iov[0].iov_len = kbuf->len; | |
4d954c25 | 2821 | return 0; |
dddb3e26 | 2822 | } |
4d954c25 JA |
2823 | if (!req->rw.len) |
2824 | return 0; | |
2825 | else if (req->rw.len > 1) | |
2826 | return -EINVAL; | |
2827 | ||
2828 | #ifdef CONFIG_COMPAT | |
2829 | if (req->ctx->compat) | |
2830 | return io_compat_import(req, iov, needs_lock); | |
2831 | #endif | |
2832 | ||
2833 | return __io_iov_buffer_select(req, iov, needs_lock); | |
2834 | } | |
2835 | ||
8452fd0c JA |
2836 | static ssize_t __io_import_iovec(int rw, struct io_kiocb *req, |
2837 | struct iovec **iovec, struct iov_iter *iter, | |
2838 | bool needs_lock) | |
2b188cc1 | 2839 | { |
9adbd45d JA |
2840 | void __user *buf = u64_to_user_ptr(req->rw.addr); |
2841 | size_t sqe_len = req->rw.len; | |
4d954c25 | 2842 | ssize_t ret; |
edafccee JA |
2843 | u8 opcode; |
2844 | ||
d625c6ee | 2845 | opcode = req->opcode; |
7d009165 | 2846 | if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) { |
edafccee | 2847 | *iovec = NULL; |
9adbd45d | 2848 | return io_import_fixed(req, rw, iter); |
edafccee | 2849 | } |
2b188cc1 | 2850 | |
bcda7baa | 2851 | /* buffer index only valid with fixed read/write, or buffer select */ |
4f4eeba8 | 2852 | if (req->buf_index && !(req->flags & REQ_F_BUFFER_SELECT)) |
9adbd45d JA |
2853 | return -EINVAL; |
2854 | ||
3a6820f2 | 2855 | if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) { |
bcda7baa | 2856 | if (req->flags & REQ_F_BUFFER_SELECT) { |
4d954c25 | 2857 | buf = io_rw_buffer_select(req, &sqe_len, needs_lock); |
867a23ea | 2858 | if (IS_ERR(buf)) |
4d954c25 | 2859 | return PTR_ERR(buf); |
3f9d6441 | 2860 | req->rw.len = sqe_len; |
bcda7baa JA |
2861 | } |
2862 | ||
3a6820f2 JA |
2863 | ret = import_single_range(rw, buf, sqe_len, *iovec, iter); |
2864 | *iovec = NULL; | |
3a901598 | 2865 | return ret < 0 ? ret : sqe_len; |
3a6820f2 JA |
2866 | } |
2867 | ||
4d954c25 JA |
2868 | if (req->flags & REQ_F_BUFFER_SELECT) { |
2869 | ret = io_iov_buffer_select(req, *iovec, needs_lock); | |
3f9d6441 JA |
2870 | if (!ret) { |
2871 | ret = (*iovec)->iov_len; | |
2872 | iov_iter_init(iter, rw, *iovec, 1, ret); | |
2873 | } | |
4d954c25 JA |
2874 | *iovec = NULL; |
2875 | return ret; | |
2876 | } | |
2877 | ||
2b188cc1 | 2878 | #ifdef CONFIG_COMPAT |
cf6fd4bd | 2879 | if (req->ctx->compat) |
2b188cc1 JA |
2880 | return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV, |
2881 | iovec, iter); | |
2882 | #endif | |
2883 | ||
2884 | return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter); | |
2885 | } | |
2886 | ||
8452fd0c JA |
2887 | static ssize_t io_import_iovec(int rw, struct io_kiocb *req, |
2888 | struct iovec **iovec, struct iov_iter *iter, | |
2889 | bool needs_lock) | |
2890 | { | |
2891 | if (!req->io) | |
2892 | return __io_import_iovec(rw, req, iovec, iter, needs_lock); | |
2893 | *iovec = NULL; | |
2894 | return iov_iter_count(&req->io->rw.iter); | |
2895 | } | |
2896 | ||
0fef9483 JA |
2897 | static inline loff_t *io_kiocb_ppos(struct kiocb *kiocb) |
2898 | { | |
2899 | return kiocb->ki_filp->f_mode & FMODE_STREAM ? NULL : &kiocb->ki_pos; | |
2900 | } | |
2901 | ||
31b51510 | 2902 | /* |
32960613 JA |
2903 | * For files that don't have ->read_iter() and ->write_iter(), handle them |
2904 | * by looping over ->read() or ->write() manually. | |
31b51510 | 2905 | */ |
32960613 JA |
2906 | static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb, |
2907 | struct iov_iter *iter) | |
2908 | { | |
2909 | ssize_t ret = 0; | |
2910 | ||
2911 | /* | |
2912 | * Don't support polled IO through this interface, and we can't | |
2913 | * support non-blocking either. For the latter, this just causes | |
2914 | * the kiocb to be handled from an async context. | |
2915 | */ | |
2916 | if (kiocb->ki_flags & IOCB_HIPRI) | |
2917 | return -EOPNOTSUPP; | |
2918 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
2919 | return -EAGAIN; | |
2920 | ||
2921 | while (iov_iter_count(iter)) { | |
311ae9e1 | 2922 | struct iovec iovec; |
32960613 JA |
2923 | ssize_t nr; |
2924 | ||
311ae9e1 PB |
2925 | if (!iov_iter_is_bvec(iter)) { |
2926 | iovec = iov_iter_iovec(iter); | |
2927 | } else { | |
2928 | /* fixed buffers import bvec */ | |
2929 | iovec.iov_base = kmap(iter->bvec->bv_page) | |
2930 | + iter->iov_offset; | |
2931 | iovec.iov_len = min(iter->count, | |
2932 | iter->bvec->bv_len - iter->iov_offset); | |
2933 | } | |
2934 | ||
32960613 JA |
2935 | if (rw == READ) { |
2936 | nr = file->f_op->read(file, iovec.iov_base, | |
0fef9483 | 2937 | iovec.iov_len, io_kiocb_ppos(kiocb)); |
32960613 JA |
2938 | } else { |
2939 | nr = file->f_op->write(file, iovec.iov_base, | |
0fef9483 | 2940 | iovec.iov_len, io_kiocb_ppos(kiocb)); |
32960613 JA |
2941 | } |
2942 | ||
311ae9e1 PB |
2943 | if (iov_iter_is_bvec(iter)) |
2944 | kunmap(iter->bvec->bv_page); | |
2945 | ||
32960613 JA |
2946 | if (nr < 0) { |
2947 | if (!ret) | |
2948 | ret = nr; | |
2949 | break; | |
2950 | } | |
2951 | ret += nr; | |
2952 | if (nr != iovec.iov_len) | |
2953 | break; | |
2954 | iov_iter_advance(iter, nr); | |
2955 | } | |
2956 | ||
2957 | return ret; | |
2958 | } | |
2959 | ||
ff6165b2 JA |
2960 | static void io_req_map_rw(struct io_kiocb *req, const struct iovec *iovec, |
2961 | const struct iovec *fast_iov, struct iov_iter *iter) | |
f67676d1 | 2962 | { |
b64e3444 PB |
2963 | struct io_async_rw *rw = &req->io->rw; |
2964 | ||
ff6165b2 JA |
2965 | memcpy(&rw->iter, iter, sizeof(*iter)); |
2966 | rw->free_iovec = NULL; | |
227c0c96 | 2967 | rw->bytes_done = 0; |
ff6165b2 JA |
2968 | /* can only be fixed buffers, no need to do anything */ |
2969 | if (iter->type == ITER_BVEC) | |
2970 | return; | |
b64e3444 | 2971 | if (!iovec) { |
ff6165b2 JA |
2972 | unsigned iov_off = 0; |
2973 | ||
2974 | rw->iter.iov = rw->fast_iov; | |
2975 | if (iter->iov != fast_iov) { | |
2976 | iov_off = iter->iov - fast_iov; | |
2977 | rw->iter.iov += iov_off; | |
2978 | } | |
2979 | if (rw->fast_iov != fast_iov) | |
2980 | memcpy(rw->fast_iov + iov_off, fast_iov + iov_off, | |
45097dae | 2981 | sizeof(struct iovec) * iter->nr_segs); |
99bc4c38 | 2982 | } else { |
ff6165b2 | 2983 | rw->free_iovec = iovec; |
99bc4c38 | 2984 | req->flags |= REQ_F_NEED_CLEANUP; |
f67676d1 JA |
2985 | } |
2986 | } | |
2987 | ||
3d9932a8 XW |
2988 | static inline int __io_alloc_async_ctx(struct io_kiocb *req) |
2989 | { | |
2990 | req->io = kmalloc(sizeof(*req->io), GFP_KERNEL); | |
2991 | return req->io == NULL; | |
2992 | } | |
2993 | ||
b7bb4f7d | 2994 | static int io_alloc_async_ctx(struct io_kiocb *req) |
f67676d1 | 2995 | { |
d3656344 JA |
2996 | if (!io_op_defs[req->opcode].async_ctx) |
2997 | return 0; | |
3d9932a8 XW |
2998 | |
2999 | return __io_alloc_async_ctx(req); | |
b7bb4f7d JA |
3000 | } |
3001 | ||
ff6165b2 JA |
3002 | static int io_setup_async_rw(struct io_kiocb *req, const struct iovec *iovec, |
3003 | const struct iovec *fast_iov, | |
227c0c96 | 3004 | struct iov_iter *iter, bool force) |
b7bb4f7d | 3005 | { |
227c0c96 | 3006 | if (!force && !io_op_defs[req->opcode].async_ctx) |
74566df3 | 3007 | return 0; |
5d204bcf | 3008 | if (!req->io) { |
3d9932a8 | 3009 | if (__io_alloc_async_ctx(req)) |
5d204bcf | 3010 | return -ENOMEM; |
b7bb4f7d | 3011 | |
ff6165b2 | 3012 | io_req_map_rw(req, iovec, fast_iov, iter); |
5d204bcf | 3013 | } |
b7bb4f7d | 3014 | return 0; |
f67676d1 JA |
3015 | } |
3016 | ||
c3e330a4 PB |
3017 | static inline int io_rw_prep_async(struct io_kiocb *req, int rw, |
3018 | bool force_nonblock) | |
3019 | { | |
ff6165b2 | 3020 | struct io_async_rw *iorw = &req->io->rw; |
c183edff | 3021 | struct iovec *iov; |
c3e330a4 PB |
3022 | ssize_t ret; |
3023 | ||
c183edff JA |
3024 | iorw->iter.iov = iov = iorw->fast_iov; |
3025 | ret = __io_import_iovec(rw, req, &iov, &iorw->iter, !force_nonblock); | |
c3e330a4 PB |
3026 | if (unlikely(ret < 0)) |
3027 | return ret; | |
3028 | ||
c183edff | 3029 | iorw->iter.iov = iov; |
ff6165b2 | 3030 | io_req_map_rw(req, iorw->iter.iov, iorw->fast_iov, &iorw->iter); |
c3e330a4 PB |
3031 | return 0; |
3032 | } | |
3033 | ||
3529d8c2 JA |
3034 | static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
3035 | bool force_nonblock) | |
f67676d1 JA |
3036 | { |
3037 | ssize_t ret; | |
3038 | ||
3529d8c2 JA |
3039 | ret = io_prep_rw(req, sqe, force_nonblock); |
3040 | if (ret) | |
3041 | return ret; | |
f67676d1 | 3042 | |
3529d8c2 JA |
3043 | if (unlikely(!(req->file->f_mode & FMODE_READ))) |
3044 | return -EBADF; | |
f67676d1 | 3045 | |
5f798bea PB |
3046 | /* either don't need iovec imported or already have it */ |
3047 | if (!req->io || req->flags & REQ_F_NEED_CLEANUP) | |
3529d8c2 | 3048 | return 0; |
c3e330a4 | 3049 | return io_rw_prep_async(req, READ, force_nonblock); |
f67676d1 JA |
3050 | } |
3051 | ||
c1dd91d1 JA |
3052 | /* |
3053 | * This is our waitqueue callback handler, registered through lock_page_async() | |
3054 | * when we initially tried to do the IO with the iocb armed our waitqueue. | |
3055 | * This gets called when the page is unlocked, and we generally expect that to | |
3056 | * happen when the page IO is completed and the page is now uptodate. This will | |
3057 | * queue a task_work based retry of the operation, attempting to copy the data | |
3058 | * again. If the latter fails because the page was NOT uptodate, then we will | |
3059 | * do a thread based blocking retry of the operation. That's the unexpected | |
3060 | * slow path. | |
3061 | */ | |
bcf5a063 JA |
3062 | static int io_async_buf_func(struct wait_queue_entry *wait, unsigned mode, |
3063 | int sync, void *arg) | |
3064 | { | |
3065 | struct wait_page_queue *wpq; | |
3066 | struct io_kiocb *req = wait->private; | |
bcf5a063 | 3067 | struct wait_page_key *key = arg; |
bcf5a063 JA |
3068 | int ret; |
3069 | ||
3070 | wpq = container_of(wait, struct wait_page_queue, wait); | |
3071 | ||
cdc8fcb4 LT |
3072 | if (!wake_page_match(wpq, key)) |
3073 | return 0; | |
3074 | ||
c8d317aa | 3075 | req->rw.kiocb.ki_flags &= ~IOCB_WAITQ; |
bcf5a063 JA |
3076 | list_del_init(&wait->entry); |
3077 | ||
e7375122 | 3078 | init_task_work(&req->task_work, io_req_task_submit); |
6d816e08 JA |
3079 | percpu_ref_get(&req->ctx->refs); |
3080 | ||
bcf5a063 JA |
3081 | /* submit ref gets dropped, acquire a new one */ |
3082 | refcount_inc(&req->refs); | |
fd7d6de2 | 3083 | ret = io_req_task_work_add(req, &req->task_work, true); |
bcf5a063 | 3084 | if (unlikely(ret)) { |
c2c4c83c JA |
3085 | struct task_struct *tsk; |
3086 | ||
bcf5a063 | 3087 | /* queue just for cancelation */ |
e7375122 | 3088 | init_task_work(&req->task_work, io_req_task_cancel); |
bcf5a063 | 3089 | tsk = io_wq_get_task(req->ctx->io_wq); |
e7375122 | 3090 | task_work_add(tsk, &req->task_work, 0); |
c2c4c83c | 3091 | wake_up_process(tsk); |
bcf5a063 | 3092 | } |
bcf5a063 JA |
3093 | return 1; |
3094 | } | |
3095 | ||
c1dd91d1 JA |
3096 | /* |
3097 | * This controls whether a given IO request should be armed for async page | |
3098 | * based retry. If we return false here, the request is handed to the async | |
3099 | * worker threads for retry. If we're doing buffered reads on a regular file, | |
3100 | * we prepare a private wait_page_queue entry and retry the operation. This | |
3101 | * will either succeed because the page is now uptodate and unlocked, or it | |
3102 | * will register a callback when the page is unlocked at IO completion. Through | |
3103 | * that callback, io_uring uses task_work to setup a retry of the operation. | |
3104 | * That retry will attempt the buffered read again. The retry will generally | |
3105 | * succeed, or in rare cases where it fails, we then fall back to using the | |
3106 | * async worker threads for a blocking retry. | |
3107 | */ | |
227c0c96 | 3108 | static bool io_rw_should_retry(struct io_kiocb *req) |
f67676d1 | 3109 | { |
3b2a4439 | 3110 | struct wait_page_queue *wait = &req->io->rw.wpq; |
bcf5a063 | 3111 | struct kiocb *kiocb = &req->rw.kiocb; |
f67676d1 | 3112 | |
bcf5a063 JA |
3113 | /* never retry for NOWAIT, we just complete with -EAGAIN */ |
3114 | if (req->flags & REQ_F_NOWAIT) | |
3115 | return false; | |
f67676d1 | 3116 | |
227c0c96 | 3117 | /* Only for buffered IO */ |
3b2a4439 | 3118 | if (kiocb->ki_flags & (IOCB_DIRECT | IOCB_HIPRI)) |
bcf5a063 | 3119 | return false; |
3b2a4439 | 3120 | |
bcf5a063 JA |
3121 | /* |
3122 | * just use poll if we can, and don't attempt if the fs doesn't | |
3123 | * support callback based unlocks | |
3124 | */ | |
3125 | if (file_can_poll(req->file) || !(req->file->f_mode & FMODE_BUF_RASYNC)) | |
3126 | return false; | |
f67676d1 | 3127 | |
3b2a4439 JA |
3128 | wait->wait.func = io_async_buf_func; |
3129 | wait->wait.private = req; | |
3130 | wait->wait.flags = 0; | |
3131 | INIT_LIST_HEAD(&wait->wait.entry); | |
3132 | kiocb->ki_flags |= IOCB_WAITQ; | |
c8d317aa | 3133 | kiocb->ki_flags &= ~IOCB_NOWAIT; |
3b2a4439 | 3134 | kiocb->ki_waitq = wait; |
3529d8c2 | 3135 | |
3b2a4439 JA |
3136 | io_get_req_task(req); |
3137 | return true; | |
bcf5a063 JA |
3138 | } |
3139 | ||
3140 | static int io_iter_do_read(struct io_kiocb *req, struct iov_iter *iter) | |
3141 | { | |
3142 | if (req->file->f_op->read_iter) | |
3143 | return call_read_iter(req->file, &req->rw.kiocb, iter); | |
2dd2111d GH |
3144 | else if (req->file->f_op->read) |
3145 | return loop_rw_iter(READ, req->file, &req->rw.kiocb, iter); | |
3146 | else | |
3147 | return -EINVAL; | |
f67676d1 JA |
3148 | } |
3149 | ||
a1d7c393 JA |
3150 | static int io_read(struct io_kiocb *req, bool force_nonblock, |
3151 | struct io_comp_state *cs) | |
2b188cc1 JA |
3152 | { |
3153 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
9adbd45d | 3154 | struct kiocb *kiocb = &req->rw.kiocb; |
ff6165b2 | 3155 | struct iov_iter __iter, *iter = &__iter; |
227c0c96 | 3156 | ssize_t io_size, ret, ret2; |
31b51510 | 3157 | size_t iov_count; |
f5cac8b1 | 3158 | bool no_async; |
ff6165b2 JA |
3159 | |
3160 | if (req->io) | |
3161 | iter = &req->io->rw.iter; | |
2b188cc1 | 3162 | |
ff6165b2 | 3163 | ret = io_import_iovec(READ, req, &iovec, iter, !force_nonblock); |
06b76d44 JA |
3164 | if (ret < 0) |
3165 | return ret; | |
eefdf30f | 3166 | iov_count = iov_iter_count(iter); |
fa15bafb PB |
3167 | io_size = ret; |
3168 | req->result = io_size; | |
227c0c96 | 3169 | ret = 0; |
2b188cc1 | 3170 | |
fd6c2e4c JA |
3171 | /* Ensure we clear previously set non-block flag */ |
3172 | if (!force_nonblock) | |
29de5f6a | 3173 | kiocb->ki_flags &= ~IOCB_NOWAIT; |
fd6c2e4c | 3174 | |
24c74678 | 3175 | /* If the file doesn't support async, just async punt */ |
f5cac8b1 JA |
3176 | no_async = force_nonblock && !io_file_supports_async(req->file, READ); |
3177 | if (no_async) | |
f67676d1 | 3178 | goto copy_iov; |
9e645e11 | 3179 | |
0fef9483 | 3180 | ret = rw_verify_area(READ, req->file, io_kiocb_ppos(kiocb), iov_count); |
fa15bafb PB |
3181 | if (unlikely(ret)) |
3182 | goto out_free; | |
2b188cc1 | 3183 | |
227c0c96 | 3184 | ret = io_iter_do_read(req, iter); |
32960613 | 3185 | |
227c0c96 JA |
3186 | if (!ret) { |
3187 | goto done; | |
3188 | } else if (ret == -EIOCBQUEUED) { | |
3189 | ret = 0; | |
3190 | goto out_free; | |
3191 | } else if (ret == -EAGAIN) { | |
eefdf30f JA |
3192 | /* IOPOLL retry should happen for io-wq threads */ |
3193 | if (!force_nonblock && !(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
f91daf56 | 3194 | goto done; |
355afaeb JA |
3195 | /* no retry on NONBLOCK marked file */ |
3196 | if (req->file->f_flags & O_NONBLOCK) | |
3197 | goto done; | |
84216315 JA |
3198 | /* some cases will consume bytes even on error returns */ |
3199 | iov_iter_revert(iter, iov_count - iov_iter_count(iter)); | |
f38c7e3a JA |
3200 | ret = 0; |
3201 | goto copy_iov; | |
227c0c96 | 3202 | } else if (ret < 0) { |
00d23d51 JA |
3203 | /* make sure -ERESTARTSYS -> -EINTR is done */ |
3204 | goto done; | |
227c0c96 JA |
3205 | } |
3206 | ||
3207 | /* read it all, or we did blocking attempt. no retry. */ | |
f91daf56 JA |
3208 | if (!iov_iter_count(iter) || !force_nonblock || |
3209 | (req->file->f_flags & O_NONBLOCK)) | |
227c0c96 JA |
3210 | goto done; |
3211 | ||
3212 | io_size -= ret; | |
3213 | copy_iov: | |
3214 | ret2 = io_setup_async_rw(req, iovec, inline_vecs, iter, true); | |
3215 | if (ret2) { | |
3216 | ret = ret2; | |
3217 | goto out_free; | |
3218 | } | |
f5cac8b1 JA |
3219 | if (no_async) |
3220 | return -EAGAIN; | |
227c0c96 JA |
3221 | /* it's copied and will be cleaned with ->io */ |
3222 | iovec = NULL; | |
3223 | /* now use our persistent iterator, if we aren't already */ | |
3224 | iter = &req->io->rw.iter; | |
3225 | retry: | |
3226 | req->io->rw.bytes_done += ret; | |
3227 | /* if we can retry, do so with the callbacks armed */ | |
3228 | if (!io_rw_should_retry(req)) { | |
fa15bafb PB |
3229 | kiocb->ki_flags &= ~IOCB_WAITQ; |
3230 | return -EAGAIN; | |
2b188cc1 | 3231 | } |
227c0c96 JA |
3232 | |
3233 | /* | |
3234 | * Now retry read with the IOCB_WAITQ parts set in the iocb. If we | |
3235 | * get -EIOCBQUEUED, then we'll get a notification when the desired | |
3236 | * page gets unlocked. We can also get a partial read here, and if we | |
3237 | * do, then just retry at the new offset. | |
3238 | */ | |
3239 | ret = io_iter_do_read(req, iter); | |
3240 | if (ret == -EIOCBQUEUED) { | |
3241 | ret = 0; | |
3242 | goto out_free; | |
3243 | } else if (ret > 0 && ret < io_size) { | |
3244 | /* we got some bytes, but not all. retry. */ | |
3245 | goto retry; | |
3246 | } | |
3247 | done: | |
3248 | kiocb_done(kiocb, ret, cs); | |
3249 | ret = 0; | |
f67676d1 | 3250 | out_free: |
f261c168 | 3251 | /* it's reportedly faster than delegating the null check to kfree() */ |
252917c3 | 3252 | if (iovec) |
6f2cc166 | 3253 | kfree(iovec); |
2b188cc1 JA |
3254 | return ret; |
3255 | } | |
3256 | ||
3529d8c2 JA |
3257 | static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
3258 | bool force_nonblock) | |
f67676d1 JA |
3259 | { |
3260 | ssize_t ret; | |
3261 | ||
3529d8c2 JA |
3262 | ret = io_prep_rw(req, sqe, force_nonblock); |
3263 | if (ret) | |
3264 | return ret; | |
f67676d1 | 3265 | |
3529d8c2 JA |
3266 | if (unlikely(!(req->file->f_mode & FMODE_WRITE))) |
3267 | return -EBADF; | |
f67676d1 | 3268 | |
5f798bea PB |
3269 | /* either don't need iovec imported or already have it */ |
3270 | if (!req->io || req->flags & REQ_F_NEED_CLEANUP) | |
3529d8c2 | 3271 | return 0; |
c3e330a4 | 3272 | return io_rw_prep_async(req, WRITE, force_nonblock); |
f67676d1 JA |
3273 | } |
3274 | ||
a1d7c393 JA |
3275 | static int io_write(struct io_kiocb *req, bool force_nonblock, |
3276 | struct io_comp_state *cs) | |
2b188cc1 JA |
3277 | { |
3278 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
9adbd45d | 3279 | struct kiocb *kiocb = &req->rw.kiocb; |
ff6165b2 | 3280 | struct iov_iter __iter, *iter = &__iter; |
31b51510 | 3281 | size_t iov_count; |
fa15bafb | 3282 | ssize_t ret, ret2, io_size; |
2b188cc1 | 3283 | |
ff6165b2 JA |
3284 | if (req->io) |
3285 | iter = &req->io->rw.iter; | |
3286 | ||
3287 | ret = io_import_iovec(WRITE, req, &iovec, iter, !force_nonblock); | |
06b76d44 JA |
3288 | if (ret < 0) |
3289 | return ret; | |
eefdf30f | 3290 | iov_count = iov_iter_count(iter); |
fa15bafb PB |
3291 | io_size = ret; |
3292 | req->result = io_size; | |
2b188cc1 | 3293 | |
fd6c2e4c JA |
3294 | /* Ensure we clear previously set non-block flag */ |
3295 | if (!force_nonblock) | |
9adbd45d | 3296 | req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT; |
fd6c2e4c | 3297 | |
24c74678 | 3298 | /* If the file doesn't support async, just async punt */ |
af197f50 | 3299 | if (force_nonblock && !io_file_supports_async(req->file, WRITE)) |
f67676d1 | 3300 | goto copy_iov; |
31b51510 | 3301 | |
10d59345 JA |
3302 | /* file path doesn't support NOWAIT for non-direct_IO */ |
3303 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) && | |
3304 | (req->flags & REQ_F_ISREG)) | |
f67676d1 | 3305 | goto copy_iov; |
31b51510 | 3306 | |
0fef9483 | 3307 | ret = rw_verify_area(WRITE, req->file, io_kiocb_ppos(kiocb), iov_count); |
fa15bafb PB |
3308 | if (unlikely(ret)) |
3309 | goto out_free; | |
4ed734b0 | 3310 | |
fa15bafb PB |
3311 | /* |
3312 | * Open-code file_start_write here to grab freeze protection, | |
3313 | * which will be released by another thread in | |
3314 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
3315 | * released so that it doesn't complain about the held lock when | |
3316 | * we return to userspace. | |
3317 | */ | |
3318 | if (req->flags & REQ_F_ISREG) { | |
3319 | __sb_start_write(file_inode(req->file)->i_sb, | |
3320 | SB_FREEZE_WRITE, true); | |
3321 | __sb_writers_release(file_inode(req->file)->i_sb, | |
3322 | SB_FREEZE_WRITE); | |
3323 | } | |
3324 | kiocb->ki_flags |= IOCB_WRITE; | |
4ed734b0 | 3325 | |
fa15bafb | 3326 | if (req->file->f_op->write_iter) |
ff6165b2 | 3327 | ret2 = call_write_iter(req->file, kiocb, iter); |
2dd2111d | 3328 | else if (req->file->f_op->write) |
ff6165b2 | 3329 | ret2 = loop_rw_iter(WRITE, req->file, kiocb, iter); |
2dd2111d GH |
3330 | else |
3331 | ret2 = -EINVAL; | |
4ed734b0 | 3332 | |
fa15bafb PB |
3333 | /* |
3334 | * Raw bdev writes will return -EOPNOTSUPP for IOCB_NOWAIT. Just | |
3335 | * retry them without IOCB_NOWAIT. | |
3336 | */ | |
3337 | if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT)) | |
3338 | ret2 = -EAGAIN; | |
355afaeb JA |
3339 | /* no retry on NONBLOCK marked file */ |
3340 | if (ret2 == -EAGAIN && (req->file->f_flags & O_NONBLOCK)) | |
3341 | goto done; | |
fa15bafb | 3342 | if (!force_nonblock || ret2 != -EAGAIN) { |
eefdf30f JA |
3343 | /* IOPOLL retry should happen for io-wq threads */ |
3344 | if ((req->ctx->flags & IORING_SETUP_IOPOLL) && ret2 == -EAGAIN) | |
3345 | goto copy_iov; | |
355afaeb | 3346 | done: |
fa15bafb PB |
3347 | kiocb_done(kiocb, ret2, cs); |
3348 | } else { | |
f67676d1 | 3349 | copy_iov: |
84216315 JA |
3350 | /* some cases will consume bytes even on error returns */ |
3351 | iov_iter_revert(iter, iov_count - iov_iter_count(iter)); | |
227c0c96 | 3352 | ret = io_setup_async_rw(req, iovec, inline_vecs, iter, false); |
ff6165b2 JA |
3353 | if (!ret) |
3354 | return -EAGAIN; | |
2b188cc1 | 3355 | } |
31b51510 | 3356 | out_free: |
f261c168 | 3357 | /* it's reportedly faster than delegating the null check to kfree() */ |
252917c3 | 3358 | if (iovec) |
6f2cc166 | 3359 | kfree(iovec); |
2b188cc1 JA |
3360 | return ret; |
3361 | } | |
3362 | ||
f2a8d5c7 PB |
3363 | static int __io_splice_prep(struct io_kiocb *req, |
3364 | const struct io_uring_sqe *sqe) | |
7d67af2c PB |
3365 | { |
3366 | struct io_splice* sp = &req->splice; | |
3367 | unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL; | |
3368 | int ret; | |
3369 | ||
3370 | if (req->flags & REQ_F_NEED_CLEANUP) | |
3371 | return 0; | |
3232dd02 PB |
3372 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3373 | return -EINVAL; | |
7d67af2c PB |
3374 | |
3375 | sp->file_in = NULL; | |
7d67af2c PB |
3376 | sp->len = READ_ONCE(sqe->len); |
3377 | sp->flags = READ_ONCE(sqe->splice_flags); | |
3378 | ||
3379 | if (unlikely(sp->flags & ~valid_flags)) | |
3380 | return -EINVAL; | |
3381 | ||
3382 | ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in, | |
3383 | (sp->flags & SPLICE_F_FD_IN_FIXED)); | |
3384 | if (ret) | |
3385 | return ret; | |
3386 | req->flags |= REQ_F_NEED_CLEANUP; | |
3387 | ||
7cdaf587 XW |
3388 | if (!S_ISREG(file_inode(sp->file_in)->i_mode)) { |
3389 | /* | |
3390 | * Splice operation will be punted aync, and here need to | |
3391 | * modify io_wq_work.flags, so initialize io_wq_work firstly. | |
3392 | */ | |
3393 | io_req_init_async(req); | |
7d67af2c | 3394 | req->work.flags |= IO_WQ_WORK_UNBOUND; |
7cdaf587 | 3395 | } |
7d67af2c PB |
3396 | |
3397 | return 0; | |
3398 | } | |
3399 | ||
f2a8d5c7 PB |
3400 | static int io_tee_prep(struct io_kiocb *req, |
3401 | const struct io_uring_sqe *sqe) | |
3402 | { | |
3403 | if (READ_ONCE(sqe->splice_off_in) || READ_ONCE(sqe->off)) | |
3404 | return -EINVAL; | |
3405 | return __io_splice_prep(req, sqe); | |
3406 | } | |
3407 | ||
3408 | static int io_tee(struct io_kiocb *req, bool force_nonblock) | |
3409 | { | |
3410 | struct io_splice *sp = &req->splice; | |
3411 | struct file *in = sp->file_in; | |
3412 | struct file *out = sp->file_out; | |
3413 | unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED; | |
3414 | long ret = 0; | |
3415 | ||
3416 | if (force_nonblock) | |
3417 | return -EAGAIN; | |
3418 | if (sp->len) | |
3419 | ret = do_tee(in, out, sp->len, flags); | |
3420 | ||
3421 | io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED)); | |
3422 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
3423 | ||
f2a8d5c7 PB |
3424 | if (ret != sp->len) |
3425 | req_set_fail_links(req); | |
e1e16097 | 3426 | io_req_complete(req, ret); |
f2a8d5c7 PB |
3427 | return 0; |
3428 | } | |
3429 | ||
3430 | static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
3431 | { | |
3432 | struct io_splice* sp = &req->splice; | |
3433 | ||
3434 | sp->off_in = READ_ONCE(sqe->splice_off_in); | |
3435 | sp->off_out = READ_ONCE(sqe->off); | |
3436 | return __io_splice_prep(req, sqe); | |
3437 | } | |
3438 | ||
014db007 | 3439 | static int io_splice(struct io_kiocb *req, bool force_nonblock) |
7d67af2c PB |
3440 | { |
3441 | struct io_splice *sp = &req->splice; | |
3442 | struct file *in = sp->file_in; | |
3443 | struct file *out = sp->file_out; | |
3444 | unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED; | |
3445 | loff_t *poff_in, *poff_out; | |
c9687426 | 3446 | long ret = 0; |
7d67af2c | 3447 | |
2fb3e822 PB |
3448 | if (force_nonblock) |
3449 | return -EAGAIN; | |
7d67af2c PB |
3450 | |
3451 | poff_in = (sp->off_in == -1) ? NULL : &sp->off_in; | |
3452 | poff_out = (sp->off_out == -1) ? NULL : &sp->off_out; | |
c9687426 | 3453 | |
948a7749 | 3454 | if (sp->len) |
c9687426 | 3455 | ret = do_splice(in, poff_in, out, poff_out, sp->len, flags); |
7d67af2c PB |
3456 | |
3457 | io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED)); | |
3458 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
3459 | ||
7d67af2c PB |
3460 | if (ret != sp->len) |
3461 | req_set_fail_links(req); | |
e1e16097 | 3462 | io_req_complete(req, ret); |
7d67af2c PB |
3463 | return 0; |
3464 | } | |
3465 | ||
2b188cc1 JA |
3466 | /* |
3467 | * IORING_OP_NOP just posts a completion event, nothing else. | |
3468 | */ | |
229a7b63 | 3469 | static int io_nop(struct io_kiocb *req, struct io_comp_state *cs) |
2b188cc1 JA |
3470 | { |
3471 | struct io_ring_ctx *ctx = req->ctx; | |
2b188cc1 | 3472 | |
def596e9 JA |
3473 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
3474 | return -EINVAL; | |
3475 | ||
229a7b63 | 3476 | __io_req_complete(req, 0, 0, cs); |
2b188cc1 JA |
3477 | return 0; |
3478 | } | |
3479 | ||
3529d8c2 | 3480 | static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
c992fe29 | 3481 | { |
6b06314c | 3482 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 3483 | |
09bb8394 JA |
3484 | if (!req->file) |
3485 | return -EBADF; | |
c992fe29 | 3486 | |
6b06314c | 3487 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 3488 | return -EINVAL; |
edafccee | 3489 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
3490 | return -EINVAL; |
3491 | ||
8ed8d3c3 JA |
3492 | req->sync.flags = READ_ONCE(sqe->fsync_flags); |
3493 | if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC)) | |
3494 | return -EINVAL; | |
3495 | ||
3496 | req->sync.off = READ_ONCE(sqe->off); | |
3497 | req->sync.len = READ_ONCE(sqe->len); | |
c992fe29 CH |
3498 | return 0; |
3499 | } | |
3500 | ||
ac45abc0 | 3501 | static int io_fsync(struct io_kiocb *req, bool force_nonblock) |
8ed8d3c3 | 3502 | { |
8ed8d3c3 | 3503 | loff_t end = req->sync.off + req->sync.len; |
8ed8d3c3 JA |
3504 | int ret; |
3505 | ||
ac45abc0 PB |
3506 | /* fsync always requires a blocking context */ |
3507 | if (force_nonblock) | |
3508 | return -EAGAIN; | |
3509 | ||
9adbd45d | 3510 | ret = vfs_fsync_range(req->file, req->sync.off, |
8ed8d3c3 JA |
3511 | end > 0 ? end : LLONG_MAX, |
3512 | req->sync.flags & IORING_FSYNC_DATASYNC); | |
3513 | if (ret < 0) | |
3514 | req_set_fail_links(req); | |
e1e16097 | 3515 | io_req_complete(req, ret); |
c992fe29 CH |
3516 | return 0; |
3517 | } | |
3518 | ||
d63d1b5e JA |
3519 | static int io_fallocate_prep(struct io_kiocb *req, |
3520 | const struct io_uring_sqe *sqe) | |
3521 | { | |
3522 | if (sqe->ioprio || sqe->buf_index || sqe->rw_flags) | |
3523 | return -EINVAL; | |
3232dd02 PB |
3524 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3525 | return -EINVAL; | |
d63d1b5e JA |
3526 | |
3527 | req->sync.off = READ_ONCE(sqe->off); | |
3528 | req->sync.len = READ_ONCE(sqe->addr); | |
3529 | req->sync.mode = READ_ONCE(sqe->len); | |
3530 | return 0; | |
3531 | } | |
3532 | ||
014db007 | 3533 | static int io_fallocate(struct io_kiocb *req, bool force_nonblock) |
5d17b4a4 | 3534 | { |
ac45abc0 PB |
3535 | int ret; |
3536 | ||
d63d1b5e | 3537 | /* fallocate always requiring blocking context */ |
ac45abc0 | 3538 | if (force_nonblock) |
5d17b4a4 | 3539 | return -EAGAIN; |
ac45abc0 PB |
3540 | ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off, |
3541 | req->sync.len); | |
ac45abc0 PB |
3542 | if (ret < 0) |
3543 | req_set_fail_links(req); | |
e1e16097 | 3544 | io_req_complete(req, ret); |
5d17b4a4 JA |
3545 | return 0; |
3546 | } | |
3547 | ||
ec65fea5 | 3548 | static int __io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
b7bb4f7d | 3549 | { |
f8748881 | 3550 | const char __user *fname; |
15b71abe | 3551 | int ret; |
b7bb4f7d | 3552 | |
ec65fea5 | 3553 | if (unlikely(sqe->ioprio || sqe->buf_index)) |
15b71abe | 3554 | return -EINVAL; |
ec65fea5 | 3555 | if (unlikely(req->flags & REQ_F_FIXED_FILE)) |
cf3040ca | 3556 | return -EBADF; |
03b1230c | 3557 | |
ec65fea5 PB |
3558 | /* open.how should be already initialised */ |
3559 | if (!(req->open.how.flags & O_PATH) && force_o_largefile()) | |
08a1d26e | 3560 | req->open.how.flags |= O_LARGEFILE; |
3529d8c2 | 3561 | |
25e72d10 PB |
3562 | req->open.dfd = READ_ONCE(sqe->fd); |
3563 | fname = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
f8748881 | 3564 | req->open.filename = getname(fname); |
15b71abe JA |
3565 | if (IS_ERR(req->open.filename)) { |
3566 | ret = PTR_ERR(req->open.filename); | |
3567 | req->open.filename = NULL; | |
3568 | return ret; | |
3569 | } | |
4022e7af | 3570 | req->open.nofile = rlimit(RLIMIT_NOFILE); |
8fef80bf | 3571 | req->flags |= REQ_F_NEED_CLEANUP; |
15b71abe | 3572 | return 0; |
03b1230c JA |
3573 | } |
3574 | ||
ec65fea5 PB |
3575 | static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
3576 | { | |
3577 | u64 flags, mode; | |
3578 | ||
4eb8dded JA |
3579 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) |
3580 | return -EINVAL; | |
ec65fea5 PB |
3581 | if (req->flags & REQ_F_NEED_CLEANUP) |
3582 | return 0; | |
3583 | mode = READ_ONCE(sqe->len); | |
3584 | flags = READ_ONCE(sqe->open_flags); | |
3585 | req->open.how = build_open_how(flags, mode); | |
3586 | return __io_openat_prep(req, sqe); | |
3587 | } | |
3588 | ||
cebdb986 | 3589 | static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
aa1fa28f | 3590 | { |
cebdb986 | 3591 | struct open_how __user *how; |
cebdb986 | 3592 | size_t len; |
0fa03c62 JA |
3593 | int ret; |
3594 | ||
4eb8dded JA |
3595 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) |
3596 | return -EINVAL; | |
0bdbdd08 PB |
3597 | if (req->flags & REQ_F_NEED_CLEANUP) |
3598 | return 0; | |
cebdb986 JA |
3599 | how = u64_to_user_ptr(READ_ONCE(sqe->addr2)); |
3600 | len = READ_ONCE(sqe->len); | |
cebdb986 JA |
3601 | if (len < OPEN_HOW_SIZE_VER0) |
3602 | return -EINVAL; | |
3529d8c2 | 3603 | |
cebdb986 JA |
3604 | ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how, |
3605 | len); | |
3606 | if (ret) | |
3607 | return ret; | |
3529d8c2 | 3608 | |
ec65fea5 | 3609 | return __io_openat_prep(req, sqe); |
cebdb986 JA |
3610 | } |
3611 | ||
014db007 | 3612 | static int io_openat2(struct io_kiocb *req, bool force_nonblock) |
15b71abe JA |
3613 | { |
3614 | struct open_flags op; | |
15b71abe JA |
3615 | struct file *file; |
3616 | int ret; | |
3617 | ||
f86cd20c | 3618 | if (force_nonblock) |
15b71abe | 3619 | return -EAGAIN; |
15b71abe | 3620 | |
cebdb986 | 3621 | ret = build_open_flags(&req->open.how, &op); |
15b71abe JA |
3622 | if (ret) |
3623 | goto err; | |
3624 | ||
4022e7af | 3625 | ret = __get_unused_fd_flags(req->open.how.flags, req->open.nofile); |
15b71abe JA |
3626 | if (ret < 0) |
3627 | goto err; | |
3628 | ||
3629 | file = do_filp_open(req->open.dfd, req->open.filename, &op); | |
3630 | if (IS_ERR(file)) { | |
3631 | put_unused_fd(ret); | |
3632 | ret = PTR_ERR(file); | |
3633 | } else { | |
3634 | fsnotify_open(file); | |
3635 | fd_install(ret, file); | |
3636 | } | |
3637 | err: | |
3638 | putname(req->open.filename); | |
8fef80bf | 3639 | req->flags &= ~REQ_F_NEED_CLEANUP; |
15b71abe JA |
3640 | if (ret < 0) |
3641 | req_set_fail_links(req); | |
e1e16097 | 3642 | io_req_complete(req, ret); |
15b71abe JA |
3643 | return 0; |
3644 | } | |
3645 | ||
014db007 | 3646 | static int io_openat(struct io_kiocb *req, bool force_nonblock) |
cebdb986 | 3647 | { |
014db007 | 3648 | return io_openat2(req, force_nonblock); |
cebdb986 JA |
3649 | } |
3650 | ||
067524e9 JA |
3651 | static int io_remove_buffers_prep(struct io_kiocb *req, |
3652 | const struct io_uring_sqe *sqe) | |
3653 | { | |
3654 | struct io_provide_buf *p = &req->pbuf; | |
3655 | u64 tmp; | |
3656 | ||
3657 | if (sqe->ioprio || sqe->rw_flags || sqe->addr || sqe->len || sqe->off) | |
3658 | return -EINVAL; | |
3659 | ||
3660 | tmp = READ_ONCE(sqe->fd); | |
3661 | if (!tmp || tmp > USHRT_MAX) | |
3662 | return -EINVAL; | |
3663 | ||
3664 | memset(p, 0, sizeof(*p)); | |
3665 | p->nbufs = tmp; | |
3666 | p->bgid = READ_ONCE(sqe->buf_group); | |
3667 | return 0; | |
3668 | } | |
3669 | ||
3670 | static int __io_remove_buffers(struct io_ring_ctx *ctx, struct io_buffer *buf, | |
3671 | int bgid, unsigned nbufs) | |
3672 | { | |
3673 | unsigned i = 0; | |
3674 | ||
3675 | /* shouldn't happen */ | |
3676 | if (!nbufs) | |
3677 | return 0; | |
3678 | ||
3679 | /* the head kbuf is the list itself */ | |
3680 | while (!list_empty(&buf->list)) { | |
3681 | struct io_buffer *nxt; | |
3682 | ||
3683 | nxt = list_first_entry(&buf->list, struct io_buffer, list); | |
3684 | list_del(&nxt->list); | |
3685 | kfree(nxt); | |
3686 | if (++i == nbufs) | |
3687 | return i; | |
3688 | } | |
3689 | i++; | |
3690 | kfree(buf); | |
3691 | idr_remove(&ctx->io_buffer_idr, bgid); | |
3692 | ||
3693 | return i; | |
3694 | } | |
3695 | ||
229a7b63 JA |
3696 | static int io_remove_buffers(struct io_kiocb *req, bool force_nonblock, |
3697 | struct io_comp_state *cs) | |
067524e9 JA |
3698 | { |
3699 | struct io_provide_buf *p = &req->pbuf; | |
3700 | struct io_ring_ctx *ctx = req->ctx; | |
3701 | struct io_buffer *head; | |
3702 | int ret = 0; | |
3703 | ||
3704 | io_ring_submit_lock(ctx, !force_nonblock); | |
3705 | ||
3706 | lockdep_assert_held(&ctx->uring_lock); | |
3707 | ||
3708 | ret = -ENOENT; | |
3709 | head = idr_find(&ctx->io_buffer_idr, p->bgid); | |
3710 | if (head) | |
3711 | ret = __io_remove_buffers(ctx, head, p->bgid, p->nbufs); | |
3712 | ||
3713 | io_ring_submit_lock(ctx, !force_nonblock); | |
3714 | if (ret < 0) | |
3715 | req_set_fail_links(req); | |
229a7b63 | 3716 | __io_req_complete(req, ret, 0, cs); |
067524e9 JA |
3717 | return 0; |
3718 | } | |
3719 | ||
ddf0322d JA |
3720 | static int io_provide_buffers_prep(struct io_kiocb *req, |
3721 | const struct io_uring_sqe *sqe) | |
3722 | { | |
3723 | struct io_provide_buf *p = &req->pbuf; | |
3724 | u64 tmp; | |
3725 | ||
3726 | if (sqe->ioprio || sqe->rw_flags) | |
3727 | return -EINVAL; | |
3728 | ||
3729 | tmp = READ_ONCE(sqe->fd); | |
3730 | if (!tmp || tmp > USHRT_MAX) | |
3731 | return -E2BIG; | |
3732 | p->nbufs = tmp; | |
3733 | p->addr = READ_ONCE(sqe->addr); | |
3734 | p->len = READ_ONCE(sqe->len); | |
3735 | ||
efe68c1c | 3736 | if (!access_ok(u64_to_user_ptr(p->addr), (p->len * p->nbufs))) |
ddf0322d JA |
3737 | return -EFAULT; |
3738 | ||
3739 | p->bgid = READ_ONCE(sqe->buf_group); | |
3740 | tmp = READ_ONCE(sqe->off); | |
3741 | if (tmp > USHRT_MAX) | |
3742 | return -E2BIG; | |
3743 | p->bid = tmp; | |
3744 | return 0; | |
3745 | } | |
3746 | ||
3747 | static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head) | |
3748 | { | |
3749 | struct io_buffer *buf; | |
3750 | u64 addr = pbuf->addr; | |
3751 | int i, bid = pbuf->bid; | |
3752 | ||
3753 | for (i = 0; i < pbuf->nbufs; i++) { | |
3754 | buf = kmalloc(sizeof(*buf), GFP_KERNEL); | |
3755 | if (!buf) | |
3756 | break; | |
3757 | ||
3758 | buf->addr = addr; | |
3759 | buf->len = pbuf->len; | |
3760 | buf->bid = bid; | |
3761 | addr += pbuf->len; | |
3762 | bid++; | |
3763 | if (!*head) { | |
3764 | INIT_LIST_HEAD(&buf->list); | |
3765 | *head = buf; | |
3766 | } else { | |
3767 | list_add_tail(&buf->list, &(*head)->list); | |
3768 | } | |
3769 | } | |
3770 | ||
3771 | return i ? i : -ENOMEM; | |
3772 | } | |
3773 | ||
229a7b63 JA |
3774 | static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock, |
3775 | struct io_comp_state *cs) | |
ddf0322d JA |
3776 | { |
3777 | struct io_provide_buf *p = &req->pbuf; | |
3778 | struct io_ring_ctx *ctx = req->ctx; | |
3779 | struct io_buffer *head, *list; | |
3780 | int ret = 0; | |
3781 | ||
3782 | io_ring_submit_lock(ctx, !force_nonblock); | |
3783 | ||
3784 | lockdep_assert_held(&ctx->uring_lock); | |
3785 | ||
3786 | list = head = idr_find(&ctx->io_buffer_idr, p->bgid); | |
3787 | ||
3788 | ret = io_add_buffers(p, &head); | |
3789 | if (ret < 0) | |
3790 | goto out; | |
3791 | ||
3792 | if (!list) { | |
3793 | ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1, | |
3794 | GFP_KERNEL); | |
3795 | if (ret < 0) { | |
067524e9 | 3796 | __io_remove_buffers(ctx, head, p->bgid, -1U); |
ddf0322d JA |
3797 | goto out; |
3798 | } | |
3799 | } | |
3800 | out: | |
3801 | io_ring_submit_unlock(ctx, !force_nonblock); | |
3802 | if (ret < 0) | |
3803 | req_set_fail_links(req); | |
229a7b63 | 3804 | __io_req_complete(req, ret, 0, cs); |
ddf0322d | 3805 | return 0; |
cebdb986 JA |
3806 | } |
3807 | ||
3e4827b0 JA |
3808 | static int io_epoll_ctl_prep(struct io_kiocb *req, |
3809 | const struct io_uring_sqe *sqe) | |
3810 | { | |
3811 | #if defined(CONFIG_EPOLL) | |
3812 | if (sqe->ioprio || sqe->buf_index) | |
3813 | return -EINVAL; | |
6ca56f84 | 3814 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL))) |
3232dd02 | 3815 | return -EINVAL; |
3e4827b0 JA |
3816 | |
3817 | req->epoll.epfd = READ_ONCE(sqe->fd); | |
3818 | req->epoll.op = READ_ONCE(sqe->len); | |
3819 | req->epoll.fd = READ_ONCE(sqe->off); | |
3820 | ||
3821 | if (ep_op_has_event(req->epoll.op)) { | |
3822 | struct epoll_event __user *ev; | |
3823 | ||
3824 | ev = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
3825 | if (copy_from_user(&req->epoll.event, ev, sizeof(*ev))) | |
3826 | return -EFAULT; | |
3827 | } | |
3828 | ||
3829 | return 0; | |
3830 | #else | |
3831 | return -EOPNOTSUPP; | |
3832 | #endif | |
3833 | } | |
3834 | ||
229a7b63 JA |
3835 | static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock, |
3836 | struct io_comp_state *cs) | |
3e4827b0 JA |
3837 | { |
3838 | #if defined(CONFIG_EPOLL) | |
3839 | struct io_epoll *ie = &req->epoll; | |
3840 | int ret; | |
3841 | ||
3842 | ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock); | |
3843 | if (force_nonblock && ret == -EAGAIN) | |
3844 | return -EAGAIN; | |
3845 | ||
3846 | if (ret < 0) | |
3847 | req_set_fail_links(req); | |
229a7b63 | 3848 | __io_req_complete(req, ret, 0, cs); |
3e4827b0 JA |
3849 | return 0; |
3850 | #else | |
3851 | return -EOPNOTSUPP; | |
3852 | #endif | |
3853 | } | |
3854 | ||
c1ca757b JA |
3855 | static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
3856 | { | |
3857 | #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU) | |
3858 | if (sqe->ioprio || sqe->buf_index || sqe->off) | |
3859 | return -EINVAL; | |
3232dd02 PB |
3860 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3861 | return -EINVAL; | |
c1ca757b JA |
3862 | |
3863 | req->madvise.addr = READ_ONCE(sqe->addr); | |
3864 | req->madvise.len = READ_ONCE(sqe->len); | |
3865 | req->madvise.advice = READ_ONCE(sqe->fadvise_advice); | |
3866 | return 0; | |
3867 | #else | |
3868 | return -EOPNOTSUPP; | |
3869 | #endif | |
3870 | } | |
3871 | ||
014db007 | 3872 | static int io_madvise(struct io_kiocb *req, bool force_nonblock) |
c1ca757b JA |
3873 | { |
3874 | #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU) | |
3875 | struct io_madvise *ma = &req->madvise; | |
3876 | int ret; | |
3877 | ||
3878 | if (force_nonblock) | |
3879 | return -EAGAIN; | |
3880 | ||
3881 | ret = do_madvise(ma->addr, ma->len, ma->advice); | |
3882 | if (ret < 0) | |
3883 | req_set_fail_links(req); | |
e1e16097 | 3884 | io_req_complete(req, ret); |
c1ca757b JA |
3885 | return 0; |
3886 | #else | |
3887 | return -EOPNOTSUPP; | |
3888 | #endif | |
3889 | } | |
3890 | ||
4840e418 JA |
3891 | static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
3892 | { | |
3893 | if (sqe->ioprio || sqe->buf_index || sqe->addr) | |
3894 | return -EINVAL; | |
3232dd02 PB |
3895 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3896 | return -EINVAL; | |
4840e418 JA |
3897 | |
3898 | req->fadvise.offset = READ_ONCE(sqe->off); | |
3899 | req->fadvise.len = READ_ONCE(sqe->len); | |
3900 | req->fadvise.advice = READ_ONCE(sqe->fadvise_advice); | |
3901 | return 0; | |
3902 | } | |
3903 | ||
014db007 | 3904 | static int io_fadvise(struct io_kiocb *req, bool force_nonblock) |
4840e418 JA |
3905 | { |
3906 | struct io_fadvise *fa = &req->fadvise; | |
3907 | int ret; | |
3908 | ||
3e69426d JA |
3909 | if (force_nonblock) { |
3910 | switch (fa->advice) { | |
3911 | case POSIX_FADV_NORMAL: | |
3912 | case POSIX_FADV_RANDOM: | |
3913 | case POSIX_FADV_SEQUENTIAL: | |
3914 | break; | |
3915 | default: | |
3916 | return -EAGAIN; | |
3917 | } | |
3918 | } | |
4840e418 JA |
3919 | |
3920 | ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice); | |
3921 | if (ret < 0) | |
3922 | req_set_fail_links(req); | |
e1e16097 | 3923 | io_req_complete(req, ret); |
4840e418 JA |
3924 | return 0; |
3925 | } | |
3926 | ||
eddc7ef5 JA |
3927 | static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
3928 | { | |
6ca56f84 | 3929 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL))) |
3232dd02 | 3930 | return -EINVAL; |
eddc7ef5 JA |
3931 | if (sqe->ioprio || sqe->buf_index) |
3932 | return -EINVAL; | |
9c280f90 | 3933 | if (req->flags & REQ_F_FIXED_FILE) |
cf3040ca | 3934 | return -EBADF; |
eddc7ef5 | 3935 | |
1d9e1288 BM |
3936 | req->statx.dfd = READ_ONCE(sqe->fd); |
3937 | req->statx.mask = READ_ONCE(sqe->len); | |
e62753e4 | 3938 | req->statx.filename = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
1d9e1288 BM |
3939 | req->statx.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2)); |
3940 | req->statx.flags = READ_ONCE(sqe->statx_flags); | |
eddc7ef5 JA |
3941 | |
3942 | return 0; | |
3943 | } | |
3944 | ||
014db007 | 3945 | static int io_statx(struct io_kiocb *req, bool force_nonblock) |
eddc7ef5 | 3946 | { |
1d9e1288 | 3947 | struct io_statx *ctx = &req->statx; |
eddc7ef5 JA |
3948 | int ret; |
3949 | ||
5b0bbee4 JA |
3950 | if (force_nonblock) { |
3951 | /* only need file table for an actual valid fd */ | |
3952 | if (ctx->dfd == -1 || ctx->dfd == AT_FDCWD) | |
3953 | req->flags |= REQ_F_NO_FILE_TABLE; | |
eddc7ef5 | 3954 | return -EAGAIN; |
5b0bbee4 | 3955 | } |
eddc7ef5 | 3956 | |
e62753e4 BM |
3957 | ret = do_statx(ctx->dfd, ctx->filename, ctx->flags, ctx->mask, |
3958 | ctx->buffer); | |
eddc7ef5 | 3959 | |
eddc7ef5 JA |
3960 | if (ret < 0) |
3961 | req_set_fail_links(req); | |
e1e16097 | 3962 | io_req_complete(req, ret); |
eddc7ef5 JA |
3963 | return 0; |
3964 | } | |
3965 | ||
b5dba59e JA |
3966 | static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
3967 | { | |
3968 | /* | |
3969 | * If we queue this for async, it must not be cancellable. That would | |
7cdaf587 XW |
3970 | * leave the 'file' in an undeterminate state, and here need to modify |
3971 | * io_wq_work.flags, so initialize io_wq_work firstly. | |
b5dba59e | 3972 | */ |
7cdaf587 | 3973 | io_req_init_async(req); |
b5dba59e JA |
3974 | req->work.flags |= IO_WQ_WORK_NO_CANCEL; |
3975 | ||
3232dd02 PB |
3976 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) |
3977 | return -EINVAL; | |
b5dba59e JA |
3978 | if (sqe->ioprio || sqe->off || sqe->addr || sqe->len || |
3979 | sqe->rw_flags || sqe->buf_index) | |
3980 | return -EINVAL; | |
9c280f90 | 3981 | if (req->flags & REQ_F_FIXED_FILE) |
cf3040ca | 3982 | return -EBADF; |
b5dba59e JA |
3983 | |
3984 | req->close.fd = READ_ONCE(sqe->fd); | |
fd2206e4 JA |
3985 | if ((req->file && req->file->f_op == &io_uring_fops) || |
3986 | req->close.fd == req->ctx->ring_fd) | |
3987 | return -EBADF; | |
b5dba59e | 3988 | |
3af73b28 | 3989 | req->close.put_file = NULL; |
b5dba59e | 3990 | return 0; |
b5dba59e JA |
3991 | } |
3992 | ||
229a7b63 JA |
3993 | static int io_close(struct io_kiocb *req, bool force_nonblock, |
3994 | struct io_comp_state *cs) | |
b5dba59e | 3995 | { |
3af73b28 | 3996 | struct io_close *close = &req->close; |
b5dba59e JA |
3997 | int ret; |
3998 | ||
3af73b28 PB |
3999 | /* might be already done during nonblock submission */ |
4000 | if (!close->put_file) { | |
4001 | ret = __close_fd_get_file(close->fd, &close->put_file); | |
4002 | if (ret < 0) | |
4003 | return (ret == -ENOENT) ? -EBADF : ret; | |
4004 | } | |
b5dba59e JA |
4005 | |
4006 | /* if the file has a flush method, be safe and punt to async */ | |
3af73b28 | 4007 | if (close->put_file->f_op->flush && force_nonblock) { |
24c74678 PB |
4008 | /* was never set, but play safe */ |
4009 | req->flags &= ~REQ_F_NOWAIT; | |
0bf0eefd | 4010 | /* avoid grabbing files - we don't need the files */ |
24c74678 | 4011 | req->flags |= REQ_F_NO_FILE_TABLE; |
0bf0eefd | 4012 | return -EAGAIN; |
a2100672 | 4013 | } |
b5dba59e | 4014 | |
3af73b28 PB |
4015 | /* No ->flush() or already async, safely close from here */ |
4016 | ret = filp_close(close->put_file, req->work.files); | |
4017 | if (ret < 0) | |
4018 | req_set_fail_links(req); | |
3af73b28 PB |
4019 | fput(close->put_file); |
4020 | close->put_file = NULL; | |
229a7b63 | 4021 | __io_req_complete(req, ret, 0, cs); |
1a417f4e | 4022 | return 0; |
b5dba59e JA |
4023 | } |
4024 | ||
3529d8c2 | 4025 | static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
5d17b4a4 JA |
4026 | { |
4027 | struct io_ring_ctx *ctx = req->ctx; | |
5d17b4a4 JA |
4028 | |
4029 | if (!req->file) | |
4030 | return -EBADF; | |
5d17b4a4 JA |
4031 | |
4032 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
4033 | return -EINVAL; | |
4034 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
4035 | return -EINVAL; | |
4036 | ||
8ed8d3c3 JA |
4037 | req->sync.off = READ_ONCE(sqe->off); |
4038 | req->sync.len = READ_ONCE(sqe->len); | |
4039 | req->sync.flags = READ_ONCE(sqe->sync_range_flags); | |
8ed8d3c3 JA |
4040 | return 0; |
4041 | } | |
4042 | ||
ac45abc0 | 4043 | static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock) |
8ed8d3c3 | 4044 | { |
8ed8d3c3 JA |
4045 | int ret; |
4046 | ||
ac45abc0 PB |
4047 | /* sync_file_range always requires a blocking context */ |
4048 | if (force_nonblock) | |
4049 | return -EAGAIN; | |
4050 | ||
9adbd45d | 4051 | ret = sync_file_range(req->file, req->sync.off, req->sync.len, |
8ed8d3c3 JA |
4052 | req->sync.flags); |
4053 | if (ret < 0) | |
4054 | req_set_fail_links(req); | |
e1e16097 | 4055 | io_req_complete(req, ret); |
5d17b4a4 JA |
4056 | return 0; |
4057 | } | |
4058 | ||
469956e8 | 4059 | #if defined(CONFIG_NET) |
02d27d89 PB |
4060 | static int io_setup_async_msg(struct io_kiocb *req, |
4061 | struct io_async_msghdr *kmsg) | |
4062 | { | |
4063 | if (req->io) | |
4064 | return -EAGAIN; | |
4065 | if (io_alloc_async_ctx(req)) { | |
4066 | if (kmsg->iov != kmsg->fast_iov) | |
4067 | kfree(kmsg->iov); | |
4068 | return -ENOMEM; | |
4069 | } | |
4070 | req->flags |= REQ_F_NEED_CLEANUP; | |
4071 | memcpy(&req->io->msg, kmsg, sizeof(*kmsg)); | |
4072 | return -EAGAIN; | |
4073 | } | |
4074 | ||
2ae523ed PB |
4075 | static int io_sendmsg_copy_hdr(struct io_kiocb *req, |
4076 | struct io_async_msghdr *iomsg) | |
4077 | { | |
4078 | iomsg->iov = iomsg->fast_iov; | |
4079 | iomsg->msg.msg_name = &iomsg->addr; | |
4080 | return sendmsg_copy_msghdr(&iomsg->msg, req->sr_msg.umsg, | |
4081 | req->sr_msg.msg_flags, &iomsg->iov); | |
4082 | } | |
4083 | ||
3529d8c2 | 4084 | static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
03b1230c | 4085 | { |
e47293fd | 4086 | struct io_sr_msg *sr = &req->sr_msg; |
3529d8c2 | 4087 | struct io_async_ctx *io = req->io; |
99bc4c38 | 4088 | int ret; |
03b1230c | 4089 | |
d2b6f48b PB |
4090 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4091 | return -EINVAL; | |
4092 | ||
e47293fd | 4093 | sr->msg_flags = READ_ONCE(sqe->msg_flags); |
270a5940 | 4094 | sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
fddaface | 4095 | sr->len = READ_ONCE(sqe->len); |
3529d8c2 | 4096 | |
d8768362 JA |
4097 | #ifdef CONFIG_COMPAT |
4098 | if (req->ctx->compat) | |
4099 | sr->msg_flags |= MSG_CMSG_COMPAT; | |
4100 | #endif | |
4101 | ||
fddaface | 4102 | if (!io || req->opcode == IORING_OP_SEND) |
3529d8c2 | 4103 | return 0; |
5f798bea PB |
4104 | /* iovec is already imported */ |
4105 | if (req->flags & REQ_F_NEED_CLEANUP) | |
4106 | return 0; | |
3529d8c2 | 4107 | |
2ae523ed | 4108 | ret = io_sendmsg_copy_hdr(req, &io->msg); |
99bc4c38 PB |
4109 | if (!ret) |
4110 | req->flags |= REQ_F_NEED_CLEANUP; | |
4111 | return ret; | |
03b1230c JA |
4112 | } |
4113 | ||
229a7b63 JA |
4114 | static int io_sendmsg(struct io_kiocb *req, bool force_nonblock, |
4115 | struct io_comp_state *cs) | |
aa1fa28f | 4116 | { |
6b754c8b | 4117 | struct io_async_msghdr iomsg, *kmsg; |
0fa03c62 | 4118 | struct socket *sock; |
7a7cacba | 4119 | unsigned flags; |
0fa03c62 JA |
4120 | int ret; |
4121 | ||
0fa03c62 | 4122 | sock = sock_from_file(req->file, &ret); |
7a7cacba PB |
4123 | if (unlikely(!sock)) |
4124 | return ret; | |
3529d8c2 | 4125 | |
7a7cacba PB |
4126 | if (req->io) { |
4127 | kmsg = &req->io->msg; | |
4128 | kmsg->msg.msg_name = &req->io->msg.addr; | |
4129 | /* if iov is set, it's allocated already */ | |
4130 | if (!kmsg->iov) | |
4131 | kmsg->iov = kmsg->fast_iov; | |
4132 | kmsg->msg.msg_iter.iov = kmsg->iov; | |
4133 | } else { | |
4134 | ret = io_sendmsg_copy_hdr(req, &iomsg); | |
4135 | if (ret) | |
4136 | return ret; | |
4137 | kmsg = &iomsg; | |
0fa03c62 | 4138 | } |
0fa03c62 | 4139 | |
7a7cacba PB |
4140 | flags = req->sr_msg.msg_flags; |
4141 | if (flags & MSG_DONTWAIT) | |
4142 | req->flags |= REQ_F_NOWAIT; | |
4143 | else if (force_nonblock) | |
4144 | flags |= MSG_DONTWAIT; | |
e47293fd | 4145 | |
7a7cacba PB |
4146 | ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags); |
4147 | if (force_nonblock && ret == -EAGAIN) | |
4148 | return io_setup_async_msg(req, kmsg); | |
4149 | if (ret == -ERESTARTSYS) | |
4150 | ret = -EINTR; | |
0fa03c62 | 4151 | |
6b754c8b | 4152 | if (kmsg->iov != kmsg->fast_iov) |
0b416c3e | 4153 | kfree(kmsg->iov); |
99bc4c38 | 4154 | req->flags &= ~REQ_F_NEED_CLEANUP; |
4e88d6e7 JA |
4155 | if (ret < 0) |
4156 | req_set_fail_links(req); | |
229a7b63 | 4157 | __io_req_complete(req, ret, 0, cs); |
5d17b4a4 | 4158 | return 0; |
03b1230c | 4159 | } |
aa1fa28f | 4160 | |
229a7b63 JA |
4161 | static int io_send(struct io_kiocb *req, bool force_nonblock, |
4162 | struct io_comp_state *cs) | |
fddaface | 4163 | { |
7a7cacba PB |
4164 | struct io_sr_msg *sr = &req->sr_msg; |
4165 | struct msghdr msg; | |
4166 | struct iovec iov; | |
fddaface | 4167 | struct socket *sock; |
7a7cacba | 4168 | unsigned flags; |
fddaface JA |
4169 | int ret; |
4170 | ||
fddaface | 4171 | sock = sock_from_file(req->file, &ret); |
7a7cacba PB |
4172 | if (unlikely(!sock)) |
4173 | return ret; | |
fddaface | 4174 | |
7a7cacba PB |
4175 | ret = import_single_range(WRITE, sr->buf, sr->len, &iov, &msg.msg_iter); |
4176 | if (unlikely(ret)) | |
14c32eee | 4177 | return ret;; |
fddaface | 4178 | |
7a7cacba PB |
4179 | msg.msg_name = NULL; |
4180 | msg.msg_control = NULL; | |
4181 | msg.msg_controllen = 0; | |
4182 | msg.msg_namelen = 0; | |
fddaface | 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; | |
fddaface | 4189 | |
7a7cacba PB |
4190 | msg.msg_flags = flags; |
4191 | ret = sock_sendmsg(sock, &msg); | |
4192 | if (force_nonblock && ret == -EAGAIN) | |
4193 | return -EAGAIN; | |
4194 | if (ret == -ERESTARTSYS) | |
4195 | ret = -EINTR; | |
fddaface | 4196 | |
fddaface JA |
4197 | if (ret < 0) |
4198 | req_set_fail_links(req); | |
229a7b63 | 4199 | __io_req_complete(req, ret, 0, cs); |
fddaface | 4200 | return 0; |
fddaface JA |
4201 | } |
4202 | ||
1400e697 PB |
4203 | static int __io_recvmsg_copy_hdr(struct io_kiocb *req, |
4204 | struct io_async_msghdr *iomsg) | |
52de1fe1 JA |
4205 | { |
4206 | struct io_sr_msg *sr = &req->sr_msg; | |
4207 | struct iovec __user *uiov; | |
4208 | size_t iov_len; | |
4209 | int ret; | |
4210 | ||
1400e697 PB |
4211 | ret = __copy_msghdr_from_user(&iomsg->msg, sr->umsg, |
4212 | &iomsg->uaddr, &uiov, &iov_len); | |
52de1fe1 JA |
4213 | if (ret) |
4214 | return ret; | |
4215 | ||
4216 | if (req->flags & REQ_F_BUFFER_SELECT) { | |
4217 | if (iov_len > 1) | |
4218 | return -EINVAL; | |
1400e697 | 4219 | if (copy_from_user(iomsg->iov, uiov, sizeof(*uiov))) |
52de1fe1 | 4220 | return -EFAULT; |
1400e697 PB |
4221 | sr->len = iomsg->iov[0].iov_len; |
4222 | iov_iter_init(&iomsg->msg.msg_iter, READ, iomsg->iov, 1, | |
52de1fe1 | 4223 | sr->len); |
1400e697 | 4224 | iomsg->iov = NULL; |
52de1fe1 JA |
4225 | } else { |
4226 | ret = import_iovec(READ, uiov, iov_len, UIO_FASTIOV, | |
1400e697 | 4227 | &iomsg->iov, &iomsg->msg.msg_iter); |
52de1fe1 JA |
4228 | if (ret > 0) |
4229 | ret = 0; | |
4230 | } | |
4231 | ||
4232 | return ret; | |
4233 | } | |
4234 | ||
4235 | #ifdef CONFIG_COMPAT | |
4236 | static int __io_compat_recvmsg_copy_hdr(struct io_kiocb *req, | |
1400e697 | 4237 | struct io_async_msghdr *iomsg) |
52de1fe1 JA |
4238 | { |
4239 | struct compat_msghdr __user *msg_compat; | |
4240 | struct io_sr_msg *sr = &req->sr_msg; | |
4241 | struct compat_iovec __user *uiov; | |
4242 | compat_uptr_t ptr; | |
4243 | compat_size_t len; | |
4244 | int ret; | |
4245 | ||
270a5940 | 4246 | msg_compat = (struct compat_msghdr __user *) sr->umsg; |
1400e697 | 4247 | ret = __get_compat_msghdr(&iomsg->msg, msg_compat, &iomsg->uaddr, |
52de1fe1 JA |
4248 | &ptr, &len); |
4249 | if (ret) | |
4250 | return ret; | |
4251 | ||
4252 | uiov = compat_ptr(ptr); | |
4253 | if (req->flags & REQ_F_BUFFER_SELECT) { | |
4254 | compat_ssize_t clen; | |
4255 | ||
4256 | if (len > 1) | |
4257 | return -EINVAL; | |
4258 | if (!access_ok(uiov, sizeof(*uiov))) | |
4259 | return -EFAULT; | |
4260 | if (__get_user(clen, &uiov->iov_len)) | |
4261 | return -EFAULT; | |
4262 | if (clen < 0) | |
4263 | return -EINVAL; | |
1400e697 PB |
4264 | sr->len = iomsg->iov[0].iov_len; |
4265 | iomsg->iov = NULL; | |
52de1fe1 JA |
4266 | } else { |
4267 | ret = compat_import_iovec(READ, uiov, len, UIO_FASTIOV, | |
1400e697 PB |
4268 | &iomsg->iov, |
4269 | &iomsg->msg.msg_iter); | |
52de1fe1 JA |
4270 | if (ret < 0) |
4271 | return ret; | |
4272 | } | |
4273 | ||
4274 | return 0; | |
4275 | } | |
4276 | #endif | |
4277 | ||
1400e697 PB |
4278 | static int io_recvmsg_copy_hdr(struct io_kiocb *req, |
4279 | struct io_async_msghdr *iomsg) | |
52de1fe1 | 4280 | { |
1400e697 PB |
4281 | iomsg->msg.msg_name = &iomsg->addr; |
4282 | iomsg->iov = iomsg->fast_iov; | |
52de1fe1 JA |
4283 | |
4284 | #ifdef CONFIG_COMPAT | |
4285 | if (req->ctx->compat) | |
1400e697 | 4286 | return __io_compat_recvmsg_copy_hdr(req, iomsg); |
fddaface | 4287 | #endif |
52de1fe1 | 4288 | |
1400e697 | 4289 | return __io_recvmsg_copy_hdr(req, iomsg); |
52de1fe1 JA |
4290 | } |
4291 | ||
bcda7baa | 4292 | static struct io_buffer *io_recv_buffer_select(struct io_kiocb *req, |
7fbb1b54 | 4293 | bool needs_lock) |
bcda7baa JA |
4294 | { |
4295 | struct io_sr_msg *sr = &req->sr_msg; | |
4296 | struct io_buffer *kbuf; | |
4297 | ||
bcda7baa JA |
4298 | kbuf = io_buffer_select(req, &sr->len, sr->bgid, sr->kbuf, needs_lock); |
4299 | if (IS_ERR(kbuf)) | |
4300 | return kbuf; | |
4301 | ||
4302 | sr->kbuf = kbuf; | |
4303 | req->flags |= REQ_F_BUFFER_SELECTED; | |
bcda7baa | 4304 | return kbuf; |
fddaface JA |
4305 | } |
4306 | ||
7fbb1b54 PB |
4307 | static inline unsigned int io_put_recv_kbuf(struct io_kiocb *req) |
4308 | { | |
4309 | return io_put_kbuf(req, req->sr_msg.kbuf); | |
4310 | } | |
4311 | ||
3529d8c2 JA |
4312 | static int io_recvmsg_prep(struct io_kiocb *req, |
4313 | const struct io_uring_sqe *sqe) | |
aa1fa28f | 4314 | { |
e47293fd | 4315 | struct io_sr_msg *sr = &req->sr_msg; |
3529d8c2 | 4316 | struct io_async_ctx *io = req->io; |
99bc4c38 | 4317 | int ret; |
3529d8c2 | 4318 | |
d2b6f48b PB |
4319 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4320 | return -EINVAL; | |
4321 | ||
3529d8c2 | 4322 | sr->msg_flags = READ_ONCE(sqe->msg_flags); |
270a5940 | 4323 | sr->umsg = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
0b7b21e4 | 4324 | sr->len = READ_ONCE(sqe->len); |
bcda7baa | 4325 | sr->bgid = READ_ONCE(sqe->buf_group); |
06b76d44 | 4326 | |
d8768362 JA |
4327 | #ifdef CONFIG_COMPAT |
4328 | if (req->ctx->compat) | |
4329 | sr->msg_flags |= MSG_CMSG_COMPAT; | |
4330 | #endif | |
4331 | ||
fddaface | 4332 | if (!io || req->opcode == IORING_OP_RECV) |
06b76d44 | 4333 | return 0; |
5f798bea PB |
4334 | /* iovec is already imported */ |
4335 | if (req->flags & REQ_F_NEED_CLEANUP) | |
4336 | return 0; | |
03b1230c | 4337 | |
1400e697 | 4338 | ret = io_recvmsg_copy_hdr(req, &io->msg); |
99bc4c38 PB |
4339 | if (!ret) |
4340 | req->flags |= REQ_F_NEED_CLEANUP; | |
4341 | return ret; | |
aa1fa28f JA |
4342 | } |
4343 | ||
229a7b63 JA |
4344 | static int io_recvmsg(struct io_kiocb *req, bool force_nonblock, |
4345 | struct io_comp_state *cs) | |
aa1fa28f | 4346 | { |
6b754c8b | 4347 | struct io_async_msghdr iomsg, *kmsg; |
03b1230c | 4348 | struct socket *sock; |
7fbb1b54 | 4349 | struct io_buffer *kbuf; |
7a7cacba | 4350 | unsigned flags; |
52de1fe1 | 4351 | int ret, cflags = 0; |
03b1230c | 4352 | |
03b1230c | 4353 | sock = sock_from_file(req->file, &ret); |
7a7cacba PB |
4354 | if (unlikely(!sock)) |
4355 | return ret; | |
3529d8c2 | 4356 | |
7a7cacba PB |
4357 | if (req->io) { |
4358 | kmsg = &req->io->msg; | |
4359 | kmsg->msg.msg_name = &req->io->msg.addr; | |
4360 | /* if iov is set, it's allocated already */ | |
4361 | if (!kmsg->iov) | |
4362 | kmsg->iov = kmsg->fast_iov; | |
4363 | kmsg->msg.msg_iter.iov = kmsg->iov; | |
4364 | } else { | |
4365 | ret = io_recvmsg_copy_hdr(req, &iomsg); | |
4366 | if (ret) | |
681fda8d | 4367 | return ret; |
7a7cacba PB |
4368 | kmsg = &iomsg; |
4369 | } | |
03b1230c | 4370 | |
bc02ef33 | 4371 | if (req->flags & REQ_F_BUFFER_SELECT) { |
7fbb1b54 | 4372 | kbuf = io_recv_buffer_select(req, !force_nonblock); |
bc02ef33 | 4373 | if (IS_ERR(kbuf)) |
52de1fe1 | 4374 | return PTR_ERR(kbuf); |
7a7cacba PB |
4375 | kmsg->fast_iov[0].iov_base = u64_to_user_ptr(kbuf->addr); |
4376 | iov_iter_init(&kmsg->msg.msg_iter, READ, kmsg->iov, | |
4377 | 1, req->sr_msg.len); | |
4378 | } | |
52de1fe1 | 4379 | |
7a7cacba PB |
4380 | flags = req->sr_msg.msg_flags; |
4381 | if (flags & MSG_DONTWAIT) | |
4382 | req->flags |= REQ_F_NOWAIT; | |
4383 | else if (force_nonblock) | |
4384 | flags |= MSG_DONTWAIT; | |
e47293fd | 4385 | |
7a7cacba PB |
4386 | ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.umsg, |
4387 | kmsg->uaddr, flags); | |
0e1b6fe3 PB |
4388 | if (force_nonblock && ret == -EAGAIN) |
4389 | return io_setup_async_msg(req, kmsg); | |
7a7cacba PB |
4390 | if (ret == -ERESTARTSYS) |
4391 | ret = -EINTR; | |
03b1230c | 4392 | |
7fbb1b54 PB |
4393 | if (req->flags & REQ_F_BUFFER_SELECTED) |
4394 | cflags = io_put_recv_kbuf(req); | |
6b754c8b | 4395 | if (kmsg->iov != kmsg->fast_iov) |
0b416c3e | 4396 | kfree(kmsg->iov); |
99bc4c38 | 4397 | req->flags &= ~REQ_F_NEED_CLEANUP; |
4e88d6e7 JA |
4398 | if (ret < 0) |
4399 | req_set_fail_links(req); | |
229a7b63 | 4400 | __io_req_complete(req, ret, cflags, cs); |
03b1230c | 4401 | return 0; |
0fa03c62 | 4402 | } |
5d17b4a4 | 4403 | |
229a7b63 JA |
4404 | static int io_recv(struct io_kiocb *req, bool force_nonblock, |
4405 | struct io_comp_state *cs) | |
fddaface | 4406 | { |
6b754c8b | 4407 | struct io_buffer *kbuf; |
7a7cacba PB |
4408 | struct io_sr_msg *sr = &req->sr_msg; |
4409 | struct msghdr msg; | |
4410 | void __user *buf = sr->buf; | |
fddaface | 4411 | struct socket *sock; |
7a7cacba PB |
4412 | struct iovec iov; |
4413 | unsigned flags; | |
bcda7baa | 4414 | int ret, cflags = 0; |
fddaface | 4415 | |
fddaface | 4416 | sock = sock_from_file(req->file, &ret); |
7a7cacba PB |
4417 | if (unlikely(!sock)) |
4418 | return ret; | |
fddaface | 4419 | |
bc02ef33 | 4420 | if (req->flags & REQ_F_BUFFER_SELECT) { |
7fbb1b54 | 4421 | kbuf = io_recv_buffer_select(req, !force_nonblock); |
bcda7baa JA |
4422 | if (IS_ERR(kbuf)) |
4423 | return PTR_ERR(kbuf); | |
7a7cacba | 4424 | buf = u64_to_user_ptr(kbuf->addr); |
bc02ef33 | 4425 | } |
bcda7baa | 4426 | |
7a7cacba | 4427 | ret = import_single_range(READ, buf, sr->len, &iov, &msg.msg_iter); |
14c32eee PB |
4428 | if (unlikely(ret)) |
4429 | goto out_free; | |
fddaface | 4430 | |
7a7cacba PB |
4431 | msg.msg_name = NULL; |
4432 | msg.msg_control = NULL; | |
4433 | msg.msg_controllen = 0; | |
4434 | msg.msg_namelen = 0; | |
4435 | msg.msg_iocb = NULL; | |
4436 | msg.msg_flags = 0; | |
fddaface | 4437 | |
7a7cacba PB |
4438 | flags = req->sr_msg.msg_flags; |
4439 | if (flags & MSG_DONTWAIT) | |
4440 | req->flags |= REQ_F_NOWAIT; | |
4441 | else if (force_nonblock) | |
4442 | flags |= MSG_DONTWAIT; | |
4443 | ||
4444 | ret = sock_recvmsg(sock, &msg, flags); | |
4445 | if (force_nonblock && ret == -EAGAIN) | |
4446 | return -EAGAIN; | |
4447 | if (ret == -ERESTARTSYS) | |
4448 | ret = -EINTR; | |
14c32eee | 4449 | out_free: |
7fbb1b54 PB |
4450 | if (req->flags & REQ_F_BUFFER_SELECTED) |
4451 | cflags = io_put_recv_kbuf(req); | |
fddaface JA |
4452 | if (ret < 0) |
4453 | req_set_fail_links(req); | |
229a7b63 | 4454 | __io_req_complete(req, ret, cflags, cs); |
fddaface | 4455 | return 0; |
fddaface JA |
4456 | } |
4457 | ||
3529d8c2 | 4458 | static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
17f2fe35 | 4459 | { |
8ed8d3c3 JA |
4460 | struct io_accept *accept = &req->accept; |
4461 | ||
17f2fe35 JA |
4462 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) |
4463 | return -EINVAL; | |
8042d6ce | 4464 | if (sqe->ioprio || sqe->len || sqe->buf_index) |
17f2fe35 JA |
4465 | return -EINVAL; |
4466 | ||
d55e5f5b JA |
4467 | accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
4468 | accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2)); | |
8ed8d3c3 | 4469 | accept->flags = READ_ONCE(sqe->accept_flags); |
09952e3e | 4470 | accept->nofile = rlimit(RLIMIT_NOFILE); |
8ed8d3c3 | 4471 | return 0; |
8ed8d3c3 | 4472 | } |
17f2fe35 | 4473 | |
229a7b63 JA |
4474 | static int io_accept(struct io_kiocb *req, bool force_nonblock, |
4475 | struct io_comp_state *cs) | |
8ed8d3c3 JA |
4476 | { |
4477 | struct io_accept *accept = &req->accept; | |
ac45abc0 | 4478 | unsigned int file_flags = force_nonblock ? O_NONBLOCK : 0; |
8ed8d3c3 JA |
4479 | int ret; |
4480 | ||
e697deed JX |
4481 | if (req->file->f_flags & O_NONBLOCK) |
4482 | req->flags |= REQ_F_NOWAIT; | |
4483 | ||
8ed8d3c3 | 4484 | ret = __sys_accept4_file(req->file, file_flags, accept->addr, |
09952e3e JA |
4485 | accept->addr_len, accept->flags, |
4486 | accept->nofile); | |
8ed8d3c3 | 4487 | if (ret == -EAGAIN && force_nonblock) |
17f2fe35 | 4488 | return -EAGAIN; |
ac45abc0 PB |
4489 | if (ret < 0) { |
4490 | if (ret == -ERESTARTSYS) | |
4491 | ret = -EINTR; | |
4e88d6e7 | 4492 | req_set_fail_links(req); |
ac45abc0 | 4493 | } |
229a7b63 | 4494 | __io_req_complete(req, ret, 0, cs); |
17f2fe35 | 4495 | return 0; |
8ed8d3c3 JA |
4496 | } |
4497 | ||
3529d8c2 | 4498 | static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f499a021 | 4499 | { |
3529d8c2 JA |
4500 | struct io_connect *conn = &req->connect; |
4501 | struct io_async_ctx *io = req->io; | |
f499a021 | 4502 | |
3fbb51c1 JA |
4503 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) |
4504 | return -EINVAL; | |
4505 | if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags) | |
4506 | return -EINVAL; | |
4507 | ||
3529d8c2 JA |
4508 | conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
4509 | conn->addr_len = READ_ONCE(sqe->addr2); | |
4510 | ||
4511 | if (!io) | |
4512 | return 0; | |
4513 | ||
4514 | return move_addr_to_kernel(conn->addr, conn->addr_len, | |
3fbb51c1 | 4515 | &io->connect.address); |
f499a021 JA |
4516 | } |
4517 | ||
229a7b63 JA |
4518 | static int io_connect(struct io_kiocb *req, bool force_nonblock, |
4519 | struct io_comp_state *cs) | |
f8e85cf2 | 4520 | { |
f499a021 | 4521 | struct io_async_ctx __io, *io; |
f8e85cf2 | 4522 | unsigned file_flags; |
3fbb51c1 | 4523 | int ret; |
f8e85cf2 | 4524 | |
f499a021 JA |
4525 | if (req->io) { |
4526 | io = req->io; | |
4527 | } else { | |
3529d8c2 JA |
4528 | ret = move_addr_to_kernel(req->connect.addr, |
4529 | req->connect.addr_len, | |
4530 | &__io.connect.address); | |
f499a021 JA |
4531 | if (ret) |
4532 | goto out; | |
4533 | io = &__io; | |
4534 | } | |
4535 | ||
3fbb51c1 JA |
4536 | file_flags = force_nonblock ? O_NONBLOCK : 0; |
4537 | ||
4538 | ret = __sys_connect_file(req->file, &io->connect.address, | |
4539 | req->connect.addr_len, file_flags); | |
87f80d62 | 4540 | if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) { |
b7bb4f7d JA |
4541 | if (req->io) |
4542 | return -EAGAIN; | |
4543 | if (io_alloc_async_ctx(req)) { | |
f499a021 JA |
4544 | ret = -ENOMEM; |
4545 | goto out; | |
4546 | } | |
b7bb4f7d | 4547 | memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect)); |
f8e85cf2 | 4548 | return -EAGAIN; |
f499a021 | 4549 | } |
f8e85cf2 JA |
4550 | if (ret == -ERESTARTSYS) |
4551 | ret = -EINTR; | |
f499a021 | 4552 | out: |
4e88d6e7 JA |
4553 | if (ret < 0) |
4554 | req_set_fail_links(req); | |
229a7b63 | 4555 | __io_req_complete(req, ret, 0, cs); |
f8e85cf2 | 4556 | return 0; |
469956e8 Y |
4557 | } |
4558 | #else /* !CONFIG_NET */ | |
4559 | static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
4560 | { | |
f8e85cf2 | 4561 | return -EOPNOTSUPP; |
f8e85cf2 JA |
4562 | } |
4563 | ||
1e16c2f9 RD |
4564 | static int io_sendmsg(struct io_kiocb *req, bool force_nonblock, |
4565 | struct io_comp_state *cs) | |
469956e8 Y |
4566 | { |
4567 | return -EOPNOTSUPP; | |
4568 | } | |
4569 | ||
1e16c2f9 RD |
4570 | static int io_send(struct io_kiocb *req, bool force_nonblock, |
4571 | struct io_comp_state *cs) | |
469956e8 Y |
4572 | { |
4573 | return -EOPNOTSUPP; | |
4574 | } | |
4575 | ||
4576 | static int io_recvmsg_prep(struct io_kiocb *req, | |
4577 | const struct io_uring_sqe *sqe) | |
4578 | { | |
4579 | return -EOPNOTSUPP; | |
4580 | } | |
4581 | ||
1e16c2f9 RD |
4582 | static int io_recvmsg(struct io_kiocb *req, bool force_nonblock, |
4583 | struct io_comp_state *cs) | |
469956e8 Y |
4584 | { |
4585 | return -EOPNOTSUPP; | |
4586 | } | |
4587 | ||
1e16c2f9 RD |
4588 | static int io_recv(struct io_kiocb *req, bool force_nonblock, |
4589 | struct io_comp_state *cs) | |
469956e8 Y |
4590 | { |
4591 | return -EOPNOTSUPP; | |
4592 | } | |
4593 | ||
4594 | static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) | |
4595 | { | |
4596 | return -EOPNOTSUPP; | |
4597 | } | |
4598 | ||
1e16c2f9 RD |
4599 | static int io_accept(struct io_kiocb *req, bool force_nonblock, |
4600 | struct io_comp_state *cs) | |
469956e8 Y |
4601 | { |
4602 | return -EOPNOTSUPP; | |
4603 | } | |
ce593a6c | 4604 | |
469956e8 Y |
4605 | static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4606 | { | |
4607 | return -EOPNOTSUPP; | |
4608 | } | |
4609 | ||
1e16c2f9 RD |
4610 | static int io_connect(struct io_kiocb *req, bool force_nonblock, |
4611 | struct io_comp_state *cs) | |
469956e8 | 4612 | { |
f8e85cf2 | 4613 | return -EOPNOTSUPP; |
ce593a6c | 4614 | } |
469956e8 | 4615 | #endif /* CONFIG_NET */ |
f8e85cf2 | 4616 | |
d7718a9d JA |
4617 | struct io_poll_table { |
4618 | struct poll_table_struct pt; | |
4619 | struct io_kiocb *req; | |
4620 | int error; | |
4621 | }; | |
ce593a6c | 4622 | |
d7718a9d JA |
4623 | static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll, |
4624 | __poll_t mask, task_work_func_t func) | |
4625 | { | |
fd7d6de2 | 4626 | bool twa_signal_ok; |
aa96bf8a | 4627 | int ret; |
d7718a9d JA |
4628 | |
4629 | /* for instances that support it check for an event match first: */ | |
4630 | if (mask && !(mask & poll->events)) | |
4631 | return 0; | |
4632 | ||
4633 | trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask); | |
4634 | ||
4635 | list_del_init(&poll->wait.entry); | |
4636 | ||
d7718a9d JA |
4637 | req->result = mask; |
4638 | init_task_work(&req->task_work, func); | |
6d816e08 JA |
4639 | percpu_ref_get(&req->ctx->refs); |
4640 | ||
fd7d6de2 JA |
4641 | /* |
4642 | * If we using the signalfd wait_queue_head for this wakeup, then | |
4643 | * it's not safe to use TWA_SIGNAL as we could be recursing on the | |
4644 | * tsk->sighand->siglock on doing the wakeup. Should not be needed | |
4645 | * either, as the normal wakeup will suffice. | |
4646 | */ | |
4647 | twa_signal_ok = (poll->head != &req->task->sighand->signalfd_wqh); | |
4648 | ||
d7718a9d | 4649 | /* |
e3aabf95 JA |
4650 | * If this fails, then the task is exiting. When a task exits, the |
4651 | * work gets canceled, so just cancel this request as well instead | |
4652 | * of executing it. We can't safely execute it anyway, as we may not | |
4653 | * have the needed state needed for it anyway. | |
d7718a9d | 4654 | */ |
fd7d6de2 | 4655 | ret = io_req_task_work_add(req, &req->task_work, twa_signal_ok); |
aa96bf8a | 4656 | if (unlikely(ret)) { |
c2c4c83c JA |
4657 | struct task_struct *tsk; |
4658 | ||
e3aabf95 | 4659 | WRITE_ONCE(poll->canceled, true); |
aa96bf8a | 4660 | tsk = io_wq_get_task(req->ctx->io_wq); |
ce593a6c JA |
4661 | task_work_add(tsk, &req->task_work, 0); |
4662 | wake_up_process(tsk); | |
aa96bf8a | 4663 | } |
d7718a9d JA |
4664 | return 1; |
4665 | } | |
4666 | ||
74ce6ce4 JA |
4667 | static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll) |
4668 | __acquires(&req->ctx->completion_lock) | |
4669 | { | |
4670 | struct io_ring_ctx *ctx = req->ctx; | |
4671 | ||
4672 | if (!req->result && !READ_ONCE(poll->canceled)) { | |
4673 | struct poll_table_struct pt = { ._key = poll->events }; | |
4674 | ||
4675 | req->result = vfs_poll(req->file, &pt) & poll->events; | |
4676 | } | |
4677 | ||
4678 | spin_lock_irq(&ctx->completion_lock); | |
4679 | if (!req->result && !READ_ONCE(poll->canceled)) { | |
4680 | add_wait_queue(poll->head, &poll->wait); | |
4681 | return true; | |
4682 | } | |
4683 | ||
4684 | return false; | |
4685 | } | |
4686 | ||
d4e7cd36 | 4687 | static struct io_poll_iocb *io_poll_get_double(struct io_kiocb *req) |
18bceab1 | 4688 | { |
d4e7cd36 JA |
4689 | /* pure poll stashes this in ->io, poll driven retry elsewhere */ |
4690 | if (req->opcode == IORING_OP_POLL_ADD) | |
4691 | return (struct io_poll_iocb *) req->io; | |
4692 | return req->apoll->double_poll; | |
4693 | } | |
4694 | ||
4695 | static struct io_poll_iocb *io_poll_get_single(struct io_kiocb *req) | |
4696 | { | |
4697 | if (req->opcode == IORING_OP_POLL_ADD) | |
4698 | return &req->poll; | |
4699 | return &req->apoll->poll; | |
4700 | } | |
4701 | ||
4702 | static void io_poll_remove_double(struct io_kiocb *req) | |
4703 | { | |
4704 | struct io_poll_iocb *poll = io_poll_get_double(req); | |
18bceab1 JA |
4705 | |
4706 | lockdep_assert_held(&req->ctx->completion_lock); | |
4707 | ||
4708 | if (poll && poll->head) { | |
4709 | struct wait_queue_head *head = poll->head; | |
4710 | ||
4711 | spin_lock(&head->lock); | |
4712 | list_del_init(&poll->wait.entry); | |
4713 | if (poll->wait.private) | |
4714 | refcount_dec(&req->refs); | |
4715 | poll->head = NULL; | |
4716 | spin_unlock(&head->lock); | |
4717 | } | |
4718 | } | |
4719 | ||
4720 | static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error) | |
4721 | { | |
4722 | struct io_ring_ctx *ctx = req->ctx; | |
4723 | ||
d4e7cd36 | 4724 | io_poll_remove_double(req); |
18bceab1 JA |
4725 | req->poll.done = true; |
4726 | io_cqring_fill_event(req, error ? error : mangle_poll(mask)); | |
4727 | io_commit_cqring(ctx); | |
4728 | } | |
4729 | ||
4730 | static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt) | |
4731 | { | |
4732 | struct io_ring_ctx *ctx = req->ctx; | |
4733 | ||
4734 | if (io_poll_rewait(req, &req->poll)) { | |
4735 | spin_unlock_irq(&ctx->completion_lock); | |
4736 | return; | |
4737 | } | |
4738 | ||
4739 | hash_del(&req->hash_node); | |
4740 | io_poll_complete(req, req->result, 0); | |
4741 | req->flags |= REQ_F_COMP_LOCKED; | |
9b5f7bd9 | 4742 | *nxt = io_put_req_find_next(req); |
18bceab1 JA |
4743 | spin_unlock_irq(&ctx->completion_lock); |
4744 | ||
4745 | io_cqring_ev_posted(ctx); | |
4746 | } | |
4747 | ||
4748 | static void io_poll_task_func(struct callback_head *cb) | |
4749 | { | |
4750 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); | |
6d816e08 | 4751 | struct io_ring_ctx *ctx = req->ctx; |
18bceab1 JA |
4752 | struct io_kiocb *nxt = NULL; |
4753 | ||
4754 | io_poll_task_handler(req, &nxt); | |
ea1164e5 PB |
4755 | if (nxt) |
4756 | __io_req_task_submit(nxt); | |
6d816e08 | 4757 | percpu_ref_put(&ctx->refs); |
18bceab1 JA |
4758 | } |
4759 | ||
4760 | static int io_poll_double_wake(struct wait_queue_entry *wait, unsigned mode, | |
4761 | int sync, void *key) | |
4762 | { | |
4763 | struct io_kiocb *req = wait->private; | |
d4e7cd36 | 4764 | struct io_poll_iocb *poll = io_poll_get_single(req); |
18bceab1 JA |
4765 | __poll_t mask = key_to_poll(key); |
4766 | ||
4767 | /* for instances that support it check for an event match first: */ | |
4768 | if (mask && !(mask & poll->events)) | |
4769 | return 0; | |
4770 | ||
8706e04e JA |
4771 | list_del_init(&wait->entry); |
4772 | ||
807abcb0 | 4773 | if (poll && poll->head) { |
18bceab1 JA |
4774 | bool done; |
4775 | ||
807abcb0 JA |
4776 | spin_lock(&poll->head->lock); |
4777 | done = list_empty(&poll->wait.entry); | |
18bceab1 | 4778 | if (!done) |
807abcb0 | 4779 | list_del_init(&poll->wait.entry); |
d4e7cd36 JA |
4780 | /* make sure double remove sees this as being gone */ |
4781 | wait->private = NULL; | |
807abcb0 | 4782 | spin_unlock(&poll->head->lock); |
18bceab1 JA |
4783 | if (!done) |
4784 | __io_async_wake(req, poll, mask, io_poll_task_func); | |
4785 | } | |
4786 | refcount_dec(&req->refs); | |
4787 | return 1; | |
4788 | } | |
4789 | ||
4790 | static void io_init_poll_iocb(struct io_poll_iocb *poll, __poll_t events, | |
4791 | wait_queue_func_t wake_func) | |
4792 | { | |
4793 | poll->head = NULL; | |
4794 | poll->done = false; | |
4795 | poll->canceled = false; | |
4796 | poll->events = events; | |
4797 | INIT_LIST_HEAD(&poll->wait.entry); | |
4798 | init_waitqueue_func_entry(&poll->wait, wake_func); | |
4799 | } | |
4800 | ||
4801 | static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt, | |
807abcb0 JA |
4802 | struct wait_queue_head *head, |
4803 | struct io_poll_iocb **poll_ptr) | |
18bceab1 JA |
4804 | { |
4805 | struct io_kiocb *req = pt->req; | |
4806 | ||
4807 | /* | |
4808 | * If poll->head is already set, it's because the file being polled | |
4809 | * uses multiple waitqueues for poll handling (eg one for read, one | |
4810 | * for write). Setup a separate io_poll_iocb if this happens. | |
4811 | */ | |
4812 | if (unlikely(poll->head)) { | |
4813 | /* already have a 2nd entry, fail a third attempt */ | |
807abcb0 | 4814 | if (*poll_ptr) { |
18bceab1 JA |
4815 | pt->error = -EINVAL; |
4816 | return; | |
4817 | } | |
4818 | poll = kmalloc(sizeof(*poll), GFP_ATOMIC); | |
4819 | if (!poll) { | |
4820 | pt->error = -ENOMEM; | |
4821 | return; | |
4822 | } | |
4823 | io_init_poll_iocb(poll, req->poll.events, io_poll_double_wake); | |
4824 | refcount_inc(&req->refs); | |
4825 | poll->wait.private = req; | |
807abcb0 | 4826 | *poll_ptr = poll; |
18bceab1 JA |
4827 | } |
4828 | ||
4829 | pt->error = 0; | |
4830 | poll->head = head; | |
a31eb4a2 JX |
4831 | |
4832 | if (poll->events & EPOLLEXCLUSIVE) | |
4833 | add_wait_queue_exclusive(head, &poll->wait); | |
4834 | else | |
4835 | add_wait_queue(head, &poll->wait); | |
18bceab1 JA |
4836 | } |
4837 | ||
4838 | static void io_async_queue_proc(struct file *file, struct wait_queue_head *head, | |
4839 | struct poll_table_struct *p) | |
4840 | { | |
4841 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
807abcb0 | 4842 | struct async_poll *apoll = pt->req->apoll; |
18bceab1 | 4843 | |
807abcb0 | 4844 | __io_queue_proc(&apoll->poll, pt, head, &apoll->double_poll); |
18bceab1 JA |
4845 | } |
4846 | ||
d7718a9d JA |
4847 | static void io_async_task_func(struct callback_head *cb) |
4848 | { | |
4849 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); | |
4850 | struct async_poll *apoll = req->apoll; | |
4851 | struct io_ring_ctx *ctx = req->ctx; | |
4852 | ||
4853 | trace_io_uring_task_run(req->ctx, req->opcode, req->user_data); | |
4854 | ||
74ce6ce4 | 4855 | if (io_poll_rewait(req, &apoll->poll)) { |
d7718a9d | 4856 | spin_unlock_irq(&ctx->completion_lock); |
6d816e08 | 4857 | percpu_ref_put(&ctx->refs); |
74ce6ce4 | 4858 | return; |
d7718a9d JA |
4859 | } |
4860 | ||
31067255 | 4861 | /* If req is still hashed, it cannot have been canceled. Don't check. */ |
0be0b0e3 | 4862 | if (hash_hashed(&req->hash_node)) |
74ce6ce4 | 4863 | hash_del(&req->hash_node); |
2bae047e | 4864 | |
d4e7cd36 | 4865 | io_poll_remove_double(req); |
74ce6ce4 JA |
4866 | spin_unlock_irq(&ctx->completion_lock); |
4867 | ||
0be0b0e3 PB |
4868 | if (!READ_ONCE(apoll->poll.canceled)) |
4869 | __io_req_task_submit(req); | |
4870 | else | |
4871 | __io_req_task_cancel(req, -ECANCELED); | |
aa340845 | 4872 | |
6d816e08 | 4873 | percpu_ref_put(&ctx->refs); |
807abcb0 | 4874 | kfree(apoll->double_poll); |
31067255 | 4875 | kfree(apoll); |
d7718a9d JA |
4876 | } |
4877 | ||
4878 | static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
4879 | void *key) | |
4880 | { | |
4881 | struct io_kiocb *req = wait->private; | |
4882 | struct io_poll_iocb *poll = &req->apoll->poll; | |
4883 | ||
4884 | trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data, | |
4885 | key_to_poll(key)); | |
4886 | ||
4887 | return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func); | |
4888 | } | |
4889 | ||
4890 | static void io_poll_req_insert(struct io_kiocb *req) | |
4891 | { | |
4892 | struct io_ring_ctx *ctx = req->ctx; | |
4893 | struct hlist_head *list; | |
4894 | ||
4895 | list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)]; | |
4896 | hlist_add_head(&req->hash_node, list); | |
4897 | } | |
4898 | ||
4899 | static __poll_t __io_arm_poll_handler(struct io_kiocb *req, | |
4900 | struct io_poll_iocb *poll, | |
4901 | struct io_poll_table *ipt, __poll_t mask, | |
4902 | wait_queue_func_t wake_func) | |
4903 | __acquires(&ctx->completion_lock) | |
4904 | { | |
4905 | struct io_ring_ctx *ctx = req->ctx; | |
4906 | bool cancel = false; | |
4907 | ||
18bceab1 | 4908 | io_init_poll_iocb(poll, mask, wake_func); |
b90cd197 | 4909 | poll->file = req->file; |
18bceab1 | 4910 | poll->wait.private = req; |
d7718a9d JA |
4911 | |
4912 | ipt->pt._key = mask; | |
4913 | ipt->req = req; | |
4914 | ipt->error = -EINVAL; | |
4915 | ||
d7718a9d JA |
4916 | mask = vfs_poll(req->file, &ipt->pt) & poll->events; |
4917 | ||
4918 | spin_lock_irq(&ctx->completion_lock); | |
4919 | if (likely(poll->head)) { | |
4920 | spin_lock(&poll->head->lock); | |
4921 | if (unlikely(list_empty(&poll->wait.entry))) { | |
4922 | if (ipt->error) | |
4923 | cancel = true; | |
4924 | ipt->error = 0; | |
4925 | mask = 0; | |
4926 | } | |
4927 | if (mask || ipt->error) | |
4928 | list_del_init(&poll->wait.entry); | |
4929 | else if (cancel) | |
4930 | WRITE_ONCE(poll->canceled, true); | |
4931 | else if (!poll->done) /* actually waiting for an event */ | |
4932 | io_poll_req_insert(req); | |
4933 | spin_unlock(&poll->head->lock); | |
4934 | } | |
4935 | ||
4936 | return mask; | |
4937 | } | |
4938 | ||
4939 | static bool io_arm_poll_handler(struct io_kiocb *req) | |
4940 | { | |
4941 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
4942 | struct io_ring_ctx *ctx = req->ctx; | |
4943 | struct async_poll *apoll; | |
4944 | struct io_poll_table ipt; | |
4945 | __poll_t mask, ret; | |
9dab14b8 | 4946 | int rw; |
d7718a9d JA |
4947 | |
4948 | if (!req->file || !file_can_poll(req->file)) | |
4949 | return false; | |
24c74678 | 4950 | if (req->flags & REQ_F_POLLED) |
d7718a9d | 4951 | return false; |
9dab14b8 JA |
4952 | if (def->pollin) |
4953 | rw = READ; | |
4954 | else if (def->pollout) | |
4955 | rw = WRITE; | |
4956 | else | |
4957 | return false; | |
4958 | /* if we can't nonblock try, then no point in arming a poll handler */ | |
4959 | if (!io_file_supports_async(req->file, rw)) | |
d7718a9d JA |
4960 | return false; |
4961 | ||
4962 | apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC); | |
4963 | if (unlikely(!apoll)) | |
4964 | return false; | |
807abcb0 | 4965 | apoll->double_poll = NULL; |
d7718a9d JA |
4966 | |
4967 | req->flags |= REQ_F_POLLED; | |
4dd2824d | 4968 | io_get_req_task(req); |
d7718a9d JA |
4969 | req->apoll = apoll; |
4970 | INIT_HLIST_NODE(&req->hash_node); | |
4971 | ||
8755d97a | 4972 | mask = 0; |
d7718a9d | 4973 | if (def->pollin) |
8755d97a | 4974 | mask |= POLLIN | POLLRDNORM; |
d7718a9d JA |
4975 | if (def->pollout) |
4976 | mask |= POLLOUT | POLLWRNORM; | |
4977 | mask |= POLLERR | POLLPRI; | |
4978 | ||
4979 | ipt.pt._qproc = io_async_queue_proc; | |
4980 | ||
4981 | ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, | |
4982 | io_async_wake); | |
a36da65c | 4983 | if (ret || ipt.error) { |
d4e7cd36 | 4984 | io_poll_remove_double(req); |
d7718a9d | 4985 | spin_unlock_irq(&ctx->completion_lock); |
807abcb0 | 4986 | kfree(apoll->double_poll); |
d7718a9d JA |
4987 | kfree(apoll); |
4988 | return false; | |
4989 | } | |
4990 | spin_unlock_irq(&ctx->completion_lock); | |
4991 | trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask, | |
4992 | apoll->poll.events); | |
4993 | return true; | |
4994 | } | |
4995 | ||
4996 | static bool __io_poll_remove_one(struct io_kiocb *req, | |
4997 | struct io_poll_iocb *poll) | |
221c5eb2 | 4998 | { |
b41e9852 | 4999 | bool do_complete = false; |
221c5eb2 JA |
5000 | |
5001 | spin_lock(&poll->head->lock); | |
5002 | WRITE_ONCE(poll->canceled, true); | |
392edb45 JA |
5003 | if (!list_empty(&poll->wait.entry)) { |
5004 | list_del_init(&poll->wait.entry); | |
b41e9852 | 5005 | do_complete = true; |
221c5eb2 JA |
5006 | } |
5007 | spin_unlock(&poll->head->lock); | |
3bfa5bcb | 5008 | hash_del(&req->hash_node); |
d7718a9d JA |
5009 | return do_complete; |
5010 | } | |
5011 | ||
5012 | static bool io_poll_remove_one(struct io_kiocb *req) | |
5013 | { | |
5014 | bool do_complete; | |
5015 | ||
d4e7cd36 JA |
5016 | io_poll_remove_double(req); |
5017 | ||
d7718a9d JA |
5018 | if (req->opcode == IORING_OP_POLL_ADD) { |
5019 | do_complete = __io_poll_remove_one(req, &req->poll); | |
5020 | } else { | |
3bfa5bcb JA |
5021 | struct async_poll *apoll = req->apoll; |
5022 | ||
d7718a9d | 5023 | /* non-poll requests have submit ref still */ |
3bfa5bcb JA |
5024 | do_complete = __io_poll_remove_one(req, &apoll->poll); |
5025 | if (do_complete) { | |
d7718a9d | 5026 | io_put_req(req); |
807abcb0 | 5027 | kfree(apoll->double_poll); |
3bfa5bcb JA |
5028 | kfree(apoll); |
5029 | } | |
b1f573bd XW |
5030 | } |
5031 | ||
b41e9852 JA |
5032 | if (do_complete) { |
5033 | io_cqring_fill_event(req, -ECANCELED); | |
5034 | io_commit_cqring(req->ctx); | |
5035 | req->flags |= REQ_F_COMP_LOCKED; | |
f254ac04 | 5036 | req_set_fail_links(req); |
b41e9852 JA |
5037 | io_put_req(req); |
5038 | } | |
5039 | ||
5040 | return do_complete; | |
221c5eb2 JA |
5041 | } |
5042 | ||
f3606e3a | 5043 | static void io_poll_remove_all(struct io_ring_ctx *ctx, struct task_struct *tsk) |
221c5eb2 | 5044 | { |
78076bb6 | 5045 | struct hlist_node *tmp; |
221c5eb2 | 5046 | struct io_kiocb *req; |
8e2e1faf | 5047 | int posted = 0, i; |
221c5eb2 JA |
5048 | |
5049 | spin_lock_irq(&ctx->completion_lock); | |
78076bb6 JA |
5050 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { |
5051 | struct hlist_head *list; | |
5052 | ||
5053 | list = &ctx->cancel_hash[i]; | |
f3606e3a JA |
5054 | hlist_for_each_entry_safe(req, tmp, list, hash_node) { |
5055 | if (io_task_match(req, tsk)) | |
5056 | posted += io_poll_remove_one(req); | |
5057 | } | |
221c5eb2 JA |
5058 | } |
5059 | spin_unlock_irq(&ctx->completion_lock); | |
b41e9852 | 5060 | |
8e2e1faf JA |
5061 | if (posted) |
5062 | io_cqring_ev_posted(ctx); | |
221c5eb2 JA |
5063 | } |
5064 | ||
47f46768 JA |
5065 | static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr) |
5066 | { | |
78076bb6 | 5067 | struct hlist_head *list; |
47f46768 JA |
5068 | struct io_kiocb *req; |
5069 | ||
78076bb6 JA |
5070 | list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)]; |
5071 | hlist_for_each_entry(req, list, hash_node) { | |
b41e9852 JA |
5072 | if (sqe_addr != req->user_data) |
5073 | continue; | |
5074 | if (io_poll_remove_one(req)) | |
eac406c6 | 5075 | return 0; |
b41e9852 | 5076 | return -EALREADY; |
47f46768 JA |
5077 | } |
5078 | ||
5079 | return -ENOENT; | |
5080 | } | |
5081 | ||
3529d8c2 JA |
5082 | static int io_poll_remove_prep(struct io_kiocb *req, |
5083 | const struct io_uring_sqe *sqe) | |
0969e783 | 5084 | { |
0969e783 JA |
5085 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5086 | return -EINVAL; | |
5087 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index || | |
5088 | sqe->poll_events) | |
5089 | return -EINVAL; | |
5090 | ||
5091 | req->poll.addr = READ_ONCE(sqe->addr); | |
0969e783 JA |
5092 | return 0; |
5093 | } | |
5094 | ||
221c5eb2 JA |
5095 | /* |
5096 | * Find a running poll command that matches one specified in sqe->addr, | |
5097 | * and remove it if found. | |
5098 | */ | |
fc4df999 | 5099 | static int io_poll_remove(struct io_kiocb *req) |
221c5eb2 JA |
5100 | { |
5101 | struct io_ring_ctx *ctx = req->ctx; | |
0969e783 | 5102 | u64 addr; |
47f46768 | 5103 | int ret; |
221c5eb2 | 5104 | |
0969e783 | 5105 | addr = req->poll.addr; |
221c5eb2 | 5106 | spin_lock_irq(&ctx->completion_lock); |
0969e783 | 5107 | ret = io_poll_cancel(ctx, addr); |
221c5eb2 JA |
5108 | spin_unlock_irq(&ctx->completion_lock); |
5109 | ||
4e88d6e7 JA |
5110 | if (ret < 0) |
5111 | req_set_fail_links(req); | |
e1e16097 | 5112 | io_req_complete(req, ret); |
221c5eb2 JA |
5113 | return 0; |
5114 | } | |
5115 | ||
221c5eb2 JA |
5116 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, |
5117 | void *key) | |
5118 | { | |
c2f2eb7d JA |
5119 | struct io_kiocb *req = wait->private; |
5120 | struct io_poll_iocb *poll = &req->poll; | |
221c5eb2 | 5121 | |
d7718a9d | 5122 | return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func); |
221c5eb2 JA |
5123 | } |
5124 | ||
221c5eb2 JA |
5125 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, |
5126 | struct poll_table_struct *p) | |
5127 | { | |
5128 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
5129 | ||
807abcb0 | 5130 | __io_queue_proc(&pt->req->poll, pt, head, (struct io_poll_iocb **) &pt->req->io); |
eac406c6 JA |
5131 | } |
5132 | ||
3529d8c2 | 5133 | static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
221c5eb2 JA |
5134 | { |
5135 | struct io_poll_iocb *poll = &req->poll; | |
5769a351 | 5136 | u32 events; |
221c5eb2 JA |
5137 | |
5138 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
5139 | return -EINVAL; | |
5140 | if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
5141 | return -EINVAL; | |
09bb8394 JA |
5142 | if (!poll->file) |
5143 | return -EBADF; | |
221c5eb2 | 5144 | |
5769a351 JX |
5145 | events = READ_ONCE(sqe->poll32_events); |
5146 | #ifdef __BIG_ENDIAN | |
5147 | events = swahw32(events); | |
5148 | #endif | |
a31eb4a2 JX |
5149 | poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP | |
5150 | (events & EPOLLEXCLUSIVE); | |
b41e9852 | 5151 | |
4dd2824d | 5152 | io_get_req_task(req); |
0969e783 JA |
5153 | return 0; |
5154 | } | |
5155 | ||
014db007 | 5156 | static int io_poll_add(struct io_kiocb *req) |
0969e783 JA |
5157 | { |
5158 | struct io_poll_iocb *poll = &req->poll; | |
5159 | struct io_ring_ctx *ctx = req->ctx; | |
5160 | struct io_poll_table ipt; | |
0969e783 | 5161 | __poll_t mask; |
0969e783 | 5162 | |
78076bb6 | 5163 | INIT_HLIST_NODE(&req->hash_node); |
d7718a9d | 5164 | ipt.pt._qproc = io_poll_queue_proc; |
36703247 | 5165 | |
d7718a9d JA |
5166 | mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events, |
5167 | io_poll_wake); | |
221c5eb2 | 5168 | |
8c838788 | 5169 | if (mask) { /* no async, we'd stolen it */ |
221c5eb2 | 5170 | ipt.error = 0; |
b0dd8a41 | 5171 | io_poll_complete(req, mask, 0); |
221c5eb2 | 5172 | } |
221c5eb2 JA |
5173 | spin_unlock_irq(&ctx->completion_lock); |
5174 | ||
8c838788 JA |
5175 | if (mask) { |
5176 | io_cqring_ev_posted(ctx); | |
014db007 | 5177 | io_put_req(req); |
221c5eb2 | 5178 | } |
8c838788 | 5179 | return ipt.error; |
221c5eb2 JA |
5180 | } |
5181 | ||
5262f567 JA |
5182 | static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer) |
5183 | { | |
ad8a48ac JA |
5184 | struct io_timeout_data *data = container_of(timer, |
5185 | struct io_timeout_data, timer); | |
5186 | struct io_kiocb *req = data->req; | |
5187 | struct io_ring_ctx *ctx = req->ctx; | |
5262f567 JA |
5188 | unsigned long flags; |
5189 | ||
5262f567 | 5190 | spin_lock_irqsave(&ctx->completion_lock, flags); |
01cec8c1 PB |
5191 | atomic_set(&req->ctx->cq_timeouts, |
5192 | atomic_read(&req->ctx->cq_timeouts) + 1); | |
5193 | ||
ef03681a | 5194 | /* |
11365043 JA |
5195 | * We could be racing with timeout deletion. If the list is empty, |
5196 | * then timeout lookup already found it and will be handling it. | |
ef03681a | 5197 | */ |
135fcde8 PB |
5198 | if (!list_empty(&req->timeout.list)) |
5199 | list_del_init(&req->timeout.list); | |
5262f567 | 5200 | |
78e19bbe | 5201 | io_cqring_fill_event(req, -ETIME); |
5262f567 JA |
5202 | io_commit_cqring(ctx); |
5203 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
5204 | ||
5205 | io_cqring_ev_posted(ctx); | |
4e88d6e7 | 5206 | req_set_fail_links(req); |
5262f567 JA |
5207 | io_put_req(req); |
5208 | return HRTIMER_NORESTART; | |
5209 | } | |
5210 | ||
f254ac04 JA |
5211 | static int __io_timeout_cancel(struct io_kiocb *req) |
5212 | { | |
5213 | int ret; | |
5214 | ||
5215 | list_del_init(&req->timeout.list); | |
5216 | ||
5217 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); | |
5218 | if (ret == -1) | |
5219 | return -EALREADY; | |
5220 | ||
5221 | req_set_fail_links(req); | |
5222 | req->flags |= REQ_F_COMP_LOCKED; | |
5223 | io_cqring_fill_event(req, -ECANCELED); | |
5224 | io_put_req(req); | |
5225 | return 0; | |
5226 | } | |
5227 | ||
47f46768 JA |
5228 | static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data) |
5229 | { | |
5230 | struct io_kiocb *req; | |
5231 | int ret = -ENOENT; | |
5232 | ||
135fcde8 | 5233 | list_for_each_entry(req, &ctx->timeout_list, timeout.list) { |
47f46768 | 5234 | if (user_data == req->user_data) { |
47f46768 JA |
5235 | ret = 0; |
5236 | break; | |
5237 | } | |
5238 | } | |
5239 | ||
5240 | if (ret == -ENOENT) | |
5241 | return ret; | |
5242 | ||
f254ac04 | 5243 | return __io_timeout_cancel(req); |
47f46768 JA |
5244 | } |
5245 | ||
3529d8c2 JA |
5246 | static int io_timeout_remove_prep(struct io_kiocb *req, |
5247 | const struct io_uring_sqe *sqe) | |
b29472ee | 5248 | { |
b29472ee JA |
5249 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5250 | return -EINVAL; | |
61710e43 DA |
5251 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) |
5252 | return -EINVAL; | |
5253 | if (sqe->ioprio || sqe->buf_index || sqe->len) | |
b29472ee JA |
5254 | return -EINVAL; |
5255 | ||
5256 | req->timeout.addr = READ_ONCE(sqe->addr); | |
5257 | req->timeout.flags = READ_ONCE(sqe->timeout_flags); | |
5258 | if (req->timeout.flags) | |
5259 | return -EINVAL; | |
5260 | ||
b29472ee JA |
5261 | return 0; |
5262 | } | |
5263 | ||
11365043 JA |
5264 | /* |
5265 | * Remove or update an existing timeout command | |
5266 | */ | |
fc4df999 | 5267 | static int io_timeout_remove(struct io_kiocb *req) |
11365043 JA |
5268 | { |
5269 | struct io_ring_ctx *ctx = req->ctx; | |
47f46768 | 5270 | int ret; |
11365043 | 5271 | |
11365043 | 5272 | spin_lock_irq(&ctx->completion_lock); |
b29472ee | 5273 | ret = io_timeout_cancel(ctx, req->timeout.addr); |
11365043 | 5274 | |
47f46768 | 5275 | io_cqring_fill_event(req, ret); |
11365043 JA |
5276 | io_commit_cqring(ctx); |
5277 | spin_unlock_irq(&ctx->completion_lock); | |
5262f567 | 5278 | io_cqring_ev_posted(ctx); |
4e88d6e7 JA |
5279 | if (ret < 0) |
5280 | req_set_fail_links(req); | |
ec9c02ad | 5281 | io_put_req(req); |
11365043 | 5282 | return 0; |
5262f567 JA |
5283 | } |
5284 | ||
3529d8c2 | 5285 | static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2d28390a | 5286 | bool is_timeout_link) |
5262f567 | 5287 | { |
ad8a48ac | 5288 | struct io_timeout_data *data; |
a41525ab | 5289 | unsigned flags; |
56080b02 | 5290 | u32 off = READ_ONCE(sqe->off); |
5262f567 | 5291 | |
ad8a48ac | 5292 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5262f567 | 5293 | return -EINVAL; |
ad8a48ac | 5294 | if (sqe->ioprio || sqe->buf_index || sqe->len != 1) |
a41525ab | 5295 | return -EINVAL; |
56080b02 | 5296 | if (off && is_timeout_link) |
2d28390a | 5297 | return -EINVAL; |
a41525ab JA |
5298 | flags = READ_ONCE(sqe->timeout_flags); |
5299 | if (flags & ~IORING_TIMEOUT_ABS) | |
5262f567 | 5300 | return -EINVAL; |
bdf20073 | 5301 | |
bfe68a22 | 5302 | req->timeout.off = off; |
26a61679 | 5303 | |
3529d8c2 | 5304 | if (!req->io && io_alloc_async_ctx(req)) |
26a61679 JA |
5305 | return -ENOMEM; |
5306 | ||
5307 | data = &req->io->timeout; | |
ad8a48ac | 5308 | data->req = req; |
ad8a48ac JA |
5309 | |
5310 | if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr))) | |
5262f567 JA |
5311 | return -EFAULT; |
5312 | ||
11365043 | 5313 | if (flags & IORING_TIMEOUT_ABS) |
ad8a48ac | 5314 | data->mode = HRTIMER_MODE_ABS; |
11365043 | 5315 | else |
ad8a48ac | 5316 | data->mode = HRTIMER_MODE_REL; |
11365043 | 5317 | |
ad8a48ac JA |
5318 | hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode); |
5319 | return 0; | |
5320 | } | |
5321 | ||
fc4df999 | 5322 | static int io_timeout(struct io_kiocb *req) |
ad8a48ac | 5323 | { |
ad8a48ac | 5324 | struct io_ring_ctx *ctx = req->ctx; |
bfe68a22 | 5325 | struct io_timeout_data *data = &req->io->timeout; |
ad8a48ac | 5326 | struct list_head *entry; |
bfe68a22 | 5327 | u32 tail, off = req->timeout.off; |
ad8a48ac | 5328 | |
733f5c95 | 5329 | spin_lock_irq(&ctx->completion_lock); |
93bd25bb | 5330 | |
5262f567 JA |
5331 | /* |
5332 | * sqe->off holds how many events that need to occur for this | |
93bd25bb JA |
5333 | * timeout event to be satisfied. If it isn't set, then this is |
5334 | * a pure timeout request, sequence isn't used. | |
5262f567 | 5335 | */ |
8eb7e2d0 | 5336 | if (io_is_timeout_noseq(req)) { |
93bd25bb JA |
5337 | entry = ctx->timeout_list.prev; |
5338 | goto add; | |
5339 | } | |
5262f567 | 5340 | |
bfe68a22 PB |
5341 | tail = ctx->cached_cq_tail - atomic_read(&ctx->cq_timeouts); |
5342 | req->timeout.target_seq = tail + off; | |
5262f567 JA |
5343 | |
5344 | /* | |
5345 | * Insertion sort, ensuring the first entry in the list is always | |
5346 | * the one we need first. | |
5347 | */ | |
5262f567 | 5348 | list_for_each_prev(entry, &ctx->timeout_list) { |
135fcde8 PB |
5349 | struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, |
5350 | timeout.list); | |
5262f567 | 5351 | |
8eb7e2d0 | 5352 | if (io_is_timeout_noseq(nxt)) |
93bd25bb | 5353 | continue; |
bfe68a22 PB |
5354 | /* nxt.seq is behind @tail, otherwise would've been completed */ |
5355 | if (off >= nxt->timeout.target_seq - tail) | |
5262f567 JA |
5356 | break; |
5357 | } | |
93bd25bb | 5358 | add: |
135fcde8 | 5359 | list_add(&req->timeout.list, entry); |
ad8a48ac JA |
5360 | data->timer.function = io_timeout_fn; |
5361 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode); | |
5262f567 | 5362 | spin_unlock_irq(&ctx->completion_lock); |
5262f567 JA |
5363 | return 0; |
5364 | } | |
5262f567 | 5365 | |
62755e35 JA |
5366 | static bool io_cancel_cb(struct io_wq_work *work, void *data) |
5367 | { | |
5368 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
5369 | ||
5370 | return req->user_data == (unsigned long) data; | |
5371 | } | |
5372 | ||
e977d6d3 | 5373 | static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr) |
62755e35 | 5374 | { |
62755e35 | 5375 | enum io_wq_cancel cancel_ret; |
62755e35 JA |
5376 | int ret = 0; |
5377 | ||
4f26bda1 | 5378 | cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr, false); |
62755e35 JA |
5379 | switch (cancel_ret) { |
5380 | case IO_WQ_CANCEL_OK: | |
5381 | ret = 0; | |
5382 | break; | |
5383 | case IO_WQ_CANCEL_RUNNING: | |
5384 | ret = -EALREADY; | |
5385 | break; | |
5386 | case IO_WQ_CANCEL_NOTFOUND: | |
5387 | ret = -ENOENT; | |
5388 | break; | |
5389 | } | |
5390 | ||
e977d6d3 JA |
5391 | return ret; |
5392 | } | |
5393 | ||
47f46768 JA |
5394 | static void io_async_find_and_cancel(struct io_ring_ctx *ctx, |
5395 | struct io_kiocb *req, __u64 sqe_addr, | |
014db007 | 5396 | int success_ret) |
47f46768 JA |
5397 | { |
5398 | unsigned long flags; | |
5399 | int ret; | |
5400 | ||
5401 | ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr); | |
5402 | if (ret != -ENOENT) { | |
5403 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
5404 | goto done; | |
5405 | } | |
5406 | ||
5407 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
5408 | ret = io_timeout_cancel(ctx, sqe_addr); | |
5409 | if (ret != -ENOENT) | |
5410 | goto done; | |
5411 | ret = io_poll_cancel(ctx, sqe_addr); | |
5412 | done: | |
b0dd8a41 JA |
5413 | if (!ret) |
5414 | ret = success_ret; | |
47f46768 JA |
5415 | io_cqring_fill_event(req, ret); |
5416 | io_commit_cqring(ctx); | |
5417 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
5418 | io_cqring_ev_posted(ctx); | |
5419 | ||
4e88d6e7 JA |
5420 | if (ret < 0) |
5421 | req_set_fail_links(req); | |
014db007 | 5422 | io_put_req(req); |
47f46768 JA |
5423 | } |
5424 | ||
3529d8c2 JA |
5425 | static int io_async_cancel_prep(struct io_kiocb *req, |
5426 | const struct io_uring_sqe *sqe) | |
e977d6d3 | 5427 | { |
fbf23849 | 5428 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
e977d6d3 | 5429 | return -EINVAL; |
61710e43 DA |
5430 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) |
5431 | return -EINVAL; | |
5432 | if (sqe->ioprio || sqe->off || sqe->len || sqe->cancel_flags) | |
e977d6d3 JA |
5433 | return -EINVAL; |
5434 | ||
fbf23849 JA |
5435 | req->cancel.addr = READ_ONCE(sqe->addr); |
5436 | return 0; | |
5437 | } | |
5438 | ||
014db007 | 5439 | static int io_async_cancel(struct io_kiocb *req) |
fbf23849 JA |
5440 | { |
5441 | struct io_ring_ctx *ctx = req->ctx; | |
fbf23849 | 5442 | |
014db007 | 5443 | io_async_find_and_cancel(ctx, req, req->cancel.addr, 0); |
5262f567 JA |
5444 | return 0; |
5445 | } | |
5446 | ||
05f3fb3c JA |
5447 | static int io_files_update_prep(struct io_kiocb *req, |
5448 | const struct io_uring_sqe *sqe) | |
5449 | { | |
6ca56f84 JA |
5450 | if (unlikely(req->ctx->flags & IORING_SETUP_SQPOLL)) |
5451 | return -EINVAL; | |
61710e43 DA |
5452 | if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) |
5453 | return -EINVAL; | |
5454 | if (sqe->ioprio || sqe->rw_flags) | |
05f3fb3c JA |
5455 | return -EINVAL; |
5456 | ||
5457 | req->files_update.offset = READ_ONCE(sqe->off); | |
5458 | req->files_update.nr_args = READ_ONCE(sqe->len); | |
5459 | if (!req->files_update.nr_args) | |
5460 | return -EINVAL; | |
5461 | req->files_update.arg = READ_ONCE(sqe->addr); | |
5462 | return 0; | |
5463 | } | |
5464 | ||
229a7b63 JA |
5465 | static int io_files_update(struct io_kiocb *req, bool force_nonblock, |
5466 | struct io_comp_state *cs) | |
fbf23849 JA |
5467 | { |
5468 | struct io_ring_ctx *ctx = req->ctx; | |
05f3fb3c JA |
5469 | struct io_uring_files_update up; |
5470 | int ret; | |
fbf23849 | 5471 | |
f86cd20c | 5472 | if (force_nonblock) |
05f3fb3c | 5473 | return -EAGAIN; |
05f3fb3c JA |
5474 | |
5475 | up.offset = req->files_update.offset; | |
5476 | up.fds = req->files_update.arg; | |
5477 | ||
5478 | mutex_lock(&ctx->uring_lock); | |
5479 | ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args); | |
5480 | mutex_unlock(&ctx->uring_lock); | |
5481 | ||
5482 | if (ret < 0) | |
5483 | req_set_fail_links(req); | |
229a7b63 | 5484 | __io_req_complete(req, ret, 0, cs); |
5262f567 JA |
5485 | return 0; |
5486 | } | |
5487 | ||
3529d8c2 JA |
5488 | static int io_req_defer_prep(struct io_kiocb *req, |
5489 | const struct io_uring_sqe *sqe) | |
f67676d1 | 5490 | { |
e781573e | 5491 | ssize_t ret = 0; |
f67676d1 | 5492 | |
f1d96a8f PB |
5493 | if (!sqe) |
5494 | return 0; | |
5495 | ||
327d6d96 PB |
5496 | if (io_alloc_async_ctx(req)) |
5497 | return -EAGAIN; | |
f56040b8 PB |
5498 | ret = io_prep_work_files(req); |
5499 | if (unlikely(ret)) | |
5500 | return ret; | |
cccf0ee8 | 5501 | |
202700e1 JA |
5502 | io_prep_async_work(req); |
5503 | ||
d625c6ee | 5504 | switch (req->opcode) { |
e781573e JA |
5505 | case IORING_OP_NOP: |
5506 | break; | |
f67676d1 JA |
5507 | case IORING_OP_READV: |
5508 | case IORING_OP_READ_FIXED: | |
3a6820f2 | 5509 | case IORING_OP_READ: |
3529d8c2 | 5510 | ret = io_read_prep(req, sqe, true); |
f67676d1 JA |
5511 | break; |
5512 | case IORING_OP_WRITEV: | |
5513 | case IORING_OP_WRITE_FIXED: | |
3a6820f2 | 5514 | case IORING_OP_WRITE: |
3529d8c2 | 5515 | ret = io_write_prep(req, sqe, true); |
f67676d1 | 5516 | break; |
0969e783 | 5517 | case IORING_OP_POLL_ADD: |
3529d8c2 | 5518 | ret = io_poll_add_prep(req, sqe); |
0969e783 JA |
5519 | break; |
5520 | case IORING_OP_POLL_REMOVE: | |
3529d8c2 | 5521 | ret = io_poll_remove_prep(req, sqe); |
0969e783 | 5522 | break; |
8ed8d3c3 | 5523 | case IORING_OP_FSYNC: |
3529d8c2 | 5524 | ret = io_prep_fsync(req, sqe); |
8ed8d3c3 JA |
5525 | break; |
5526 | case IORING_OP_SYNC_FILE_RANGE: | |
3529d8c2 | 5527 | ret = io_prep_sfr(req, sqe); |
8ed8d3c3 | 5528 | break; |
03b1230c | 5529 | case IORING_OP_SENDMSG: |
fddaface | 5530 | case IORING_OP_SEND: |
3529d8c2 | 5531 | ret = io_sendmsg_prep(req, sqe); |
03b1230c JA |
5532 | break; |
5533 | case IORING_OP_RECVMSG: | |
fddaface | 5534 | case IORING_OP_RECV: |
3529d8c2 | 5535 | ret = io_recvmsg_prep(req, sqe); |
03b1230c | 5536 | break; |
f499a021 | 5537 | case IORING_OP_CONNECT: |
3529d8c2 | 5538 | ret = io_connect_prep(req, sqe); |
f499a021 | 5539 | break; |
2d28390a | 5540 | case IORING_OP_TIMEOUT: |
3529d8c2 | 5541 | ret = io_timeout_prep(req, sqe, false); |
b7bb4f7d | 5542 | break; |
b29472ee | 5543 | case IORING_OP_TIMEOUT_REMOVE: |
3529d8c2 | 5544 | ret = io_timeout_remove_prep(req, sqe); |
b29472ee | 5545 | break; |
fbf23849 | 5546 | case IORING_OP_ASYNC_CANCEL: |
3529d8c2 | 5547 | ret = io_async_cancel_prep(req, sqe); |
fbf23849 | 5548 | break; |
2d28390a | 5549 | case IORING_OP_LINK_TIMEOUT: |
3529d8c2 | 5550 | ret = io_timeout_prep(req, sqe, true); |
b7bb4f7d | 5551 | break; |
8ed8d3c3 | 5552 | case IORING_OP_ACCEPT: |
3529d8c2 | 5553 | ret = io_accept_prep(req, sqe); |
8ed8d3c3 | 5554 | break; |
d63d1b5e JA |
5555 | case IORING_OP_FALLOCATE: |
5556 | ret = io_fallocate_prep(req, sqe); | |
5557 | break; | |
15b71abe JA |
5558 | case IORING_OP_OPENAT: |
5559 | ret = io_openat_prep(req, sqe); | |
5560 | break; | |
b5dba59e JA |
5561 | case IORING_OP_CLOSE: |
5562 | ret = io_close_prep(req, sqe); | |
5563 | break; | |
05f3fb3c JA |
5564 | case IORING_OP_FILES_UPDATE: |
5565 | ret = io_files_update_prep(req, sqe); | |
5566 | break; | |
eddc7ef5 JA |
5567 | case IORING_OP_STATX: |
5568 | ret = io_statx_prep(req, sqe); | |
5569 | break; | |
4840e418 JA |
5570 | case IORING_OP_FADVISE: |
5571 | ret = io_fadvise_prep(req, sqe); | |
5572 | break; | |
c1ca757b JA |
5573 | case IORING_OP_MADVISE: |
5574 | ret = io_madvise_prep(req, sqe); | |
5575 | break; | |
cebdb986 JA |
5576 | case IORING_OP_OPENAT2: |
5577 | ret = io_openat2_prep(req, sqe); | |
5578 | break; | |
3e4827b0 JA |
5579 | case IORING_OP_EPOLL_CTL: |
5580 | ret = io_epoll_ctl_prep(req, sqe); | |
5581 | break; | |
7d67af2c PB |
5582 | case IORING_OP_SPLICE: |
5583 | ret = io_splice_prep(req, sqe); | |
5584 | break; | |
ddf0322d JA |
5585 | case IORING_OP_PROVIDE_BUFFERS: |
5586 | ret = io_provide_buffers_prep(req, sqe); | |
5587 | break; | |
067524e9 JA |
5588 | case IORING_OP_REMOVE_BUFFERS: |
5589 | ret = io_remove_buffers_prep(req, sqe); | |
5590 | break; | |
f2a8d5c7 PB |
5591 | case IORING_OP_TEE: |
5592 | ret = io_tee_prep(req, sqe); | |
5593 | break; | |
f67676d1 | 5594 | default: |
e781573e JA |
5595 | printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n", |
5596 | req->opcode); | |
5597 | ret = -EINVAL; | |
b7bb4f7d | 5598 | break; |
f67676d1 JA |
5599 | } |
5600 | ||
b7bb4f7d | 5601 | return ret; |
f67676d1 JA |
5602 | } |
5603 | ||
9cf7c104 PB |
5604 | static u32 io_get_sequence(struct io_kiocb *req) |
5605 | { | |
5606 | struct io_kiocb *pos; | |
5607 | struct io_ring_ctx *ctx = req->ctx; | |
5608 | u32 total_submitted, nr_reqs = 1; | |
5609 | ||
5610 | if (req->flags & REQ_F_LINK_HEAD) | |
5611 | list_for_each_entry(pos, &req->link_list, link_list) | |
5612 | nr_reqs++; | |
5613 | ||
5614 | total_submitted = ctx->cached_sq_head - ctx->cached_sq_dropped; | |
5615 | return total_submitted - nr_reqs; | |
5616 | } | |
5617 | ||
3529d8c2 | 5618 | static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
de0617e4 | 5619 | { |
a197f664 | 5620 | struct io_ring_ctx *ctx = req->ctx; |
27dc8338 | 5621 | struct io_defer_entry *de; |
f67676d1 | 5622 | int ret; |
9cf7c104 | 5623 | u32 seq; |
de0617e4 | 5624 | |
9d858b21 | 5625 | /* Still need defer if there is pending req in defer list. */ |
9cf7c104 PB |
5626 | if (likely(list_empty_careful(&ctx->defer_list) && |
5627 | !(req->flags & REQ_F_IO_DRAIN))) | |
5628 | return 0; | |
5629 | ||
5630 | seq = io_get_sequence(req); | |
5631 | /* Still a chance to pass the sequence check */ | |
5632 | if (!req_need_defer(req, seq) && list_empty_careful(&ctx->defer_list)) | |
de0617e4 JA |
5633 | return 0; |
5634 | ||
650b5481 | 5635 | if (!req->io) { |
650b5481 | 5636 | ret = io_req_defer_prep(req, sqe); |
327d6d96 | 5637 | if (ret) |
650b5481 PB |
5638 | return ret; |
5639 | } | |
cbdcb435 | 5640 | io_prep_async_link(req); |
27dc8338 PB |
5641 | de = kmalloc(sizeof(*de), GFP_KERNEL); |
5642 | if (!de) | |
5643 | return -ENOMEM; | |
2d28390a | 5644 | |
de0617e4 | 5645 | spin_lock_irq(&ctx->completion_lock); |
9cf7c104 | 5646 | if (!req_need_defer(req, seq) && list_empty(&ctx->defer_list)) { |
de0617e4 | 5647 | spin_unlock_irq(&ctx->completion_lock); |
27dc8338 | 5648 | kfree(de); |
ae34817b PB |
5649 | io_queue_async_work(req); |
5650 | return -EIOCBQUEUED; | |
de0617e4 JA |
5651 | } |
5652 | ||
915967f6 | 5653 | trace_io_uring_defer(ctx, req, req->user_data); |
27dc8338 | 5654 | de->req = req; |
9cf7c104 | 5655 | de->seq = seq; |
27dc8338 | 5656 | list_add_tail(&de->list, &ctx->defer_list); |
de0617e4 JA |
5657 | spin_unlock_irq(&ctx->completion_lock); |
5658 | return -EIOCBQUEUED; | |
5659 | } | |
5660 | ||
f573d384 JA |
5661 | static void io_req_drop_files(struct io_kiocb *req) |
5662 | { | |
5663 | struct io_ring_ctx *ctx = req->ctx; | |
5664 | unsigned long flags; | |
5665 | ||
5666 | spin_lock_irqsave(&ctx->inflight_lock, flags); | |
5667 | list_del(&req->inflight_entry); | |
5668 | if (waitqueue_active(&ctx->inflight_wait)) | |
5669 | wake_up(&ctx->inflight_wait); | |
5670 | spin_unlock_irqrestore(&ctx->inflight_lock, flags); | |
5671 | req->flags &= ~REQ_F_INFLIGHT; | |
5672 | req->work.files = NULL; | |
5673 | } | |
5674 | ||
3ca405eb | 5675 | static void __io_clean_op(struct io_kiocb *req) |
99bc4c38 PB |
5676 | { |
5677 | struct io_async_ctx *io = req->io; | |
5678 | ||
0e1b6fe3 PB |
5679 | if (req->flags & REQ_F_BUFFER_SELECTED) { |
5680 | switch (req->opcode) { | |
5681 | case IORING_OP_READV: | |
5682 | case IORING_OP_READ_FIXED: | |
5683 | case IORING_OP_READ: | |
bcda7baa | 5684 | kfree((void *)(unsigned long)req->rw.addr); |
0e1b6fe3 PB |
5685 | break; |
5686 | case IORING_OP_RECVMSG: | |
5687 | case IORING_OP_RECV: | |
bcda7baa | 5688 | kfree(req->sr_msg.kbuf); |
0e1b6fe3 PB |
5689 | break; |
5690 | } | |
5691 | req->flags &= ~REQ_F_BUFFER_SELECTED; | |
99bc4c38 PB |
5692 | } |
5693 | ||
0e1b6fe3 PB |
5694 | if (req->flags & REQ_F_NEED_CLEANUP) { |
5695 | switch (req->opcode) { | |
5696 | case IORING_OP_READV: | |
5697 | case IORING_OP_READ_FIXED: | |
5698 | case IORING_OP_READ: | |
5699 | case IORING_OP_WRITEV: | |
5700 | case IORING_OP_WRITE_FIXED: | |
5701 | case IORING_OP_WRITE: | |
ff6165b2 JA |
5702 | if (io->rw.free_iovec) |
5703 | kfree(io->rw.free_iovec); | |
0e1b6fe3 PB |
5704 | break; |
5705 | case IORING_OP_RECVMSG: | |
5706 | case IORING_OP_SENDMSG: | |
5707 | if (io->msg.iov != io->msg.fast_iov) | |
5708 | kfree(io->msg.iov); | |
5709 | break; | |
5710 | case IORING_OP_SPLICE: | |
5711 | case IORING_OP_TEE: | |
5712 | io_put_file(req, req->splice.file_in, | |
5713 | (req->splice.flags & SPLICE_F_FD_IN_FIXED)); | |
5714 | break; | |
f3cd4850 JA |
5715 | case IORING_OP_OPENAT: |
5716 | case IORING_OP_OPENAT2: | |
5717 | if (req->open.filename) | |
5718 | putname(req->open.filename); | |
5719 | break; | |
0e1b6fe3 PB |
5720 | } |
5721 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
99bc4c38 | 5722 | } |
bb175342 | 5723 | |
f573d384 JA |
5724 | if (req->flags & REQ_F_INFLIGHT) |
5725 | io_req_drop_files(req); | |
99bc4c38 PB |
5726 | } |
5727 | ||
3529d8c2 | 5728 | static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
f13fad7b | 5729 | bool force_nonblock, struct io_comp_state *cs) |
2b188cc1 | 5730 | { |
a197f664 | 5731 | struct io_ring_ctx *ctx = req->ctx; |
d625c6ee | 5732 | int ret; |
2b188cc1 | 5733 | |
d625c6ee | 5734 | switch (req->opcode) { |
2b188cc1 | 5735 | case IORING_OP_NOP: |
229a7b63 | 5736 | ret = io_nop(req, cs); |
2b188cc1 JA |
5737 | break; |
5738 | case IORING_OP_READV: | |
edafccee | 5739 | case IORING_OP_READ_FIXED: |
3a6820f2 | 5740 | case IORING_OP_READ: |
3529d8c2 JA |
5741 | if (sqe) { |
5742 | ret = io_read_prep(req, sqe, force_nonblock); | |
5743 | if (ret < 0) | |
5744 | break; | |
5745 | } | |
a1d7c393 | 5746 | ret = io_read(req, force_nonblock, cs); |
edafccee | 5747 | break; |
3529d8c2 | 5748 | case IORING_OP_WRITEV: |
edafccee | 5749 | case IORING_OP_WRITE_FIXED: |
3a6820f2 | 5750 | case IORING_OP_WRITE: |
3529d8c2 JA |
5751 | if (sqe) { |
5752 | ret = io_write_prep(req, sqe, force_nonblock); | |
5753 | if (ret < 0) | |
5754 | break; | |
5755 | } | |
a1d7c393 | 5756 | ret = io_write(req, force_nonblock, cs); |
2b188cc1 | 5757 | break; |
c992fe29 | 5758 | case IORING_OP_FSYNC: |
3529d8c2 JA |
5759 | if (sqe) { |
5760 | ret = io_prep_fsync(req, sqe); | |
5761 | if (ret < 0) | |
5762 | break; | |
5763 | } | |
014db007 | 5764 | ret = io_fsync(req, force_nonblock); |
c992fe29 | 5765 | break; |
221c5eb2 | 5766 | case IORING_OP_POLL_ADD: |
3529d8c2 JA |
5767 | if (sqe) { |
5768 | ret = io_poll_add_prep(req, sqe); | |
5769 | if (ret) | |
5770 | break; | |
5771 | } | |
014db007 | 5772 | ret = io_poll_add(req); |
221c5eb2 JA |
5773 | break; |
5774 | case IORING_OP_POLL_REMOVE: | |
3529d8c2 JA |
5775 | if (sqe) { |
5776 | ret = io_poll_remove_prep(req, sqe); | |
5777 | if (ret < 0) | |
5778 | break; | |
5779 | } | |
fc4df999 | 5780 | ret = io_poll_remove(req); |
221c5eb2 | 5781 | break; |
5d17b4a4 | 5782 | case IORING_OP_SYNC_FILE_RANGE: |
3529d8c2 JA |
5783 | if (sqe) { |
5784 | ret = io_prep_sfr(req, sqe); | |
5785 | if (ret < 0) | |
5786 | break; | |
5787 | } | |
014db007 | 5788 | ret = io_sync_file_range(req, force_nonblock); |
5d17b4a4 | 5789 | break; |
0fa03c62 | 5790 | case IORING_OP_SENDMSG: |
fddaface | 5791 | case IORING_OP_SEND: |
3529d8c2 JA |
5792 | if (sqe) { |
5793 | ret = io_sendmsg_prep(req, sqe); | |
5794 | if (ret < 0) | |
5795 | break; | |
5796 | } | |
fddaface | 5797 | if (req->opcode == IORING_OP_SENDMSG) |
229a7b63 | 5798 | ret = io_sendmsg(req, force_nonblock, cs); |
fddaface | 5799 | else |
229a7b63 | 5800 | ret = io_send(req, force_nonblock, cs); |
0fa03c62 | 5801 | break; |
aa1fa28f | 5802 | case IORING_OP_RECVMSG: |
fddaface | 5803 | case IORING_OP_RECV: |
3529d8c2 JA |
5804 | if (sqe) { |
5805 | ret = io_recvmsg_prep(req, sqe); | |
5806 | if (ret) | |
5807 | break; | |
5808 | } | |
fddaface | 5809 | if (req->opcode == IORING_OP_RECVMSG) |
229a7b63 | 5810 | ret = io_recvmsg(req, force_nonblock, cs); |
fddaface | 5811 | else |
229a7b63 | 5812 | ret = io_recv(req, force_nonblock, cs); |
aa1fa28f | 5813 | break; |
5262f567 | 5814 | case IORING_OP_TIMEOUT: |
3529d8c2 JA |
5815 | if (sqe) { |
5816 | ret = io_timeout_prep(req, sqe, false); | |
5817 | if (ret) | |
5818 | break; | |
5819 | } | |
fc4df999 | 5820 | ret = io_timeout(req); |
5262f567 | 5821 | break; |
11365043 | 5822 | case IORING_OP_TIMEOUT_REMOVE: |
3529d8c2 JA |
5823 | if (sqe) { |
5824 | ret = io_timeout_remove_prep(req, sqe); | |
5825 | if (ret) | |
5826 | break; | |
5827 | } | |
fc4df999 | 5828 | ret = io_timeout_remove(req); |
11365043 | 5829 | break; |
17f2fe35 | 5830 | case IORING_OP_ACCEPT: |
3529d8c2 JA |
5831 | if (sqe) { |
5832 | ret = io_accept_prep(req, sqe); | |
5833 | if (ret) | |
5834 | break; | |
5835 | } | |
229a7b63 | 5836 | ret = io_accept(req, force_nonblock, cs); |
17f2fe35 | 5837 | break; |
f8e85cf2 | 5838 | case IORING_OP_CONNECT: |
3529d8c2 JA |
5839 | if (sqe) { |
5840 | ret = io_connect_prep(req, sqe); | |
5841 | if (ret) | |
5842 | break; | |
5843 | } | |
229a7b63 | 5844 | ret = io_connect(req, force_nonblock, cs); |
f8e85cf2 | 5845 | break; |
62755e35 | 5846 | case IORING_OP_ASYNC_CANCEL: |
3529d8c2 JA |
5847 | if (sqe) { |
5848 | ret = io_async_cancel_prep(req, sqe); | |
5849 | if (ret) | |
5850 | break; | |
5851 | } | |
014db007 | 5852 | ret = io_async_cancel(req); |
62755e35 | 5853 | break; |
d63d1b5e JA |
5854 | case IORING_OP_FALLOCATE: |
5855 | if (sqe) { | |
5856 | ret = io_fallocate_prep(req, sqe); | |
5857 | if (ret) | |
5858 | break; | |
5859 | } | |
014db007 | 5860 | ret = io_fallocate(req, force_nonblock); |
d63d1b5e | 5861 | break; |
15b71abe JA |
5862 | case IORING_OP_OPENAT: |
5863 | if (sqe) { | |
5864 | ret = io_openat_prep(req, sqe); | |
5865 | if (ret) | |
5866 | break; | |
5867 | } | |
014db007 | 5868 | ret = io_openat(req, force_nonblock); |
15b71abe | 5869 | break; |
b5dba59e JA |
5870 | case IORING_OP_CLOSE: |
5871 | if (sqe) { | |
5872 | ret = io_close_prep(req, sqe); | |
5873 | if (ret) | |
5874 | break; | |
5875 | } | |
229a7b63 | 5876 | ret = io_close(req, force_nonblock, cs); |
b5dba59e | 5877 | break; |
05f3fb3c JA |
5878 | case IORING_OP_FILES_UPDATE: |
5879 | if (sqe) { | |
5880 | ret = io_files_update_prep(req, sqe); | |
5881 | if (ret) | |
5882 | break; | |
5883 | } | |
229a7b63 | 5884 | ret = io_files_update(req, force_nonblock, cs); |
05f3fb3c | 5885 | break; |
eddc7ef5 JA |
5886 | case IORING_OP_STATX: |
5887 | if (sqe) { | |
5888 | ret = io_statx_prep(req, sqe); | |
5889 | if (ret) | |
5890 | break; | |
5891 | } | |
014db007 | 5892 | ret = io_statx(req, force_nonblock); |
eddc7ef5 | 5893 | break; |
4840e418 JA |
5894 | case IORING_OP_FADVISE: |
5895 | if (sqe) { | |
5896 | ret = io_fadvise_prep(req, sqe); | |
5897 | if (ret) | |
5898 | break; | |
5899 | } | |
014db007 | 5900 | ret = io_fadvise(req, force_nonblock); |
4840e418 | 5901 | break; |
c1ca757b JA |
5902 | case IORING_OP_MADVISE: |
5903 | if (sqe) { | |
5904 | ret = io_madvise_prep(req, sqe); | |
5905 | if (ret) | |
5906 | break; | |
5907 | } | |
014db007 | 5908 | ret = io_madvise(req, force_nonblock); |
c1ca757b | 5909 | break; |
cebdb986 JA |
5910 | case IORING_OP_OPENAT2: |
5911 | if (sqe) { | |
5912 | ret = io_openat2_prep(req, sqe); | |
5913 | if (ret) | |
5914 | break; | |
5915 | } | |
014db007 | 5916 | ret = io_openat2(req, force_nonblock); |
cebdb986 | 5917 | break; |
3e4827b0 JA |
5918 | case IORING_OP_EPOLL_CTL: |
5919 | if (sqe) { | |
5920 | ret = io_epoll_ctl_prep(req, sqe); | |
5921 | if (ret) | |
5922 | break; | |
5923 | } | |
229a7b63 | 5924 | ret = io_epoll_ctl(req, force_nonblock, cs); |
3e4827b0 | 5925 | break; |
7d67af2c PB |
5926 | case IORING_OP_SPLICE: |
5927 | if (sqe) { | |
5928 | ret = io_splice_prep(req, sqe); | |
5929 | if (ret < 0) | |
5930 | break; | |
5931 | } | |
014db007 | 5932 | ret = io_splice(req, force_nonblock); |
7d67af2c | 5933 | break; |
ddf0322d JA |
5934 | case IORING_OP_PROVIDE_BUFFERS: |
5935 | if (sqe) { | |
5936 | ret = io_provide_buffers_prep(req, sqe); | |
5937 | if (ret) | |
5938 | break; | |
5939 | } | |
229a7b63 | 5940 | ret = io_provide_buffers(req, force_nonblock, cs); |
ddf0322d | 5941 | break; |
067524e9 JA |
5942 | case IORING_OP_REMOVE_BUFFERS: |
5943 | if (sqe) { | |
5944 | ret = io_remove_buffers_prep(req, sqe); | |
5945 | if (ret) | |
5946 | break; | |
5947 | } | |
229a7b63 | 5948 | ret = io_remove_buffers(req, force_nonblock, cs); |
3e4827b0 | 5949 | break; |
f2a8d5c7 PB |
5950 | case IORING_OP_TEE: |
5951 | if (sqe) { | |
5952 | ret = io_tee_prep(req, sqe); | |
5953 | if (ret < 0) | |
5954 | break; | |
5955 | } | |
5956 | ret = io_tee(req, force_nonblock); | |
5957 | break; | |
2b188cc1 JA |
5958 | default: |
5959 | ret = -EINVAL; | |
5960 | break; | |
5961 | } | |
5962 | ||
def596e9 JA |
5963 | if (ret) |
5964 | return ret; | |
5965 | ||
b532576e JA |
5966 | /* If the op doesn't have a file, we're not polling for it */ |
5967 | if ((ctx->flags & IORING_SETUP_IOPOLL) && req->file) { | |
11ba820b JA |
5968 | const bool in_async = io_wq_current_is_worker(); |
5969 | ||
11ba820b JA |
5970 | /* workqueue context doesn't hold uring_lock, grab it now */ |
5971 | if (in_async) | |
5972 | mutex_lock(&ctx->uring_lock); | |
5973 | ||
def596e9 | 5974 | io_iopoll_req_issued(req); |
11ba820b JA |
5975 | |
5976 | if (in_async) | |
5977 | mutex_unlock(&ctx->uring_lock); | |
def596e9 JA |
5978 | } |
5979 | ||
5980 | return 0; | |
2b188cc1 JA |
5981 | } |
5982 | ||
f4db7182 | 5983 | static struct io_wq_work *io_wq_submit_work(struct io_wq_work *work) |
2b188cc1 JA |
5984 | { |
5985 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
6df1db6b | 5986 | struct io_kiocb *timeout; |
561fb04a | 5987 | int ret = 0; |
2b188cc1 | 5988 | |
6df1db6b PB |
5989 | timeout = io_prep_linked_timeout(req); |
5990 | if (timeout) | |
5991 | io_queue_linked_timeout(timeout); | |
d4c81f38 | 5992 | |
0c9d5ccd JA |
5993 | /* if NO_CANCEL is set, we must still run the work */ |
5994 | if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) == | |
5995 | IO_WQ_WORK_CANCEL) { | |
561fb04a | 5996 | ret = -ECANCELED; |
0c9d5ccd | 5997 | } |
31b51510 | 5998 | |
561fb04a | 5999 | if (!ret) { |
561fb04a | 6000 | do { |
f13fad7b | 6001 | ret = io_issue_sqe(req, NULL, false, NULL); |
561fb04a JA |
6002 | /* |
6003 | * We can get EAGAIN for polled IO even though we're | |
6004 | * forcing a sync submission from here, since we can't | |
6005 | * wait for request slots on the block side. | |
6006 | */ | |
6007 | if (ret != -EAGAIN) | |
6008 | break; | |
6009 | cond_resched(); | |
6010 | } while (1); | |
6011 | } | |
31b51510 | 6012 | |
561fb04a | 6013 | if (ret) { |
4e88d6e7 | 6014 | req_set_fail_links(req); |
e1e16097 | 6015 | io_req_complete(req, ret); |
edafccee | 6016 | } |
2b188cc1 | 6017 | |
f4db7182 | 6018 | return io_steal_work(req); |
2b188cc1 JA |
6019 | } |
6020 | ||
65e19f54 JA |
6021 | static inline struct file *io_file_from_index(struct io_ring_ctx *ctx, |
6022 | int index) | |
6023 | { | |
6024 | struct fixed_file_table *table; | |
6025 | ||
05f3fb3c | 6026 | table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT]; |
84695089 | 6027 | return table->files[index & IORING_FILE_TABLE_MASK]; |
65e19f54 JA |
6028 | } |
6029 | ||
8da11c19 PB |
6030 | static int io_file_get(struct io_submit_state *state, struct io_kiocb *req, |
6031 | int fd, struct file **out_file, bool fixed) | |
09bb8394 | 6032 | { |
a197f664 | 6033 | struct io_ring_ctx *ctx = req->ctx; |
8da11c19 | 6034 | struct file *file; |
09bb8394 | 6035 | |
8da11c19 | 6036 | if (fixed) { |
05f3fb3c | 6037 | if (unlikely(!ctx->file_data || |
09bb8394 JA |
6038 | (unsigned) fd >= ctx->nr_user_files)) |
6039 | return -EBADF; | |
b7620121 | 6040 | fd = array_index_nospec(fd, ctx->nr_user_files); |
8da11c19 | 6041 | file = io_file_from_index(ctx, fd); |
fd2206e4 JA |
6042 | if (file) { |
6043 | req->fixed_file_refs = ctx->file_data->cur_refs; | |
6044 | percpu_ref_get(req->fixed_file_refs); | |
6045 | } | |
09bb8394 | 6046 | } else { |
c826bd7a | 6047 | trace_io_uring_file_get(ctx, fd); |
8da11c19 | 6048 | file = __io_file_get(state, fd); |
09bb8394 JA |
6049 | } |
6050 | ||
fd2206e4 JA |
6051 | if (file || io_op_defs[req->opcode].needs_file_no_error) { |
6052 | *out_file = file; | |
6053 | return 0; | |
6054 | } | |
6055 | return -EBADF; | |
09bb8394 JA |
6056 | } |
6057 | ||
8da11c19 | 6058 | static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req, |
63ff8223 | 6059 | int fd) |
8da11c19 | 6060 | { |
8da11c19 PB |
6061 | bool fixed; |
6062 | ||
63ff8223 | 6063 | fixed = (req->flags & REQ_F_FIXED_FILE) != 0; |
0cdaf760 | 6064 | if (unlikely(!fixed && io_async_submit(req->ctx))) |
8da11c19 PB |
6065 | return -EBADF; |
6066 | ||
6067 | return io_file_get(state, req, fd, &req->file, fixed); | |
6068 | } | |
6069 | ||
a197f664 | 6070 | static int io_grab_files(struct io_kiocb *req) |
fcb323cc JA |
6071 | { |
6072 | int ret = -EBADF; | |
a197f664 | 6073 | struct io_ring_ctx *ctx = req->ctx; |
fcb323cc | 6074 | |
f56040b8 PB |
6075 | io_req_init_async(req); |
6076 | ||
5b0bbee4 | 6077 | if (req->work.files || (req->flags & REQ_F_NO_FILE_TABLE)) |
f86cd20c | 6078 | return 0; |
b14cca0c | 6079 | if (!ctx->ring_file) |
b5dba59e JA |
6080 | return -EBADF; |
6081 | ||
fcb323cc JA |
6082 | rcu_read_lock(); |
6083 | spin_lock_irq(&ctx->inflight_lock); | |
6084 | /* | |
6085 | * We use the f_ops->flush() handler to ensure that we can flush | |
6086 | * out work accessing these files if the fd is closed. Check if | |
6087 | * the fd has changed since we started down this path, and disallow | |
6088 | * this operation if it has. | |
6089 | */ | |
b14cca0c | 6090 | if (fcheck(ctx->ring_fd) == ctx->ring_file) { |
fcb323cc JA |
6091 | list_add(&req->inflight_entry, &ctx->inflight_list); |
6092 | req->flags |= REQ_F_INFLIGHT; | |
6093 | req->work.files = current->files; | |
6094 | ret = 0; | |
6095 | } | |
6096 | spin_unlock_irq(&ctx->inflight_lock); | |
6097 | rcu_read_unlock(); | |
6098 | ||
6099 | return ret; | |
6100 | } | |
6101 | ||
f56040b8 PB |
6102 | static inline int io_prep_work_files(struct io_kiocb *req) |
6103 | { | |
6104 | if (!io_op_defs[req->opcode].file_table) | |
6105 | return 0; | |
6106 | return io_grab_files(req); | |
6107 | } | |
6108 | ||
2665abfd | 6109 | static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer) |
2b188cc1 | 6110 | { |
ad8a48ac JA |
6111 | struct io_timeout_data *data = container_of(timer, |
6112 | struct io_timeout_data, timer); | |
6113 | struct io_kiocb *req = data->req; | |
2665abfd JA |
6114 | struct io_ring_ctx *ctx = req->ctx; |
6115 | struct io_kiocb *prev = NULL; | |
6116 | unsigned long flags; | |
2665abfd JA |
6117 | |
6118 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
6119 | ||
6120 | /* | |
6121 | * We don't expect the list to be empty, that will only happen if we | |
6122 | * race with the completion of the linked work. | |
6123 | */ | |
4493233e PB |
6124 | if (!list_empty(&req->link_list)) { |
6125 | prev = list_entry(req->link_list.prev, struct io_kiocb, | |
6126 | link_list); | |
5d960724 | 6127 | if (refcount_inc_not_zero(&prev->refs)) { |
4493233e | 6128 | list_del_init(&req->link_list); |
5d960724 JA |
6129 | prev->flags &= ~REQ_F_LINK_TIMEOUT; |
6130 | } else | |
76a46e06 | 6131 | prev = NULL; |
2665abfd JA |
6132 | } |
6133 | ||
6134 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
6135 | ||
6136 | if (prev) { | |
4e88d6e7 | 6137 | req_set_fail_links(prev); |
014db007 | 6138 | io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME); |
76a46e06 | 6139 | io_put_req(prev); |
47f46768 | 6140 | } else { |
e1e16097 | 6141 | io_req_complete(req, -ETIME); |
2665abfd | 6142 | } |
2665abfd JA |
6143 | return HRTIMER_NORESTART; |
6144 | } | |
6145 | ||
7271ef3a | 6146 | static void __io_queue_linked_timeout(struct io_kiocb *req) |
2665abfd | 6147 | { |
76a46e06 JA |
6148 | /* |
6149 | * If the list is now empty, then our linked request finished before | |
6150 | * we got a chance to setup the timer | |
6151 | */ | |
4493233e | 6152 | if (!list_empty(&req->link_list)) { |
2d28390a | 6153 | struct io_timeout_data *data = &req->io->timeout; |
94ae5e77 | 6154 | |
ad8a48ac JA |
6155 | data->timer.function = io_link_timeout_fn; |
6156 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), | |
6157 | data->mode); | |
2665abfd | 6158 | } |
7271ef3a JA |
6159 | } |
6160 | ||
6161 | static void io_queue_linked_timeout(struct io_kiocb *req) | |
6162 | { | |
6163 | struct io_ring_ctx *ctx = req->ctx; | |
6164 | ||
6165 | spin_lock_irq(&ctx->completion_lock); | |
6166 | __io_queue_linked_timeout(req); | |
76a46e06 | 6167 | spin_unlock_irq(&ctx->completion_lock); |
2665abfd | 6168 | |
2665abfd | 6169 | /* drop submission reference */ |
76a46e06 JA |
6170 | io_put_req(req); |
6171 | } | |
2665abfd | 6172 | |
ad8a48ac | 6173 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) |
2665abfd JA |
6174 | { |
6175 | struct io_kiocb *nxt; | |
6176 | ||
dea3b49c | 6177 | if (!(req->flags & REQ_F_LINK_HEAD)) |
2665abfd | 6178 | return NULL; |
6df1db6b | 6179 | if (req->flags & REQ_F_LINK_TIMEOUT) |
d7718a9d | 6180 | return NULL; |
2665abfd | 6181 | |
4493233e PB |
6182 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, |
6183 | link_list); | |
d625c6ee | 6184 | if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT) |
76a46e06 | 6185 | return NULL; |
2665abfd | 6186 | |
76a46e06 | 6187 | req->flags |= REQ_F_LINK_TIMEOUT; |
76a46e06 | 6188 | return nxt; |
2665abfd JA |
6189 | } |
6190 | ||
f13fad7b JA |
6191 | static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
6192 | struct io_comp_state *cs) | |
2b188cc1 | 6193 | { |
4a0a7a18 | 6194 | struct io_kiocb *linked_timeout; |
4bc4494e | 6195 | struct io_kiocb *nxt; |
193155c8 | 6196 | const struct cred *old_creds = NULL; |
e0c5c576 | 6197 | int ret; |
2b188cc1 | 6198 | |
4a0a7a18 JA |
6199 | again: |
6200 | linked_timeout = io_prep_linked_timeout(req); | |
6201 | ||
7cdaf587 XW |
6202 | if ((req->flags & REQ_F_WORK_INITIALIZED) && req->work.creds && |
6203 | req->work.creds != current_cred()) { | |
193155c8 JA |
6204 | if (old_creds) |
6205 | revert_creds(old_creds); | |
6206 | if (old_creds == req->work.creds) | |
6207 | old_creds = NULL; /* restored original creds */ | |
6208 | else | |
6209 | old_creds = override_creds(req->work.creds); | |
6210 | } | |
6211 | ||
f13fad7b | 6212 | ret = io_issue_sqe(req, sqe, true, cs); |
491381ce JA |
6213 | |
6214 | /* | |
6215 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
6216 | * doesn't support non-blocking read/write attempts | |
6217 | */ | |
24c74678 | 6218 | if (ret == -EAGAIN && !(req->flags & REQ_F_NOWAIT)) { |
f063c547 | 6219 | if (!io_arm_poll_handler(req)) { |
86a761f8 | 6220 | punt: |
f063c547 PB |
6221 | ret = io_prep_work_files(req); |
6222 | if (unlikely(ret)) | |
bbad27b2 | 6223 | goto err; |
f063c547 PB |
6224 | /* |
6225 | * Queued up for async execution, worker will release | |
6226 | * submit reference when the iocb is actually submitted. | |
6227 | */ | |
6228 | io_queue_async_work(req); | |
2b188cc1 | 6229 | } |
bbad27b2 | 6230 | |
f063c547 PB |
6231 | if (linked_timeout) |
6232 | io_queue_linked_timeout(linked_timeout); | |
4bc4494e | 6233 | goto exit; |
2b188cc1 | 6234 | } |
e65ef56d | 6235 | |
652532ad | 6236 | if (unlikely(ret)) { |
fcb323cc | 6237 | err: |
652532ad PB |
6238 | /* un-prep timeout, so it'll be killed as any other linked */ |
6239 | req->flags &= ~REQ_F_LINK_TIMEOUT; | |
4e88d6e7 | 6240 | req_set_fail_links(req); |
e65ef56d | 6241 | io_put_req(req); |
e1e16097 | 6242 | io_req_complete(req, ret); |
652532ad | 6243 | goto exit; |
9e645e11 | 6244 | } |
652532ad PB |
6245 | |
6246 | /* drop submission reference */ | |
6247 | nxt = io_put_req_find_next(req); | |
6248 | if (linked_timeout) | |
6249 | io_queue_linked_timeout(linked_timeout); | |
6250 | ||
4a0a7a18 JA |
6251 | if (nxt) { |
6252 | req = nxt; | |
86a761f8 PB |
6253 | |
6254 | if (req->flags & REQ_F_FORCE_ASYNC) | |
6255 | goto punt; | |
4a0a7a18 JA |
6256 | goto again; |
6257 | } | |
4bc4494e | 6258 | exit: |
193155c8 JA |
6259 | if (old_creds) |
6260 | revert_creds(old_creds); | |
2b188cc1 JA |
6261 | } |
6262 | ||
f13fad7b JA |
6263 | static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
6264 | struct io_comp_state *cs) | |
4fe2c963 JL |
6265 | { |
6266 | int ret; | |
6267 | ||
3529d8c2 | 6268 | ret = io_req_defer(req, sqe); |
4fe2c963 JL |
6269 | if (ret) { |
6270 | if (ret != -EIOCBQUEUED) { | |
1118591a | 6271 | fail_req: |
4e88d6e7 | 6272 | req_set_fail_links(req); |
e1e16097 JA |
6273 | io_put_req(req); |
6274 | io_req_complete(req, ret); | |
4fe2c963 | 6275 | } |
2550878f | 6276 | } else if (req->flags & REQ_F_FORCE_ASYNC) { |
bd2ab18a | 6277 | if (!req->io) { |
bd2ab18a | 6278 | ret = io_req_defer_prep(req, sqe); |
327d6d96 | 6279 | if (unlikely(ret)) |
bd2ab18a PB |
6280 | goto fail_req; |
6281 | } | |
6282 | ||
ce35a47a JA |
6283 | /* |
6284 | * Never try inline submit of IOSQE_ASYNC is set, go straight | |
6285 | * to async execution. | |
6286 | */ | |
3e863ea3 | 6287 | io_req_init_async(req); |
ce35a47a JA |
6288 | req->work.flags |= IO_WQ_WORK_CONCURRENT; |
6289 | io_queue_async_work(req); | |
6290 | } else { | |
f13fad7b | 6291 | __io_queue_sqe(req, sqe, cs); |
ce35a47a | 6292 | } |
4fe2c963 JL |
6293 | } |
6294 | ||
f13fad7b JA |
6295 | static inline void io_queue_link_head(struct io_kiocb *req, |
6296 | struct io_comp_state *cs) | |
4fe2c963 | 6297 | { |
94ae5e77 | 6298 | if (unlikely(req->flags & REQ_F_FAIL_LINK)) { |
e1e16097 JA |
6299 | io_put_req(req); |
6300 | io_req_complete(req, -ECANCELED); | |
1b4a51b6 | 6301 | } else |
f13fad7b | 6302 | io_queue_sqe(req, NULL, cs); |
4fe2c963 JL |
6303 | } |
6304 | ||
1d4240cc | 6305 | static int io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
f13fad7b | 6306 | struct io_kiocb **link, struct io_comp_state *cs) |
9e645e11 | 6307 | { |
a197f664 | 6308 | struct io_ring_ctx *ctx = req->ctx; |
ef4ff581 | 6309 | int ret; |
9e645e11 | 6310 | |
9e645e11 JA |
6311 | /* |
6312 | * If we already have a head request, queue this one for async | |
6313 | * submittal once the head completes. If we don't have a head but | |
6314 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
6315 | * submitted sync once the chain is complete. If none of those | |
6316 | * conditions are true (normal request), then just queue it. | |
6317 | */ | |
6318 | if (*link) { | |
9d76377f | 6319 | struct io_kiocb *head = *link; |
4e88d6e7 | 6320 | |
8cdf2193 PB |
6321 | /* |
6322 | * Taking sequential execution of a link, draining both sides | |
6323 | * of the link also fullfils IOSQE_IO_DRAIN semantics for all | |
6324 | * requests in the link. So, it drains the head and the | |
6325 | * next after the link request. The last one is done via | |
6326 | * drain_next flag to persist the effect across calls. | |
6327 | */ | |
ef4ff581 | 6328 | if (req->flags & REQ_F_IO_DRAIN) { |
711be031 PB |
6329 | head->flags |= REQ_F_IO_DRAIN; |
6330 | ctx->drain_next = 1; | |
6331 | } | |
3529d8c2 | 6332 | ret = io_req_defer_prep(req, sqe); |
327d6d96 | 6333 | if (unlikely(ret)) { |
4e88d6e7 | 6334 | /* fail even hard links since we don't submit */ |
9d76377f | 6335 | head->flags |= REQ_F_FAIL_LINK; |
1d4240cc | 6336 | return ret; |
2d28390a | 6337 | } |
9d76377f | 6338 | trace_io_uring_link(ctx, req, head); |
c40f6379 | 6339 | io_get_req_task(req); |
9d76377f | 6340 | list_add_tail(&req->link_list, &head->link_list); |
32fe525b PB |
6341 | |
6342 | /* last request of a link, enqueue the link */ | |
ef4ff581 | 6343 | if (!(req->flags & (REQ_F_LINK | REQ_F_HARDLINK))) { |
f13fad7b | 6344 | io_queue_link_head(head, cs); |
32fe525b PB |
6345 | *link = NULL; |
6346 | } | |
9e645e11 | 6347 | } else { |
711be031 PB |
6348 | if (unlikely(ctx->drain_next)) { |
6349 | req->flags |= REQ_F_IO_DRAIN; | |
ef4ff581 | 6350 | ctx->drain_next = 0; |
711be031 | 6351 | } |
ef4ff581 | 6352 | if (req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) { |
dea3b49c | 6353 | req->flags |= REQ_F_LINK_HEAD; |
711be031 | 6354 | INIT_LIST_HEAD(&req->link_list); |
f1d96a8f | 6355 | |
711be031 | 6356 | ret = io_req_defer_prep(req, sqe); |
327d6d96 | 6357 | if (unlikely(ret)) |
711be031 PB |
6358 | req->flags |= REQ_F_FAIL_LINK; |
6359 | *link = req; | |
6360 | } else { | |
f13fad7b | 6361 | io_queue_sqe(req, sqe, cs); |
711be031 | 6362 | } |
9e645e11 | 6363 | } |
2e6e1fde | 6364 | |
1d4240cc | 6365 | return 0; |
9e645e11 JA |
6366 | } |
6367 | ||
9a56a232 JA |
6368 | /* |
6369 | * Batched submission is done, ensure local IO is flushed out. | |
6370 | */ | |
6371 | static void io_submit_state_end(struct io_submit_state *state) | |
6372 | { | |
f13fad7b JA |
6373 | if (!list_empty(&state->comp.list)) |
6374 | io_submit_flush_completions(&state->comp); | |
9a56a232 | 6375 | blk_finish_plug(&state->plug); |
9f13c35b | 6376 | io_state_file_put(state); |
2579f913 | 6377 | if (state->free_reqs) |
6c8a3134 | 6378 | kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs); |
9a56a232 JA |
6379 | } |
6380 | ||
6381 | /* | |
6382 | * Start submission side cache. | |
6383 | */ | |
6384 | static void io_submit_state_start(struct io_submit_state *state, | |
013538bd | 6385 | struct io_ring_ctx *ctx, unsigned int max_ios) |
9a56a232 JA |
6386 | { |
6387 | blk_start_plug(&state->plug); | |
013538bd JA |
6388 | state->comp.nr = 0; |
6389 | INIT_LIST_HEAD(&state->comp.list); | |
6390 | state->comp.ctx = ctx; | |
2579f913 | 6391 | state->free_reqs = 0; |
9a56a232 JA |
6392 | state->file = NULL; |
6393 | state->ios_left = max_ios; | |
6394 | } | |
6395 | ||
2b188cc1 JA |
6396 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
6397 | { | |
75b28aff | 6398 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 6399 | |
caf582c6 PB |
6400 | /* |
6401 | * Ensure any loads from the SQEs are done at this point, | |
6402 | * since once we write the new head, the application could | |
6403 | * write new data to them. | |
6404 | */ | |
6405 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
6406 | } |
6407 | ||
2b188cc1 | 6408 | /* |
3529d8c2 | 6409 | * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory |
2b188cc1 JA |
6410 | * that is mapped by userspace. This means that care needs to be taken to |
6411 | * ensure that reads are stable, as we cannot rely on userspace always | |
6412 | * being a good citizen. If members of the sqe are validated and then later | |
6413 | * used, it's important that those reads are done through READ_ONCE() to | |
6414 | * prevent a re-load down the line. | |
6415 | */ | |
709b302f | 6416 | static const struct io_uring_sqe *io_get_sqe(struct io_ring_ctx *ctx) |
2b188cc1 | 6417 | { |
75b28aff | 6418 | u32 *sq_array = ctx->sq_array; |
2b188cc1 JA |
6419 | unsigned head; |
6420 | ||
6421 | /* | |
6422 | * The cached sq head (or cq tail) serves two purposes: | |
6423 | * | |
6424 | * 1) allows us to batch the cost of updating the user visible | |
6425 | * head updates. | |
6426 | * 2) allows the kernel side to track the head on its own, even | |
6427 | * though the application is the one updating it. | |
6428 | */ | |
ee7d46d9 | 6429 | head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]); |
709b302f PB |
6430 | if (likely(head < ctx->sq_entries)) |
6431 | return &ctx->sq_sqes[head]; | |
2b188cc1 JA |
6432 | |
6433 | /* drop invalid entries */ | |
498ccd9e | 6434 | ctx->cached_sq_dropped++; |
ee7d46d9 | 6435 | WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped); |
709b302f PB |
6436 | return NULL; |
6437 | } | |
6438 | ||
6439 | static inline void io_consume_sqe(struct io_ring_ctx *ctx) | |
6440 | { | |
6441 | ctx->cached_sq_head++; | |
2b188cc1 JA |
6442 | } |
6443 | ||
ef4ff581 PB |
6444 | #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \ |
6445 | IOSQE_IO_HARDLINK | IOSQE_ASYNC | \ | |
6446 | IOSQE_BUFFER_SELECT) | |
6447 | ||
6448 | static int io_init_req(struct io_ring_ctx *ctx, struct io_kiocb *req, | |
6449 | const struct io_uring_sqe *sqe, | |
0cdaf760 | 6450 | struct io_submit_state *state) |
0553b8bd | 6451 | { |
ef4ff581 | 6452 | unsigned int sqe_flags; |
63ff8223 | 6453 | int id; |
ef4ff581 | 6454 | |
0553b8bd PB |
6455 | req->opcode = READ_ONCE(sqe->opcode); |
6456 | req->user_data = READ_ONCE(sqe->user_data); | |
6457 | req->io = NULL; | |
6458 | req->file = NULL; | |
6459 | req->ctx = ctx; | |
6460 | req->flags = 0; | |
6461 | /* one is dropped after submission, the other at completion */ | |
6462 | refcount_set(&req->refs, 2); | |
4dd2824d | 6463 | req->task = current; |
0553b8bd | 6464 | req->result = 0; |
ef4ff581 PB |
6465 | |
6466 | if (unlikely(req->opcode >= IORING_OP_LAST)) | |
6467 | return -EINVAL; | |
6468 | ||
9d8426a0 JA |
6469 | if (unlikely(io_sq_thread_acquire_mm(ctx, req))) |
6470 | return -EFAULT; | |
ef4ff581 PB |
6471 | |
6472 | sqe_flags = READ_ONCE(sqe->flags); | |
6473 | /* enforce forwards compatibility on users */ | |
6474 | if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) | |
6475 | return -EINVAL; | |
6476 | ||
6477 | if ((sqe_flags & IOSQE_BUFFER_SELECT) && | |
6478 | !io_op_defs[req->opcode].buffer_select) | |
6479 | return -EOPNOTSUPP; | |
6480 | ||
6481 | id = READ_ONCE(sqe->personality); | |
6482 | if (id) { | |
7cdaf587 | 6483 | io_req_init_async(req); |
ef4ff581 PB |
6484 | req->work.creds = idr_find(&ctx->personality_idr, id); |
6485 | if (unlikely(!req->work.creds)) | |
6486 | return -EINVAL; | |
6487 | get_cred(req->work.creds); | |
6488 | } | |
6489 | ||
6490 | /* same numerical values with corresponding REQ_F_*, safe to copy */ | |
c11368a5 | 6491 | req->flags |= sqe_flags; |
ef4ff581 | 6492 | |
63ff8223 JA |
6493 | if (!io_op_defs[req->opcode].needs_file) |
6494 | return 0; | |
6495 | ||
6496 | return io_req_set_file(state, req, READ_ONCE(sqe->fd)); | |
0553b8bd PB |
6497 | } |
6498 | ||
fb5ccc98 | 6499 | static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr, |
0cdaf760 | 6500 | struct file *ring_file, int ring_fd) |
6c271ce2 | 6501 | { |
ac8691c4 | 6502 | struct io_submit_state state; |
9e645e11 | 6503 | struct io_kiocb *link = NULL; |
9e645e11 | 6504 | int i, submitted = 0; |
6c271ce2 | 6505 | |
c4a2ed72 | 6506 | /* if we have a backlog and couldn't flush it all, return BUSY */ |
ad3eb2c8 JA |
6507 | if (test_bit(0, &ctx->sq_check_overflow)) { |
6508 | if (!list_empty(&ctx->cq_overflow_list) && | |
6509 | !io_cqring_overflow_flush(ctx, false)) | |
6510 | return -EBUSY; | |
6511 | } | |
6c271ce2 | 6512 | |
ee7d46d9 PB |
6513 | /* make sure SQ entry isn't read before tail */ |
6514 | nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx)); | |
9ef4f124 | 6515 | |
2b85edfc PB |
6516 | if (!percpu_ref_tryget_many(&ctx->refs, nr)) |
6517 | return -EAGAIN; | |
6c271ce2 | 6518 | |
013538bd | 6519 | io_submit_state_start(&state, ctx, nr); |
6c271ce2 | 6520 | |
b14cca0c PB |
6521 | ctx->ring_fd = ring_fd; |
6522 | ctx->ring_file = ring_file; | |
6523 | ||
6c271ce2 | 6524 | for (i = 0; i < nr; i++) { |
3529d8c2 | 6525 | const struct io_uring_sqe *sqe; |
196be95c | 6526 | struct io_kiocb *req; |
1cb1edb2 | 6527 | int err; |
fb5ccc98 | 6528 | |
b1e50e54 PB |
6529 | sqe = io_get_sqe(ctx); |
6530 | if (unlikely(!sqe)) { | |
6531 | io_consume_sqe(ctx); | |
6532 | break; | |
6533 | } | |
ac8691c4 | 6534 | req = io_alloc_req(ctx, &state); |
196be95c PB |
6535 | if (unlikely(!req)) { |
6536 | if (!submitted) | |
6537 | submitted = -EAGAIN; | |
fb5ccc98 | 6538 | break; |
196be95c | 6539 | } |
fb5ccc98 | 6540 | |
ac8691c4 | 6541 | err = io_init_req(ctx, req, sqe, &state); |
709b302f | 6542 | io_consume_sqe(ctx); |
d3656344 JA |
6543 | /* will complete beyond this point, count as submitted */ |
6544 | submitted++; | |
6545 | ||
ef4ff581 | 6546 | if (unlikely(err)) { |
1cb1edb2 | 6547 | fail_req: |
e1e16097 JA |
6548 | io_put_req(req); |
6549 | io_req_complete(req, err); | |
196be95c PB |
6550 | break; |
6551 | } | |
fb5ccc98 | 6552 | |
354420f7 | 6553 | trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data, |
0cdaf760 | 6554 | true, io_async_submit(ctx)); |
f13fad7b | 6555 | err = io_submit_sqe(req, sqe, &link, &state.comp); |
1d4240cc PB |
6556 | if (err) |
6557 | goto fail_req; | |
6c271ce2 JA |
6558 | } |
6559 | ||
9466f437 PB |
6560 | if (unlikely(submitted != nr)) { |
6561 | int ref_used = (submitted == -EAGAIN) ? 0 : submitted; | |
6562 | ||
6563 | percpu_ref_put_many(&ctx->refs, nr - ref_used); | |
6564 | } | |
9e645e11 | 6565 | if (link) |
f13fad7b | 6566 | io_queue_link_head(link, &state.comp); |
ac8691c4 | 6567 | io_submit_state_end(&state); |
6c271ce2 | 6568 | |
ae9428ca PB |
6569 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
6570 | io_commit_sqring(ctx); | |
6571 | ||
6c271ce2 JA |
6572 | return submitted; |
6573 | } | |
6574 | ||
23b3628e XW |
6575 | static inline void io_ring_set_wakeup_flag(struct io_ring_ctx *ctx) |
6576 | { | |
6577 | /* Tell userspace we may need a wakeup call */ | |
6578 | spin_lock_irq(&ctx->completion_lock); | |
6579 | ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP; | |
6580 | spin_unlock_irq(&ctx->completion_lock); | |
6581 | } | |
6582 | ||
6583 | static inline void io_ring_clear_wakeup_flag(struct io_ring_ctx *ctx) | |
6584 | { | |
6585 | spin_lock_irq(&ctx->completion_lock); | |
6586 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; | |
6587 | spin_unlock_irq(&ctx->completion_lock); | |
6588 | } | |
6589 | ||
6c271ce2 JA |
6590 | static int io_sq_thread(void *data) |
6591 | { | |
6c271ce2 | 6592 | struct io_ring_ctx *ctx = data; |
181e448d | 6593 | const struct cred *old_cred; |
6c271ce2 | 6594 | DEFINE_WAIT(wait); |
6c271ce2 | 6595 | unsigned long timeout; |
bdcd3eab | 6596 | int ret = 0; |
6c271ce2 | 6597 | |
0f158b4c | 6598 | complete(&ctx->sq_thread_comp); |
a4c0b3de | 6599 | |
181e448d | 6600 | old_cred = override_creds(ctx->creds); |
6c271ce2 | 6601 | |
bdcd3eab | 6602 | timeout = jiffies + ctx->sq_thread_idle; |
2bbcd6d3 | 6603 | while (!kthread_should_park()) { |
fb5ccc98 | 6604 | unsigned int to_submit; |
6c271ce2 | 6605 | |
540e32a0 | 6606 | if (!list_empty(&ctx->iopoll_list)) { |
6c271ce2 JA |
6607 | unsigned nr_events = 0; |
6608 | ||
bdcd3eab | 6609 | mutex_lock(&ctx->uring_lock); |
540e32a0 | 6610 | if (!list_empty(&ctx->iopoll_list) && !need_resched()) |
9dedd563 | 6611 | io_do_iopoll(ctx, &nr_events, 0); |
bdcd3eab | 6612 | else |
6c271ce2 | 6613 | timeout = jiffies + ctx->sq_thread_idle; |
bdcd3eab | 6614 | mutex_unlock(&ctx->uring_lock); |
6c271ce2 JA |
6615 | } |
6616 | ||
fb5ccc98 | 6617 | to_submit = io_sqring_entries(ctx); |
c1edbf5f JA |
6618 | |
6619 | /* | |
6620 | * If submit got -EBUSY, flag us as needing the application | |
6621 | * to enter the kernel to reap and flush events. | |
6622 | */ | |
b772f07a | 6623 | if (!to_submit || ret == -EBUSY || need_resched()) { |
7143b5ac SG |
6624 | /* |
6625 | * Drop cur_mm before scheduling, we can't hold it for | |
6626 | * long periods (or over schedule()). Do this before | |
6627 | * adding ourselves to the waitqueue, as the unuse/drop | |
6628 | * may sleep. | |
6629 | */ | |
4349f30e | 6630 | io_sq_thread_drop_mm(); |
7143b5ac | 6631 | |
6c271ce2 JA |
6632 | /* |
6633 | * We're polling. If we're within the defined idle | |
6634 | * period, then let us spin without work before going | |
c1edbf5f JA |
6635 | * to sleep. The exception is if we got EBUSY doing |
6636 | * more IO, we should wait for the application to | |
6637 | * reap events and wake us up. | |
6c271ce2 | 6638 | */ |
540e32a0 | 6639 | if (!list_empty(&ctx->iopoll_list) || need_resched() || |
df069d80 JA |
6640 | (!time_after(jiffies, timeout) && ret != -EBUSY && |
6641 | !percpu_ref_is_dying(&ctx->refs))) { | |
4c6e277c | 6642 | io_run_task_work(); |
9831a90c | 6643 | cond_resched(); |
6c271ce2 JA |
6644 | continue; |
6645 | } | |
6646 | ||
6c271ce2 JA |
6647 | prepare_to_wait(&ctx->sqo_wait, &wait, |
6648 | TASK_INTERRUPTIBLE); | |
6649 | ||
bdcd3eab XW |
6650 | /* |
6651 | * While doing polled IO, before going to sleep, we need | |
540e32a0 PB |
6652 | * to check if there are new reqs added to iopoll_list, |
6653 | * it is because reqs may have been punted to io worker | |
6654 | * and will be added to iopoll_list later, hence check | |
6655 | * the iopoll_list again. | |
bdcd3eab XW |
6656 | */ |
6657 | if ((ctx->flags & IORING_SETUP_IOPOLL) && | |
540e32a0 | 6658 | !list_empty_careful(&ctx->iopoll_list)) { |
bdcd3eab XW |
6659 | finish_wait(&ctx->sqo_wait, &wait); |
6660 | continue; | |
6661 | } | |
6662 | ||
23b3628e | 6663 | io_ring_set_wakeup_flag(ctx); |
6c271ce2 | 6664 | |
fb5ccc98 | 6665 | to_submit = io_sqring_entries(ctx); |
c1edbf5f | 6666 | if (!to_submit || ret == -EBUSY) { |
2bbcd6d3 | 6667 | if (kthread_should_park()) { |
6c271ce2 JA |
6668 | finish_wait(&ctx->sqo_wait, &wait); |
6669 | break; | |
6670 | } | |
4c6e277c | 6671 | if (io_run_task_work()) { |
10bea96d | 6672 | finish_wait(&ctx->sqo_wait, &wait); |
23b3628e | 6673 | io_ring_clear_wakeup_flag(ctx); |
b41e9852 JA |
6674 | continue; |
6675 | } | |
6c271ce2 JA |
6676 | if (signal_pending(current)) |
6677 | flush_signals(current); | |
6678 | schedule(); | |
6679 | finish_wait(&ctx->sqo_wait, &wait); | |
6680 | ||
23b3628e | 6681 | io_ring_clear_wakeup_flag(ctx); |
d4ae271d | 6682 | ret = 0; |
6c271ce2 JA |
6683 | continue; |
6684 | } | |
6685 | finish_wait(&ctx->sqo_wait, &wait); | |
6686 | ||
23b3628e | 6687 | io_ring_clear_wakeup_flag(ctx); |
6c271ce2 JA |
6688 | } |
6689 | ||
8a4955ff | 6690 | mutex_lock(&ctx->uring_lock); |
6b668c9b XW |
6691 | if (likely(!percpu_ref_is_dying(&ctx->refs))) |
6692 | ret = io_submit_sqes(ctx, to_submit, NULL, -1); | |
8a4955ff | 6693 | mutex_unlock(&ctx->uring_lock); |
bdcd3eab | 6694 | timeout = jiffies + ctx->sq_thread_idle; |
6c271ce2 JA |
6695 | } |
6696 | ||
4c6e277c | 6697 | io_run_task_work(); |
b41e9852 | 6698 | |
4349f30e | 6699 | io_sq_thread_drop_mm(); |
181e448d | 6700 | revert_creds(old_cred); |
06058632 | 6701 | |
2bbcd6d3 | 6702 | kthread_parkme(); |
06058632 | 6703 | |
6c271ce2 JA |
6704 | return 0; |
6705 | } | |
6706 | ||
bda52162 JA |
6707 | struct io_wait_queue { |
6708 | struct wait_queue_entry wq; | |
6709 | struct io_ring_ctx *ctx; | |
6710 | unsigned to_wait; | |
6711 | unsigned nr_timeouts; | |
6712 | }; | |
6713 | ||
1d7bb1d5 | 6714 | static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush) |
bda52162 JA |
6715 | { |
6716 | struct io_ring_ctx *ctx = iowq->ctx; | |
6717 | ||
6718 | /* | |
d195a66e | 6719 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
6720 | * started waiting. For timeouts, we always want to return to userspace, |
6721 | * regardless of event count. | |
6722 | */ | |
1d7bb1d5 | 6723 | return io_cqring_events(ctx, noflush) >= iowq->to_wait || |
bda52162 JA |
6724 | atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
6725 | } | |
6726 | ||
6727 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
6728 | int wake_flags, void *key) | |
6729 | { | |
6730 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
6731 | wq); | |
6732 | ||
1d7bb1d5 JA |
6733 | /* use noflush == true, as we can't safely rely on locking context */ |
6734 | if (!io_should_wake(iowq, true)) | |
bda52162 JA |
6735 | return -1; |
6736 | ||
6737 | return autoremove_wake_function(curr, mode, wake_flags, key); | |
6738 | } | |
6739 | ||
2b188cc1 JA |
6740 | /* |
6741 | * Wait until events become available, if we don't already have some. The | |
6742 | * application must reap them itself, as they reside on the shared cq ring. | |
6743 | */ | |
6744 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
6745 | const sigset_t __user *sig, size_t sigsz) | |
6746 | { | |
bda52162 JA |
6747 | struct io_wait_queue iowq = { |
6748 | .wq = { | |
6749 | .private = current, | |
6750 | .func = io_wake_function, | |
6751 | .entry = LIST_HEAD_INIT(iowq.wq.entry), | |
6752 | }, | |
6753 | .ctx = ctx, | |
6754 | .to_wait = min_events, | |
6755 | }; | |
75b28aff | 6756 | struct io_rings *rings = ctx->rings; |
e9ffa5c2 | 6757 | int ret = 0; |
2b188cc1 | 6758 | |
b41e9852 JA |
6759 | do { |
6760 | if (io_cqring_events(ctx, false) >= min_events) | |
6761 | return 0; | |
4c6e277c | 6762 | if (!io_run_task_work()) |
b41e9852 | 6763 | break; |
b41e9852 | 6764 | } while (1); |
2b188cc1 JA |
6765 | |
6766 | if (sig) { | |
9e75ad5d AB |
6767 | #ifdef CONFIG_COMPAT |
6768 | if (in_compat_syscall()) | |
6769 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 6770 | sigsz); |
9e75ad5d AB |
6771 | else |
6772 | #endif | |
b772434b | 6773 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 6774 | |
2b188cc1 JA |
6775 | if (ret) |
6776 | return ret; | |
6777 | } | |
6778 | ||
bda52162 | 6779 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
c826bd7a | 6780 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 JA |
6781 | do { |
6782 | prepare_to_wait_exclusive(&ctx->wait, &iowq.wq, | |
6783 | TASK_INTERRUPTIBLE); | |
ce593a6c | 6784 | /* make sure we run task_work before checking for signals */ |
4c6e277c JA |
6785 | if (io_run_task_work()) |
6786 | continue; | |
bda52162 | 6787 | if (signal_pending(current)) { |
b7db41c9 JA |
6788 | if (current->jobctl & JOBCTL_TASK_WORK) { |
6789 | spin_lock_irq(¤t->sighand->siglock); | |
6790 | current->jobctl &= ~JOBCTL_TASK_WORK; | |
6791 | recalc_sigpending(); | |
6792 | spin_unlock_irq(¤t->sighand->siglock); | |
6793 | continue; | |
6794 | } | |
6795 | ret = -EINTR; | |
bda52162 JA |
6796 | break; |
6797 | } | |
ce593a6c JA |
6798 | if (io_should_wake(&iowq, false)) |
6799 | break; | |
6800 | schedule(); | |
bda52162 JA |
6801 | } while (1); |
6802 | finish_wait(&ctx->wait, &iowq.wq); | |
6803 | ||
b7db41c9 | 6804 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 6805 | |
75b28aff | 6806 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
6807 | } |
6808 | ||
6b06314c JA |
6809 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
6810 | { | |
6811 | #if defined(CONFIG_UNIX) | |
6812 | if (ctx->ring_sock) { | |
6813 | struct sock *sock = ctx->ring_sock->sk; | |
6814 | struct sk_buff *skb; | |
6815 | ||
6816 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
6817 | kfree_skb(skb); | |
6818 | } | |
6819 | #else | |
6820 | int i; | |
6821 | ||
65e19f54 JA |
6822 | for (i = 0; i < ctx->nr_user_files; i++) { |
6823 | struct file *file; | |
6824 | ||
6825 | file = io_file_from_index(ctx, i); | |
6826 | if (file) | |
6827 | fput(file); | |
6828 | } | |
6b06314c JA |
6829 | #endif |
6830 | } | |
6831 | ||
05f3fb3c JA |
6832 | static void io_file_ref_kill(struct percpu_ref *ref) |
6833 | { | |
6834 | struct fixed_file_data *data; | |
6835 | ||
6836 | data = container_of(ref, struct fixed_file_data, refs); | |
6837 | complete(&data->done); | |
6838 | } | |
6839 | ||
6b06314c JA |
6840 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) |
6841 | { | |
05f3fb3c | 6842 | struct fixed_file_data *data = ctx->file_data; |
05589553 | 6843 | struct fixed_file_ref_node *ref_node = NULL; |
65e19f54 JA |
6844 | unsigned nr_tables, i; |
6845 | ||
05f3fb3c | 6846 | if (!data) |
6b06314c JA |
6847 | return -ENXIO; |
6848 | ||
6a4d07cd | 6849 | spin_lock(&data->lock); |
05589553 XW |
6850 | if (!list_empty(&data->ref_list)) |
6851 | ref_node = list_first_entry(&data->ref_list, | |
6852 | struct fixed_file_ref_node, node); | |
6a4d07cd | 6853 | spin_unlock(&data->lock); |
05589553 XW |
6854 | if (ref_node) |
6855 | percpu_ref_kill(&ref_node->refs); | |
6856 | ||
6857 | percpu_ref_kill(&data->refs); | |
6858 | ||
6859 | /* wait for all refs nodes to complete */ | |
4a38aed2 | 6860 | flush_delayed_work(&ctx->file_put_work); |
2faf852d | 6861 | wait_for_completion(&data->done); |
05f3fb3c | 6862 | |
6b06314c | 6863 | __io_sqe_files_unregister(ctx); |
65e19f54 JA |
6864 | nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE); |
6865 | for (i = 0; i < nr_tables; i++) | |
05f3fb3c JA |
6866 | kfree(data->table[i].files); |
6867 | kfree(data->table); | |
05589553 XW |
6868 | percpu_ref_exit(&data->refs); |
6869 | kfree(data); | |
05f3fb3c | 6870 | ctx->file_data = NULL; |
6b06314c JA |
6871 | ctx->nr_user_files = 0; |
6872 | return 0; | |
6873 | } | |
6874 | ||
6c271ce2 JA |
6875 | static void io_sq_thread_stop(struct io_ring_ctx *ctx) |
6876 | { | |
6877 | if (ctx->sqo_thread) { | |
0f158b4c | 6878 | wait_for_completion(&ctx->sq_thread_comp); |
2bbcd6d3 RP |
6879 | /* |
6880 | * The park is a bit of a work-around, without it we get | |
6881 | * warning spews on shutdown with SQPOLL set and affinity | |
6882 | * set to a single CPU. | |
6883 | */ | |
06058632 | 6884 | kthread_park(ctx->sqo_thread); |
6c271ce2 JA |
6885 | kthread_stop(ctx->sqo_thread); |
6886 | ctx->sqo_thread = NULL; | |
6887 | } | |
6888 | } | |
6889 | ||
6b06314c JA |
6890 | static void io_finish_async(struct io_ring_ctx *ctx) |
6891 | { | |
6c271ce2 JA |
6892 | io_sq_thread_stop(ctx); |
6893 | ||
561fb04a JA |
6894 | if (ctx->io_wq) { |
6895 | io_wq_destroy(ctx->io_wq); | |
6896 | ctx->io_wq = NULL; | |
6b06314c JA |
6897 | } |
6898 | } | |
6899 | ||
6900 | #if defined(CONFIG_UNIX) | |
6b06314c JA |
6901 | /* |
6902 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
6903 | * the io_uring can be safely unregistered on process exit, even if we have | |
6904 | * loops in the file referencing. | |
6905 | */ | |
6906 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
6907 | { | |
6908 | struct sock *sk = ctx->ring_sock->sk; | |
6909 | struct scm_fp_list *fpl; | |
6910 | struct sk_buff *skb; | |
08a45173 | 6911 | int i, nr_files; |
6b06314c | 6912 | |
6b06314c JA |
6913 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); |
6914 | if (!fpl) | |
6915 | return -ENOMEM; | |
6916 | ||
6917 | skb = alloc_skb(0, GFP_KERNEL); | |
6918 | if (!skb) { | |
6919 | kfree(fpl); | |
6920 | return -ENOMEM; | |
6921 | } | |
6922 | ||
6923 | skb->sk = sk; | |
6b06314c | 6924 | |
08a45173 | 6925 | nr_files = 0; |
6b06314c JA |
6926 | fpl->user = get_uid(ctx->user); |
6927 | for (i = 0; i < nr; i++) { | |
65e19f54 JA |
6928 | struct file *file = io_file_from_index(ctx, i + offset); |
6929 | ||
6930 | if (!file) | |
08a45173 | 6931 | continue; |
65e19f54 | 6932 | fpl->fp[nr_files] = get_file(file); |
08a45173 JA |
6933 | unix_inflight(fpl->user, fpl->fp[nr_files]); |
6934 | nr_files++; | |
6b06314c JA |
6935 | } |
6936 | ||
08a45173 JA |
6937 | if (nr_files) { |
6938 | fpl->max = SCM_MAX_FD; | |
6939 | fpl->count = nr_files; | |
6940 | UNIXCB(skb).fp = fpl; | |
05f3fb3c | 6941 | skb->destructor = unix_destruct_scm; |
08a45173 JA |
6942 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); |
6943 | skb_queue_head(&sk->sk_receive_queue, skb); | |
6b06314c | 6944 | |
08a45173 JA |
6945 | for (i = 0; i < nr_files; i++) |
6946 | fput(fpl->fp[i]); | |
6947 | } else { | |
6948 | kfree_skb(skb); | |
6949 | kfree(fpl); | |
6950 | } | |
6b06314c JA |
6951 | |
6952 | return 0; | |
6953 | } | |
6954 | ||
6955 | /* | |
6956 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
6957 | * causes regular reference counting to break down. We rely on the UNIX | |
6958 | * garbage collection to take care of this problem for us. | |
6959 | */ | |
6960 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
6961 | { | |
6962 | unsigned left, total; | |
6963 | int ret = 0; | |
6964 | ||
6965 | total = 0; | |
6966 | left = ctx->nr_user_files; | |
6967 | while (left) { | |
6968 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
6969 | |
6970 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
6971 | if (ret) | |
6972 | break; | |
6973 | left -= this_files; | |
6974 | total += this_files; | |
6975 | } | |
6976 | ||
6977 | if (!ret) | |
6978 | return 0; | |
6979 | ||
6980 | while (total < ctx->nr_user_files) { | |
65e19f54 JA |
6981 | struct file *file = io_file_from_index(ctx, total); |
6982 | ||
6983 | if (file) | |
6984 | fput(file); | |
6b06314c JA |
6985 | total++; |
6986 | } | |
6987 | ||
6988 | return ret; | |
6989 | } | |
6990 | #else | |
6991 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
6992 | { | |
6993 | return 0; | |
6994 | } | |
6995 | #endif | |
6996 | ||
65e19f54 JA |
6997 | static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables, |
6998 | unsigned nr_files) | |
6999 | { | |
7000 | int i; | |
7001 | ||
7002 | for (i = 0; i < nr_tables; i++) { | |
05f3fb3c | 7003 | struct fixed_file_table *table = &ctx->file_data->table[i]; |
65e19f54 JA |
7004 | unsigned this_files; |
7005 | ||
7006 | this_files = min(nr_files, IORING_MAX_FILES_TABLE); | |
7007 | table->files = kcalloc(this_files, sizeof(struct file *), | |
7008 | GFP_KERNEL); | |
7009 | if (!table->files) | |
7010 | break; | |
7011 | nr_files -= this_files; | |
7012 | } | |
7013 | ||
7014 | if (i == nr_tables) | |
7015 | return 0; | |
7016 | ||
7017 | for (i = 0; i < nr_tables; i++) { | |
05f3fb3c | 7018 | struct fixed_file_table *table = &ctx->file_data->table[i]; |
65e19f54 JA |
7019 | kfree(table->files); |
7020 | } | |
7021 | return 1; | |
7022 | } | |
7023 | ||
05f3fb3c JA |
7024 | static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file) |
7025 | { | |
7026 | #if defined(CONFIG_UNIX) | |
7027 | struct sock *sock = ctx->ring_sock->sk; | |
7028 | struct sk_buff_head list, *head = &sock->sk_receive_queue; | |
7029 | struct sk_buff *skb; | |
7030 | int i; | |
7031 | ||
7032 | __skb_queue_head_init(&list); | |
7033 | ||
7034 | /* | |
7035 | * Find the skb that holds this file in its SCM_RIGHTS. When found, | |
7036 | * remove this entry and rearrange the file array. | |
7037 | */ | |
7038 | skb = skb_dequeue(head); | |
7039 | while (skb) { | |
7040 | struct scm_fp_list *fp; | |
7041 | ||
7042 | fp = UNIXCB(skb).fp; | |
7043 | for (i = 0; i < fp->count; i++) { | |
7044 | int left; | |
7045 | ||
7046 | if (fp->fp[i] != file) | |
7047 | continue; | |
7048 | ||
7049 | unix_notinflight(fp->user, fp->fp[i]); | |
7050 | left = fp->count - 1 - i; | |
7051 | if (left) { | |
7052 | memmove(&fp->fp[i], &fp->fp[i + 1], | |
7053 | left * sizeof(struct file *)); | |
7054 | } | |
7055 | fp->count--; | |
7056 | if (!fp->count) { | |
7057 | kfree_skb(skb); | |
7058 | skb = NULL; | |
7059 | } else { | |
7060 | __skb_queue_tail(&list, skb); | |
7061 | } | |
7062 | fput(file); | |
7063 | file = NULL; | |
7064 | break; | |
7065 | } | |
7066 | ||
7067 | if (!file) | |
7068 | break; | |
7069 | ||
7070 | __skb_queue_tail(&list, skb); | |
7071 | ||
7072 | skb = skb_dequeue(head); | |
7073 | } | |
7074 | ||
7075 | if (skb_peek(&list)) { | |
7076 | spin_lock_irq(&head->lock); | |
7077 | while ((skb = __skb_dequeue(&list)) != NULL) | |
7078 | __skb_queue_tail(head, skb); | |
7079 | spin_unlock_irq(&head->lock); | |
7080 | } | |
7081 | #else | |
7082 | fput(file); | |
7083 | #endif | |
7084 | } | |
7085 | ||
7086 | struct io_file_put { | |
05589553 | 7087 | struct list_head list; |
05f3fb3c | 7088 | struct file *file; |
05f3fb3c JA |
7089 | }; |
7090 | ||
4a38aed2 | 7091 | static void __io_file_put_work(struct fixed_file_ref_node *ref_node) |
65e19f54 | 7092 | { |
4a38aed2 JA |
7093 | struct fixed_file_data *file_data = ref_node->file_data; |
7094 | struct io_ring_ctx *ctx = file_data->ctx; | |
05f3fb3c | 7095 | struct io_file_put *pfile, *tmp; |
05589553 XW |
7096 | |
7097 | list_for_each_entry_safe(pfile, tmp, &ref_node->file_list, list) { | |
6a4d07cd | 7098 | list_del(&pfile->list); |
05589553 XW |
7099 | io_ring_file_put(ctx, pfile->file); |
7100 | kfree(pfile); | |
65e19f54 | 7101 | } |
05589553 | 7102 | |
6a4d07cd JA |
7103 | spin_lock(&file_data->lock); |
7104 | list_del(&ref_node->node); | |
7105 | spin_unlock(&file_data->lock); | |
05589553 XW |
7106 | |
7107 | percpu_ref_exit(&ref_node->refs); | |
7108 | kfree(ref_node); | |
7109 | percpu_ref_put(&file_data->refs); | |
2faf852d | 7110 | } |
65e19f54 | 7111 | |
4a38aed2 JA |
7112 | static void io_file_put_work(struct work_struct *work) |
7113 | { | |
7114 | struct io_ring_ctx *ctx; | |
7115 | struct llist_node *node; | |
7116 | ||
7117 | ctx = container_of(work, struct io_ring_ctx, file_put_work.work); | |
7118 | node = llist_del_all(&ctx->file_put_llist); | |
7119 | ||
7120 | while (node) { | |
7121 | struct fixed_file_ref_node *ref_node; | |
7122 | struct llist_node *next = node->next; | |
7123 | ||
7124 | ref_node = llist_entry(node, struct fixed_file_ref_node, llist); | |
7125 | __io_file_put_work(ref_node); | |
7126 | node = next; | |
7127 | } | |
7128 | } | |
7129 | ||
05589553 | 7130 | static void io_file_data_ref_zero(struct percpu_ref *ref) |
2faf852d | 7131 | { |
05589553 | 7132 | struct fixed_file_ref_node *ref_node; |
4a38aed2 JA |
7133 | struct io_ring_ctx *ctx; |
7134 | bool first_add; | |
7135 | int delay = HZ; | |
65e19f54 | 7136 | |
05589553 | 7137 | ref_node = container_of(ref, struct fixed_file_ref_node, refs); |
4a38aed2 | 7138 | ctx = ref_node->file_data->ctx; |
05589553 | 7139 | |
4a38aed2 JA |
7140 | if (percpu_ref_is_dying(&ctx->file_data->refs)) |
7141 | delay = 0; | |
05589553 | 7142 | |
4a38aed2 JA |
7143 | first_add = llist_add(&ref_node->llist, &ctx->file_put_llist); |
7144 | if (!delay) | |
7145 | mod_delayed_work(system_wq, &ctx->file_put_work, 0); | |
7146 | else if (first_add) | |
7147 | queue_delayed_work(system_wq, &ctx->file_put_work, delay); | |
05f3fb3c | 7148 | } |
65e19f54 | 7149 | |
05589553 XW |
7150 | static struct fixed_file_ref_node *alloc_fixed_file_ref_node( |
7151 | struct io_ring_ctx *ctx) | |
05f3fb3c | 7152 | { |
05589553 | 7153 | struct fixed_file_ref_node *ref_node; |
05f3fb3c | 7154 | |
05589553 XW |
7155 | ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL); |
7156 | if (!ref_node) | |
7157 | return ERR_PTR(-ENOMEM); | |
05f3fb3c | 7158 | |
05589553 XW |
7159 | if (percpu_ref_init(&ref_node->refs, io_file_data_ref_zero, |
7160 | 0, GFP_KERNEL)) { | |
7161 | kfree(ref_node); | |
7162 | return ERR_PTR(-ENOMEM); | |
7163 | } | |
7164 | INIT_LIST_HEAD(&ref_node->node); | |
7165 | INIT_LIST_HEAD(&ref_node->file_list); | |
05589553 XW |
7166 | ref_node->file_data = ctx->file_data; |
7167 | return ref_node; | |
05589553 XW |
7168 | } |
7169 | ||
7170 | static void destroy_fixed_file_ref_node(struct fixed_file_ref_node *ref_node) | |
7171 | { | |
7172 | percpu_ref_exit(&ref_node->refs); | |
7173 | kfree(ref_node); | |
65e19f54 JA |
7174 | } |
7175 | ||
6b06314c JA |
7176 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, |
7177 | unsigned nr_args) | |
7178 | { | |
7179 | __s32 __user *fds = (__s32 __user *) arg; | |
65e19f54 | 7180 | unsigned nr_tables; |
05f3fb3c | 7181 | struct file *file; |
6b06314c JA |
7182 | int fd, ret = 0; |
7183 | unsigned i; | |
05589553 | 7184 | struct fixed_file_ref_node *ref_node; |
6b06314c | 7185 | |
05f3fb3c | 7186 | if (ctx->file_data) |
6b06314c JA |
7187 | return -EBUSY; |
7188 | if (!nr_args) | |
7189 | return -EINVAL; | |
7190 | if (nr_args > IORING_MAX_FIXED_FILES) | |
7191 | return -EMFILE; | |
7192 | ||
05f3fb3c JA |
7193 | ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL); |
7194 | if (!ctx->file_data) | |
7195 | return -ENOMEM; | |
7196 | ctx->file_data->ctx = ctx; | |
7197 | init_completion(&ctx->file_data->done); | |
05589553 | 7198 | INIT_LIST_HEAD(&ctx->file_data->ref_list); |
f7fe9346 | 7199 | spin_lock_init(&ctx->file_data->lock); |
05f3fb3c | 7200 | |
65e19f54 | 7201 | nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE); |
05f3fb3c JA |
7202 | ctx->file_data->table = kcalloc(nr_tables, |
7203 | sizeof(struct fixed_file_table), | |
65e19f54 | 7204 | GFP_KERNEL); |
05f3fb3c JA |
7205 | if (!ctx->file_data->table) { |
7206 | kfree(ctx->file_data); | |
7207 | ctx->file_data = NULL; | |
6b06314c | 7208 | return -ENOMEM; |
05f3fb3c JA |
7209 | } |
7210 | ||
05589553 | 7211 | if (percpu_ref_init(&ctx->file_data->refs, io_file_ref_kill, |
05f3fb3c JA |
7212 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) { |
7213 | kfree(ctx->file_data->table); | |
7214 | kfree(ctx->file_data); | |
7215 | ctx->file_data = NULL; | |
6b06314c | 7216 | return -ENOMEM; |
05f3fb3c | 7217 | } |
6b06314c | 7218 | |
65e19f54 | 7219 | if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) { |
05f3fb3c JA |
7220 | percpu_ref_exit(&ctx->file_data->refs); |
7221 | kfree(ctx->file_data->table); | |
7222 | kfree(ctx->file_data); | |
7223 | ctx->file_data = NULL; | |
65e19f54 JA |
7224 | return -ENOMEM; |
7225 | } | |
7226 | ||
08a45173 | 7227 | for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { |
65e19f54 JA |
7228 | struct fixed_file_table *table; |
7229 | unsigned index; | |
7230 | ||
6b06314c JA |
7231 | ret = -EFAULT; |
7232 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) | |
7233 | break; | |
08a45173 JA |
7234 | /* allow sparse sets */ |
7235 | if (fd == -1) { | |
7236 | ret = 0; | |
7237 | continue; | |
7238 | } | |
6b06314c | 7239 | |
05f3fb3c | 7240 | table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT]; |
65e19f54 | 7241 | index = i & IORING_FILE_TABLE_MASK; |
05f3fb3c | 7242 | file = fget(fd); |
6b06314c JA |
7243 | |
7244 | ret = -EBADF; | |
05f3fb3c | 7245 | if (!file) |
6b06314c | 7246 | break; |
05f3fb3c | 7247 | |
6b06314c JA |
7248 | /* |
7249 | * Don't allow io_uring instances to be registered. If UNIX | |
7250 | * isn't enabled, then this causes a reference cycle and this | |
7251 | * instance can never get freed. If UNIX is enabled we'll | |
7252 | * handle it just fine, but there's still no point in allowing | |
7253 | * a ring fd as it doesn't support regular read/write anyway. | |
7254 | */ | |
05f3fb3c JA |
7255 | if (file->f_op == &io_uring_fops) { |
7256 | fput(file); | |
6b06314c JA |
7257 | break; |
7258 | } | |
6b06314c | 7259 | ret = 0; |
05f3fb3c | 7260 | table->files[index] = file; |
6b06314c JA |
7261 | } |
7262 | ||
7263 | if (ret) { | |
65e19f54 | 7264 | for (i = 0; i < ctx->nr_user_files; i++) { |
65e19f54 JA |
7265 | file = io_file_from_index(ctx, i); |
7266 | if (file) | |
7267 | fput(file); | |
7268 | } | |
7269 | for (i = 0; i < nr_tables; i++) | |
05f3fb3c | 7270 | kfree(ctx->file_data->table[i].files); |
6b06314c | 7271 | |
667e57da | 7272 | percpu_ref_exit(&ctx->file_data->refs); |
05f3fb3c JA |
7273 | kfree(ctx->file_data->table); |
7274 | kfree(ctx->file_data); | |
7275 | ctx->file_data = NULL; | |
6b06314c JA |
7276 | ctx->nr_user_files = 0; |
7277 | return ret; | |
7278 | } | |
7279 | ||
7280 | ret = io_sqe_files_scm(ctx); | |
05589553 | 7281 | if (ret) { |
6b06314c | 7282 | io_sqe_files_unregister(ctx); |
05589553 XW |
7283 | return ret; |
7284 | } | |
6b06314c | 7285 | |
05589553 XW |
7286 | ref_node = alloc_fixed_file_ref_node(ctx); |
7287 | if (IS_ERR(ref_node)) { | |
7288 | io_sqe_files_unregister(ctx); | |
7289 | return PTR_ERR(ref_node); | |
7290 | } | |
7291 | ||
7292 | ctx->file_data->cur_refs = &ref_node->refs; | |
6a4d07cd | 7293 | spin_lock(&ctx->file_data->lock); |
05589553 | 7294 | list_add(&ref_node->node, &ctx->file_data->ref_list); |
6a4d07cd | 7295 | spin_unlock(&ctx->file_data->lock); |
05589553 | 7296 | percpu_ref_get(&ctx->file_data->refs); |
6b06314c JA |
7297 | return ret; |
7298 | } | |
7299 | ||
c3a31e60 JA |
7300 | static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file, |
7301 | int index) | |
7302 | { | |
7303 | #if defined(CONFIG_UNIX) | |
7304 | struct sock *sock = ctx->ring_sock->sk; | |
7305 | struct sk_buff_head *head = &sock->sk_receive_queue; | |
7306 | struct sk_buff *skb; | |
7307 | ||
7308 | /* | |
7309 | * See if we can merge this file into an existing skb SCM_RIGHTS | |
7310 | * file set. If there's no room, fall back to allocating a new skb | |
7311 | * and filling it in. | |
7312 | */ | |
7313 | spin_lock_irq(&head->lock); | |
7314 | skb = skb_peek(head); | |
7315 | if (skb) { | |
7316 | struct scm_fp_list *fpl = UNIXCB(skb).fp; | |
7317 | ||
7318 | if (fpl->count < SCM_MAX_FD) { | |
7319 | __skb_unlink(skb, head); | |
7320 | spin_unlock_irq(&head->lock); | |
7321 | fpl->fp[fpl->count] = get_file(file); | |
7322 | unix_inflight(fpl->user, fpl->fp[fpl->count]); | |
7323 | fpl->count++; | |
7324 | spin_lock_irq(&head->lock); | |
7325 | __skb_queue_head(head, skb); | |
7326 | } else { | |
7327 | skb = NULL; | |
7328 | } | |
7329 | } | |
7330 | spin_unlock_irq(&head->lock); | |
7331 | ||
7332 | if (skb) { | |
7333 | fput(file); | |
7334 | return 0; | |
7335 | } | |
7336 | ||
7337 | return __io_sqe_files_scm(ctx, 1, index); | |
7338 | #else | |
7339 | return 0; | |
7340 | #endif | |
7341 | } | |
7342 | ||
a5318d3c | 7343 | static int io_queue_file_removal(struct fixed_file_data *data, |
05589553 | 7344 | struct file *file) |
05f3fb3c | 7345 | { |
a5318d3c | 7346 | struct io_file_put *pfile; |
05589553 XW |
7347 | struct percpu_ref *refs = data->cur_refs; |
7348 | struct fixed_file_ref_node *ref_node; | |
05f3fb3c | 7349 | |
05f3fb3c | 7350 | pfile = kzalloc(sizeof(*pfile), GFP_KERNEL); |
a5318d3c HD |
7351 | if (!pfile) |
7352 | return -ENOMEM; | |
05f3fb3c | 7353 | |
05589553 | 7354 | ref_node = container_of(refs, struct fixed_file_ref_node, refs); |
05f3fb3c | 7355 | pfile->file = file; |
05589553 XW |
7356 | list_add(&pfile->list, &ref_node->file_list); |
7357 | ||
a5318d3c | 7358 | return 0; |
05f3fb3c JA |
7359 | } |
7360 | ||
7361 | static int __io_sqe_files_update(struct io_ring_ctx *ctx, | |
7362 | struct io_uring_files_update *up, | |
7363 | unsigned nr_args) | |
7364 | { | |
7365 | struct fixed_file_data *data = ctx->file_data; | |
05589553 | 7366 | struct fixed_file_ref_node *ref_node; |
05f3fb3c | 7367 | struct file *file; |
c3a31e60 JA |
7368 | __s32 __user *fds; |
7369 | int fd, i, err; | |
7370 | __u32 done; | |
05589553 | 7371 | bool needs_switch = false; |
c3a31e60 | 7372 | |
05f3fb3c | 7373 | if (check_add_overflow(up->offset, nr_args, &done)) |
c3a31e60 JA |
7374 | return -EOVERFLOW; |
7375 | if (done > ctx->nr_user_files) | |
7376 | return -EINVAL; | |
7377 | ||
05589553 XW |
7378 | ref_node = alloc_fixed_file_ref_node(ctx); |
7379 | if (IS_ERR(ref_node)) | |
7380 | return PTR_ERR(ref_node); | |
7381 | ||
c3a31e60 | 7382 | done = 0; |
05f3fb3c | 7383 | fds = u64_to_user_ptr(up->fds); |
c3a31e60 | 7384 | while (nr_args) { |
65e19f54 JA |
7385 | struct fixed_file_table *table; |
7386 | unsigned index; | |
7387 | ||
c3a31e60 JA |
7388 | err = 0; |
7389 | if (copy_from_user(&fd, &fds[done], sizeof(fd))) { | |
7390 | err = -EFAULT; | |
7391 | break; | |
7392 | } | |
05f3fb3c JA |
7393 | i = array_index_nospec(up->offset, ctx->nr_user_files); |
7394 | table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT]; | |
65e19f54 JA |
7395 | index = i & IORING_FILE_TABLE_MASK; |
7396 | if (table->files[index]) { | |
98dfd502 | 7397 | file = table->files[index]; |
a5318d3c HD |
7398 | err = io_queue_file_removal(data, file); |
7399 | if (err) | |
7400 | break; | |
65e19f54 | 7401 | table->files[index] = NULL; |
05589553 | 7402 | needs_switch = true; |
c3a31e60 JA |
7403 | } |
7404 | if (fd != -1) { | |
c3a31e60 JA |
7405 | file = fget(fd); |
7406 | if (!file) { | |
7407 | err = -EBADF; | |
7408 | break; | |
7409 | } | |
7410 | /* | |
7411 | * Don't allow io_uring instances to be registered. If | |
7412 | * UNIX isn't enabled, then this causes a reference | |
7413 | * cycle and this instance can never get freed. If UNIX | |
7414 | * is enabled we'll handle it just fine, but there's | |
7415 | * still no point in allowing a ring fd as it doesn't | |
7416 | * support regular read/write anyway. | |
7417 | */ | |
7418 | if (file->f_op == &io_uring_fops) { | |
7419 | fput(file); | |
7420 | err = -EBADF; | |
7421 | break; | |
7422 | } | |
65e19f54 | 7423 | table->files[index] = file; |
c3a31e60 | 7424 | err = io_sqe_file_register(ctx, file, i); |
f3bd9dae | 7425 | if (err) { |
95d1c8e5 | 7426 | table->files[index] = NULL; |
f3bd9dae | 7427 | fput(file); |
c3a31e60 | 7428 | break; |
f3bd9dae | 7429 | } |
c3a31e60 JA |
7430 | } |
7431 | nr_args--; | |
7432 | done++; | |
05f3fb3c JA |
7433 | up->offset++; |
7434 | } | |
7435 | ||
05589553 XW |
7436 | if (needs_switch) { |
7437 | percpu_ref_kill(data->cur_refs); | |
6a4d07cd | 7438 | spin_lock(&data->lock); |
05589553 XW |
7439 | list_add(&ref_node->node, &data->ref_list); |
7440 | data->cur_refs = &ref_node->refs; | |
6a4d07cd | 7441 | spin_unlock(&data->lock); |
05589553 XW |
7442 | percpu_ref_get(&ctx->file_data->refs); |
7443 | } else | |
7444 | destroy_fixed_file_ref_node(ref_node); | |
c3a31e60 JA |
7445 | |
7446 | return done ? done : err; | |
7447 | } | |
05589553 | 7448 | |
05f3fb3c JA |
7449 | static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg, |
7450 | unsigned nr_args) | |
7451 | { | |
7452 | struct io_uring_files_update up; | |
7453 | ||
7454 | if (!ctx->file_data) | |
7455 | return -ENXIO; | |
7456 | if (!nr_args) | |
7457 | return -EINVAL; | |
7458 | if (copy_from_user(&up, arg, sizeof(up))) | |
7459 | return -EFAULT; | |
7460 | if (up.resv) | |
7461 | return -EINVAL; | |
7462 | ||
7463 | return __io_sqe_files_update(ctx, &up, nr_args); | |
7464 | } | |
c3a31e60 | 7465 | |
e9fd9396 | 7466 | static void io_free_work(struct io_wq_work *work) |
7d723065 JA |
7467 | { |
7468 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
7469 | ||
e9fd9396 | 7470 | /* Consider that io_steal_work() relies on this ref */ |
7d723065 JA |
7471 | io_put_req(req); |
7472 | } | |
7473 | ||
24369c2e PB |
7474 | static int io_init_wq_offload(struct io_ring_ctx *ctx, |
7475 | struct io_uring_params *p) | |
7476 | { | |
7477 | struct io_wq_data data; | |
7478 | struct fd f; | |
7479 | struct io_ring_ctx *ctx_attach; | |
7480 | unsigned int concurrency; | |
7481 | int ret = 0; | |
7482 | ||
7483 | data.user = ctx->user; | |
e9fd9396 | 7484 | data.free_work = io_free_work; |
f5fa38c5 | 7485 | data.do_work = io_wq_submit_work; |
24369c2e PB |
7486 | |
7487 | if (!(p->flags & IORING_SETUP_ATTACH_WQ)) { | |
7488 | /* Do QD, or 4 * CPUS, whatever is smallest */ | |
7489 | concurrency = min(ctx->sq_entries, 4 * num_online_cpus()); | |
7490 | ||
7491 | ctx->io_wq = io_wq_create(concurrency, &data); | |
7492 | if (IS_ERR(ctx->io_wq)) { | |
7493 | ret = PTR_ERR(ctx->io_wq); | |
7494 | ctx->io_wq = NULL; | |
7495 | } | |
7496 | return ret; | |
7497 | } | |
7498 | ||
7499 | f = fdget(p->wq_fd); | |
7500 | if (!f.file) | |
7501 | return -EBADF; | |
7502 | ||
7503 | if (f.file->f_op != &io_uring_fops) { | |
7504 | ret = -EINVAL; | |
7505 | goto out_fput; | |
7506 | } | |
7507 | ||
7508 | ctx_attach = f.file->private_data; | |
7509 | /* @io_wq is protected by holding the fd */ | |
7510 | if (!io_wq_get(ctx_attach->io_wq, &data)) { | |
7511 | ret = -EINVAL; | |
7512 | goto out_fput; | |
7513 | } | |
7514 | ||
7515 | ctx->io_wq = ctx_attach->io_wq; | |
7516 | out_fput: | |
7517 | fdput(f); | |
7518 | return ret; | |
7519 | } | |
7520 | ||
6c271ce2 JA |
7521 | static int io_sq_offload_start(struct io_ring_ctx *ctx, |
7522 | struct io_uring_params *p) | |
2b188cc1 JA |
7523 | { |
7524 | int ret; | |
7525 | ||
6c271ce2 | 7526 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3ec482d1 JA |
7527 | ret = -EPERM; |
7528 | if (!capable(CAP_SYS_ADMIN)) | |
7529 | goto err; | |
7530 | ||
917257da JA |
7531 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
7532 | if (!ctx->sq_thread_idle) | |
7533 | ctx->sq_thread_idle = HZ; | |
7534 | ||
6c271ce2 | 7535 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 7536 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 7537 | |
917257da | 7538 | ret = -EINVAL; |
44a9bd18 JA |
7539 | if (cpu >= nr_cpu_ids) |
7540 | goto err; | |
7889f44d | 7541 | if (!cpu_online(cpu)) |
917257da JA |
7542 | goto err; |
7543 | ||
6c271ce2 JA |
7544 | ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread, |
7545 | ctx, cpu, | |
7546 | "io_uring-sq"); | |
7547 | } else { | |
7548 | ctx->sqo_thread = kthread_create(io_sq_thread, ctx, | |
7549 | "io_uring-sq"); | |
7550 | } | |
7551 | if (IS_ERR(ctx->sqo_thread)) { | |
7552 | ret = PTR_ERR(ctx->sqo_thread); | |
7553 | ctx->sqo_thread = NULL; | |
7554 | goto err; | |
7555 | } | |
7556 | wake_up_process(ctx->sqo_thread); | |
7557 | } else if (p->flags & IORING_SETUP_SQ_AFF) { | |
7558 | /* Can't have SQ_AFF without SQPOLL */ | |
7559 | ret = -EINVAL; | |
7560 | goto err; | |
7561 | } | |
7562 | ||
24369c2e PB |
7563 | ret = io_init_wq_offload(ctx, p); |
7564 | if (ret) | |
2b188cc1 | 7565 | goto err; |
2b188cc1 JA |
7566 | |
7567 | return 0; | |
7568 | err: | |
54a91f3b | 7569 | io_finish_async(ctx); |
2b188cc1 JA |
7570 | return ret; |
7571 | } | |
7572 | ||
a087e2b5 BM |
7573 | static inline void __io_unaccount_mem(struct user_struct *user, |
7574 | unsigned long nr_pages) | |
2b188cc1 JA |
7575 | { |
7576 | atomic_long_sub(nr_pages, &user->locked_vm); | |
7577 | } | |
7578 | ||
a087e2b5 BM |
7579 | static inline int __io_account_mem(struct user_struct *user, |
7580 | unsigned long nr_pages) | |
2b188cc1 JA |
7581 | { |
7582 | unsigned long page_limit, cur_pages, new_pages; | |
7583 | ||
7584 | /* Don't allow more pages than we can safely lock */ | |
7585 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
7586 | ||
7587 | do { | |
7588 | cur_pages = atomic_long_read(&user->locked_vm); | |
7589 | new_pages = cur_pages + nr_pages; | |
7590 | if (new_pages > page_limit) | |
7591 | return -ENOMEM; | |
7592 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
7593 | new_pages) != cur_pages); | |
7594 | ||
7595 | return 0; | |
7596 | } | |
7597 | ||
2e0464d4 BM |
7598 | static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages, |
7599 | enum io_mem_account acct) | |
a087e2b5 | 7600 | { |
aad5d8da | 7601 | if (ctx->limit_mem) |
a087e2b5 | 7602 | __io_unaccount_mem(ctx->user, nr_pages); |
30975825 | 7603 | |
2aede0e4 | 7604 | if (ctx->mm_account) { |
2e0464d4 | 7605 | if (acct == ACCT_LOCKED) |
2aede0e4 | 7606 | ctx->mm_account->locked_vm -= nr_pages; |
2e0464d4 | 7607 | else if (acct == ACCT_PINNED) |
2aede0e4 | 7608 | atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm); |
2e0464d4 | 7609 | } |
a087e2b5 BM |
7610 | } |
7611 | ||
2e0464d4 BM |
7612 | static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages, |
7613 | enum io_mem_account acct) | |
a087e2b5 | 7614 | { |
30975825 BM |
7615 | int ret; |
7616 | ||
7617 | if (ctx->limit_mem) { | |
7618 | ret = __io_account_mem(ctx->user, nr_pages); | |
7619 | if (ret) | |
7620 | return ret; | |
7621 | } | |
7622 | ||
2aede0e4 | 7623 | if (ctx->mm_account) { |
2e0464d4 | 7624 | if (acct == ACCT_LOCKED) |
2aede0e4 | 7625 | ctx->mm_account->locked_vm += nr_pages; |
2e0464d4 | 7626 | else if (acct == ACCT_PINNED) |
2aede0e4 | 7627 | atomic64_add(nr_pages, &ctx->mm_account->pinned_vm); |
2e0464d4 | 7628 | } |
a087e2b5 BM |
7629 | |
7630 | return 0; | |
7631 | } | |
7632 | ||
2b188cc1 JA |
7633 | static void io_mem_free(void *ptr) |
7634 | { | |
52e04ef4 MR |
7635 | struct page *page; |
7636 | ||
7637 | if (!ptr) | |
7638 | return; | |
2b188cc1 | 7639 | |
52e04ef4 | 7640 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
7641 | if (put_page_testzero(page)) |
7642 | free_compound_page(page); | |
7643 | } | |
7644 | ||
7645 | static void *io_mem_alloc(size_t size) | |
7646 | { | |
7647 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
7648 | __GFP_NORETRY; | |
7649 | ||
7650 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
7651 | } | |
7652 | ||
75b28aff HV |
7653 | static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries, |
7654 | size_t *sq_offset) | |
7655 | { | |
7656 | struct io_rings *rings; | |
7657 | size_t off, sq_array_size; | |
7658 | ||
7659 | off = struct_size(rings, cqes, cq_entries); | |
7660 | if (off == SIZE_MAX) | |
7661 | return SIZE_MAX; | |
7662 | ||
7663 | #ifdef CONFIG_SMP | |
7664 | off = ALIGN(off, SMP_CACHE_BYTES); | |
7665 | if (off == 0) | |
7666 | return SIZE_MAX; | |
7667 | #endif | |
7668 | ||
b36200f5 DV |
7669 | if (sq_offset) |
7670 | *sq_offset = off; | |
7671 | ||
75b28aff HV |
7672 | sq_array_size = array_size(sizeof(u32), sq_entries); |
7673 | if (sq_array_size == SIZE_MAX) | |
7674 | return SIZE_MAX; | |
7675 | ||
7676 | if (check_add_overflow(off, sq_array_size, &off)) | |
7677 | return SIZE_MAX; | |
7678 | ||
75b28aff HV |
7679 | return off; |
7680 | } | |
7681 | ||
2b188cc1 JA |
7682 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) |
7683 | { | |
75b28aff | 7684 | size_t pages; |
2b188cc1 | 7685 | |
75b28aff HV |
7686 | pages = (size_t)1 << get_order( |
7687 | rings_size(sq_entries, cq_entries, NULL)); | |
7688 | pages += (size_t)1 << get_order( | |
7689 | array_size(sizeof(struct io_uring_sqe), sq_entries)); | |
2b188cc1 | 7690 | |
75b28aff | 7691 | return pages; |
2b188cc1 JA |
7692 | } |
7693 | ||
edafccee JA |
7694 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
7695 | { | |
7696 | int i, j; | |
7697 | ||
7698 | if (!ctx->user_bufs) | |
7699 | return -ENXIO; | |
7700 | ||
7701 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
7702 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
7703 | ||
7704 | for (j = 0; j < imu->nr_bvecs; j++) | |
f1f6a7dd | 7705 | unpin_user_page(imu->bvec[j].bv_page); |
edafccee | 7706 | |
2e0464d4 | 7707 | io_unaccount_mem(ctx, imu->nr_bvecs, ACCT_PINNED); |
d4ef6475 | 7708 | kvfree(imu->bvec); |
edafccee JA |
7709 | imu->nr_bvecs = 0; |
7710 | } | |
7711 | ||
7712 | kfree(ctx->user_bufs); | |
7713 | ctx->user_bufs = NULL; | |
7714 | ctx->nr_user_bufs = 0; | |
7715 | return 0; | |
7716 | } | |
7717 | ||
7718 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
7719 | void __user *arg, unsigned index) | |
7720 | { | |
7721 | struct iovec __user *src; | |
7722 | ||
7723 | #ifdef CONFIG_COMPAT | |
7724 | if (ctx->compat) { | |
7725 | struct compat_iovec __user *ciovs; | |
7726 | struct compat_iovec ciov; | |
7727 | ||
7728 | ciovs = (struct compat_iovec __user *) arg; | |
7729 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
7730 | return -EFAULT; | |
7731 | ||
d55e5f5b | 7732 | dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base); |
edafccee JA |
7733 | dst->iov_len = ciov.iov_len; |
7734 | return 0; | |
7735 | } | |
7736 | #endif | |
7737 | src = (struct iovec __user *) arg; | |
7738 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
7739 | return -EFAULT; | |
7740 | return 0; | |
7741 | } | |
7742 | ||
7743 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, | |
7744 | unsigned nr_args) | |
7745 | { | |
7746 | struct vm_area_struct **vmas = NULL; | |
7747 | struct page **pages = NULL; | |
7748 | int i, j, got_pages = 0; | |
7749 | int ret = -EINVAL; | |
7750 | ||
7751 | if (ctx->user_bufs) | |
7752 | return -EBUSY; | |
7753 | if (!nr_args || nr_args > UIO_MAXIOV) | |
7754 | return -EINVAL; | |
7755 | ||
7756 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
7757 | GFP_KERNEL); | |
7758 | if (!ctx->user_bufs) | |
7759 | return -ENOMEM; | |
7760 | ||
7761 | for (i = 0; i < nr_args; i++) { | |
7762 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
7763 | unsigned long off, start, end, ubuf; | |
7764 | int pret, nr_pages; | |
7765 | struct iovec iov; | |
7766 | size_t size; | |
7767 | ||
7768 | ret = io_copy_iov(ctx, &iov, arg, i); | |
7769 | if (ret) | |
a278682d | 7770 | goto err; |
edafccee JA |
7771 | |
7772 | /* | |
7773 | * Don't impose further limits on the size and buffer | |
7774 | * constraints here, we'll -EINVAL later when IO is | |
7775 | * submitted if they are wrong. | |
7776 | */ | |
7777 | ret = -EFAULT; | |
7778 | if (!iov.iov_base || !iov.iov_len) | |
7779 | goto err; | |
7780 | ||
7781 | /* arbitrary limit, but we need something */ | |
7782 | if (iov.iov_len > SZ_1G) | |
7783 | goto err; | |
7784 | ||
7785 | ubuf = (unsigned long) iov.iov_base; | |
7786 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
7787 | start = ubuf >> PAGE_SHIFT; | |
7788 | nr_pages = end - start; | |
7789 | ||
2e0464d4 | 7790 | ret = io_account_mem(ctx, nr_pages, ACCT_PINNED); |
a087e2b5 BM |
7791 | if (ret) |
7792 | goto err; | |
edafccee JA |
7793 | |
7794 | ret = 0; | |
7795 | if (!pages || nr_pages > got_pages) { | |
a8c73c1a DE |
7796 | kvfree(vmas); |
7797 | kvfree(pages); | |
d4ef6475 | 7798 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), |
edafccee | 7799 | GFP_KERNEL); |
d4ef6475 | 7800 | vmas = kvmalloc_array(nr_pages, |
edafccee JA |
7801 | sizeof(struct vm_area_struct *), |
7802 | GFP_KERNEL); | |
7803 | if (!pages || !vmas) { | |
7804 | ret = -ENOMEM; | |
2e0464d4 | 7805 | io_unaccount_mem(ctx, nr_pages, ACCT_PINNED); |
edafccee JA |
7806 | goto err; |
7807 | } | |
7808 | got_pages = nr_pages; | |
7809 | } | |
7810 | ||
d4ef6475 | 7811 | imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec), |
edafccee JA |
7812 | GFP_KERNEL); |
7813 | ret = -ENOMEM; | |
7814 | if (!imu->bvec) { | |
2e0464d4 | 7815 | io_unaccount_mem(ctx, nr_pages, ACCT_PINNED); |
edafccee JA |
7816 | goto err; |
7817 | } | |
7818 | ||
7819 | ret = 0; | |
d8ed45c5 | 7820 | mmap_read_lock(current->mm); |
2113b05d | 7821 | pret = pin_user_pages(ubuf, nr_pages, |
932f4a63 IW |
7822 | FOLL_WRITE | FOLL_LONGTERM, |
7823 | pages, vmas); | |
edafccee JA |
7824 | if (pret == nr_pages) { |
7825 | /* don't support file backed memory */ | |
7826 | for (j = 0; j < nr_pages; j++) { | |
7827 | struct vm_area_struct *vma = vmas[j]; | |
7828 | ||
7829 | if (vma->vm_file && | |
7830 | !is_file_hugepages(vma->vm_file)) { | |
7831 | ret = -EOPNOTSUPP; | |
7832 | break; | |
7833 | } | |
7834 | } | |
7835 | } else { | |
7836 | ret = pret < 0 ? pret : -EFAULT; | |
7837 | } | |
d8ed45c5 | 7838 | mmap_read_unlock(current->mm); |
edafccee JA |
7839 | if (ret) { |
7840 | /* | |
7841 | * if we did partial map, or found file backed vmas, | |
7842 | * release any pages we did get | |
7843 | */ | |
27c4d3a3 | 7844 | if (pret > 0) |
f1f6a7dd | 7845 | unpin_user_pages(pages, pret); |
2e0464d4 | 7846 | io_unaccount_mem(ctx, nr_pages, ACCT_PINNED); |
d4ef6475 | 7847 | kvfree(imu->bvec); |
edafccee JA |
7848 | goto err; |
7849 | } | |
7850 | ||
7851 | off = ubuf & ~PAGE_MASK; | |
7852 | size = iov.iov_len; | |
7853 | for (j = 0; j < nr_pages; j++) { | |
7854 | size_t vec_len; | |
7855 | ||
7856 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
7857 | imu->bvec[j].bv_page = pages[j]; | |
7858 | imu->bvec[j].bv_len = vec_len; | |
7859 | imu->bvec[j].bv_offset = off; | |
7860 | off = 0; | |
7861 | size -= vec_len; | |
7862 | } | |
7863 | /* store original address for later verification */ | |
7864 | imu->ubuf = ubuf; | |
7865 | imu->len = iov.iov_len; | |
7866 | imu->nr_bvecs = nr_pages; | |
7867 | ||
7868 | ctx->nr_user_bufs++; | |
7869 | } | |
d4ef6475 MR |
7870 | kvfree(pages); |
7871 | kvfree(vmas); | |
edafccee JA |
7872 | return 0; |
7873 | err: | |
d4ef6475 MR |
7874 | kvfree(pages); |
7875 | kvfree(vmas); | |
edafccee JA |
7876 | io_sqe_buffer_unregister(ctx); |
7877 | return ret; | |
7878 | } | |
7879 | ||
9b402849 JA |
7880 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
7881 | { | |
7882 | __s32 __user *fds = arg; | |
7883 | int fd; | |
7884 | ||
7885 | if (ctx->cq_ev_fd) | |
7886 | return -EBUSY; | |
7887 | ||
7888 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
7889 | return -EFAULT; | |
7890 | ||
7891 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
7892 | if (IS_ERR(ctx->cq_ev_fd)) { | |
7893 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
7894 | ctx->cq_ev_fd = NULL; | |
7895 | return ret; | |
7896 | } | |
7897 | ||
7898 | return 0; | |
7899 | } | |
7900 | ||
7901 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
7902 | { | |
7903 | if (ctx->cq_ev_fd) { | |
7904 | eventfd_ctx_put(ctx->cq_ev_fd); | |
7905 | ctx->cq_ev_fd = NULL; | |
7906 | return 0; | |
7907 | } | |
7908 | ||
7909 | return -ENXIO; | |
7910 | } | |
7911 | ||
5a2e745d JA |
7912 | static int __io_destroy_buffers(int id, void *p, void *data) |
7913 | { | |
7914 | struct io_ring_ctx *ctx = data; | |
7915 | struct io_buffer *buf = p; | |
7916 | ||
067524e9 | 7917 | __io_remove_buffers(ctx, buf, id, -1U); |
5a2e745d JA |
7918 | return 0; |
7919 | } | |
7920 | ||
7921 | static void io_destroy_buffers(struct io_ring_ctx *ctx) | |
7922 | { | |
7923 | idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx); | |
7924 | idr_destroy(&ctx->io_buffer_idr); | |
7925 | } | |
7926 | ||
2b188cc1 JA |
7927 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
7928 | { | |
6b06314c | 7929 | io_finish_async(ctx); |
5dbcad51 | 7930 | io_sqe_buffer_unregister(ctx); |
2aede0e4 JA |
7931 | |
7932 | if (ctx->sqo_task) { | |
7933 | put_task_struct(ctx->sqo_task); | |
7934 | ctx->sqo_task = NULL; | |
7935 | mmdrop(ctx->mm_account); | |
7936 | ctx->mm_account = NULL; | |
30975825 | 7937 | } |
def596e9 | 7938 | |
6b06314c | 7939 | io_sqe_files_unregister(ctx); |
9b402849 | 7940 | io_eventfd_unregister(ctx); |
5a2e745d | 7941 | io_destroy_buffers(ctx); |
41726c9a | 7942 | idr_destroy(&ctx->personality_idr); |
def596e9 | 7943 | |
2b188cc1 | 7944 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
7945 | if (ctx->ring_sock) { |
7946 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 7947 | sock_release(ctx->ring_sock); |
355e8d26 | 7948 | } |
2b188cc1 JA |
7949 | #endif |
7950 | ||
75b28aff | 7951 | io_mem_free(ctx->rings); |
2b188cc1 | 7952 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
7953 | |
7954 | percpu_ref_exit(&ctx->refs); | |
2b188cc1 | 7955 | free_uid(ctx->user); |
181e448d | 7956 | put_cred(ctx->creds); |
78076bb6 | 7957 | kfree(ctx->cancel_hash); |
0ddf92e8 | 7958 | kmem_cache_free(req_cachep, ctx->fallback_req); |
2b188cc1 JA |
7959 | kfree(ctx); |
7960 | } | |
7961 | ||
7962 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
7963 | { | |
7964 | struct io_ring_ctx *ctx = file->private_data; | |
7965 | __poll_t mask = 0; | |
7966 | ||
7967 | poll_wait(file, &ctx->cq_wait, wait); | |
4f7067c3 SB |
7968 | /* |
7969 | * synchronizes with barrier from wq_has_sleeper call in | |
7970 | * io_commit_cqring | |
7971 | */ | |
2b188cc1 | 7972 | smp_rmb(); |
75b28aff HV |
7973 | if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head != |
7974 | ctx->rings->sq_ring_entries) | |
2b188cc1 | 7975 | mask |= EPOLLOUT | EPOLLWRNORM; |
63e5d81f | 7976 | if (io_cqring_events(ctx, false)) |
2b188cc1 JA |
7977 | mask |= EPOLLIN | EPOLLRDNORM; |
7978 | ||
7979 | return mask; | |
7980 | } | |
7981 | ||
7982 | static int io_uring_fasync(int fd, struct file *file, int on) | |
7983 | { | |
7984 | struct io_ring_ctx *ctx = file->private_data; | |
7985 | ||
7986 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
7987 | } | |
7988 | ||
071698e1 JA |
7989 | static int io_remove_personalities(int id, void *p, void *data) |
7990 | { | |
7991 | struct io_ring_ctx *ctx = data; | |
7992 | const struct cred *cred; | |
7993 | ||
7994 | cred = idr_remove(&ctx->personality_idr, id); | |
7995 | if (cred) | |
7996 | put_cred(cred); | |
7997 | return 0; | |
7998 | } | |
7999 | ||
85faa7b8 JA |
8000 | static void io_ring_exit_work(struct work_struct *work) |
8001 | { | |
b2edc0a7 PB |
8002 | struct io_ring_ctx *ctx = container_of(work, struct io_ring_ctx, |
8003 | exit_work); | |
85faa7b8 | 8004 | |
56952e91 JA |
8005 | /* |
8006 | * If we're doing polled IO and end up having requests being | |
8007 | * submitted async (out-of-line), then completions can come in while | |
8008 | * we're waiting for refs to drop. We need to reap these manually, | |
8009 | * as nobody else will be looking for them. | |
8010 | */ | |
b2edc0a7 | 8011 | do { |
56952e91 JA |
8012 | if (ctx->rings) |
8013 | io_cqring_overflow_flush(ctx, true); | |
b2edc0a7 PB |
8014 | io_iopoll_try_reap_events(ctx); |
8015 | } while (!wait_for_completion_timeout(&ctx->ref_comp, HZ/20)); | |
85faa7b8 JA |
8016 | io_ring_ctx_free(ctx); |
8017 | } | |
8018 | ||
2b188cc1 JA |
8019 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
8020 | { | |
8021 | mutex_lock(&ctx->uring_lock); | |
8022 | percpu_ref_kill(&ctx->refs); | |
8023 | mutex_unlock(&ctx->uring_lock); | |
8024 | ||
f3606e3a JA |
8025 | io_kill_timeouts(ctx, NULL); |
8026 | io_poll_remove_all(ctx, NULL); | |
561fb04a JA |
8027 | |
8028 | if (ctx->io_wq) | |
8029 | io_wq_cancel_all(ctx->io_wq); | |
8030 | ||
15dff286 JA |
8031 | /* if we failed setting up the ctx, we might not have any rings */ |
8032 | if (ctx->rings) | |
8033 | io_cqring_overflow_flush(ctx, true); | |
b2edc0a7 | 8034 | io_iopoll_try_reap_events(ctx); |
071698e1 | 8035 | idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx); |
309fc03a JA |
8036 | |
8037 | /* | |
8038 | * Do this upfront, so we won't have a grace period where the ring | |
8039 | * is closed but resources aren't reaped yet. This can cause | |
8040 | * spurious failure in setting up a new ring. | |
8041 | */ | |
760618f7 JA |
8042 | io_unaccount_mem(ctx, ring_pages(ctx->sq_entries, ctx->cq_entries), |
8043 | ACCT_LOCKED); | |
309fc03a | 8044 | |
85faa7b8 | 8045 | INIT_WORK(&ctx->exit_work, io_ring_exit_work); |
fc666777 JA |
8046 | /* |
8047 | * Use system_unbound_wq to avoid spawning tons of event kworkers | |
8048 | * if we're exiting a ton of rings at the same time. It just adds | |
8049 | * noise and overhead, there's no discernable change in runtime | |
8050 | * over using system_wq. | |
8051 | */ | |
8052 | queue_work(system_unbound_wq, &ctx->exit_work); | |
2b188cc1 JA |
8053 | } |
8054 | ||
8055 | static int io_uring_release(struct inode *inode, struct file *file) | |
8056 | { | |
8057 | struct io_ring_ctx *ctx = file->private_data; | |
8058 | ||
8059 | file->private_data = NULL; | |
8060 | io_ring_ctx_wait_and_kill(ctx); | |
8061 | return 0; | |
8062 | } | |
8063 | ||
67c4d9e6 PB |
8064 | static bool io_wq_files_match(struct io_wq_work *work, void *data) |
8065 | { | |
8066 | struct files_struct *files = data; | |
8067 | ||
8068 | return work->files == files; | |
8069 | } | |
8070 | ||
f254ac04 JA |
8071 | /* |
8072 | * Returns true if 'preq' is the link parent of 'req' | |
8073 | */ | |
8074 | static bool io_match_link(struct io_kiocb *preq, struct io_kiocb *req) | |
8075 | { | |
8076 | struct io_kiocb *link; | |
8077 | ||
8078 | if (!(preq->flags & REQ_F_LINK_HEAD)) | |
8079 | return false; | |
8080 | ||
8081 | list_for_each_entry(link, &preq->link_list, link_list) { | |
8082 | if (link == req) | |
8083 | return true; | |
8084 | } | |
8085 | ||
8086 | return false; | |
8087 | } | |
8088 | ||
c127a2a1 PB |
8089 | static inline bool io_match_files(struct io_kiocb *req, |
8090 | struct files_struct *files) | |
8091 | { | |
8092 | return (req->flags & REQ_F_WORK_INITIALIZED) && req->work.files == files; | |
8093 | } | |
8094 | ||
8095 | static bool io_match_link_files(struct io_kiocb *req, | |
8096 | struct files_struct *files) | |
8097 | { | |
8098 | struct io_kiocb *link; | |
8099 | ||
8100 | if (io_match_files(req, files)) | |
8101 | return true; | |
8102 | if (req->flags & REQ_F_LINK_HEAD) { | |
8103 | list_for_each_entry(link, &req->link_list, link_list) { | |
8104 | if (io_match_files(link, files)) | |
8105 | return true; | |
8106 | } | |
8107 | } | |
8108 | return false; | |
8109 | } | |
8110 | ||
f254ac04 JA |
8111 | /* |
8112 | * We're looking to cancel 'req' because it's holding on to our files, but | |
8113 | * 'req' could be a link to another request. See if it is, and cancel that | |
8114 | * parent request if so. | |
8115 | */ | |
8116 | static bool io_poll_remove_link(struct io_ring_ctx *ctx, struct io_kiocb *req) | |
8117 | { | |
8118 | struct hlist_node *tmp; | |
8119 | struct io_kiocb *preq; | |
8120 | bool found = false; | |
8121 | int i; | |
8122 | ||
8123 | spin_lock_irq(&ctx->completion_lock); | |
8124 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { | |
8125 | struct hlist_head *list; | |
8126 | ||
8127 | list = &ctx->cancel_hash[i]; | |
8128 | hlist_for_each_entry_safe(preq, tmp, list, hash_node) { | |
8129 | found = io_match_link(preq, req); | |
8130 | if (found) { | |
8131 | io_poll_remove_one(preq); | |
8132 | break; | |
8133 | } | |
8134 | } | |
8135 | } | |
8136 | spin_unlock_irq(&ctx->completion_lock); | |
8137 | return found; | |
8138 | } | |
8139 | ||
8140 | static bool io_timeout_remove_link(struct io_ring_ctx *ctx, | |
8141 | struct io_kiocb *req) | |
8142 | { | |
8143 | struct io_kiocb *preq; | |
8144 | bool found = false; | |
8145 | ||
8146 | spin_lock_irq(&ctx->completion_lock); | |
8147 | list_for_each_entry(preq, &ctx->timeout_list, timeout.list) { | |
8148 | found = io_match_link(preq, req); | |
8149 | if (found) { | |
8150 | __io_timeout_cancel(preq); | |
8151 | break; | |
8152 | } | |
8153 | } | |
8154 | spin_unlock_irq(&ctx->completion_lock); | |
8155 | return found; | |
8156 | } | |
8157 | ||
b711d4ea JA |
8158 | static bool io_cancel_link_cb(struct io_wq_work *work, void *data) |
8159 | { | |
8160 | return io_match_link(container_of(work, struct io_kiocb, work), data); | |
8161 | } | |
8162 | ||
8163 | static void io_attempt_cancel(struct io_ring_ctx *ctx, struct io_kiocb *req) | |
8164 | { | |
8165 | enum io_wq_cancel cret; | |
8166 | ||
8167 | /* cancel this particular work, if it's running */ | |
8168 | cret = io_wq_cancel_work(ctx->io_wq, &req->work); | |
8169 | if (cret != IO_WQ_CANCEL_NOTFOUND) | |
8170 | return; | |
8171 | ||
8172 | /* find links that hold this pending, cancel those */ | |
8173 | cret = io_wq_cancel_cb(ctx->io_wq, io_cancel_link_cb, req, true); | |
8174 | if (cret != IO_WQ_CANCEL_NOTFOUND) | |
8175 | return; | |
8176 | ||
8177 | /* if we have a poll link holding this pending, cancel that */ | |
8178 | if (io_poll_remove_link(ctx, req)) | |
8179 | return; | |
8180 | ||
8181 | /* final option, timeout link is holding this req pending */ | |
8182 | io_timeout_remove_link(ctx, req); | |
8183 | } | |
8184 | ||
b7ddce3c PB |
8185 | static void io_cancel_defer_files(struct io_ring_ctx *ctx, |
8186 | struct files_struct *files) | |
8187 | { | |
8188 | struct io_defer_entry *de = NULL; | |
8189 | LIST_HEAD(list); | |
8190 | ||
8191 | spin_lock_irq(&ctx->completion_lock); | |
8192 | list_for_each_entry_reverse(de, &ctx->defer_list, list) { | |
c127a2a1 | 8193 | if (io_match_link_files(de->req, files)) { |
b7ddce3c PB |
8194 | list_cut_position(&list, &ctx->defer_list, &de->list); |
8195 | break; | |
8196 | } | |
8197 | } | |
8198 | spin_unlock_irq(&ctx->completion_lock); | |
8199 | ||
8200 | while (!list_empty(&list)) { | |
8201 | de = list_first_entry(&list, struct io_defer_entry, list); | |
8202 | list_del_init(&de->list); | |
8203 | req_set_fail_links(de->req); | |
8204 | io_put_req(de->req); | |
8205 | io_req_complete(de->req, -ECANCELED); | |
8206 | kfree(de); | |
8207 | } | |
8208 | } | |
8209 | ||
fcb323cc JA |
8210 | static void io_uring_cancel_files(struct io_ring_ctx *ctx, |
8211 | struct files_struct *files) | |
8212 | { | |
67c4d9e6 PB |
8213 | if (list_empty_careful(&ctx->inflight_list)) |
8214 | return; | |
8215 | ||
b7ddce3c | 8216 | io_cancel_defer_files(ctx, files); |
67c4d9e6 PB |
8217 | /* cancel all at once, should be faster than doing it one by one*/ |
8218 | io_wq_cancel_cb(ctx->io_wq, io_wq_files_match, files, true); | |
8219 | ||
fcb323cc | 8220 | while (!list_empty_careful(&ctx->inflight_list)) { |
d8f1b971 XW |
8221 | struct io_kiocb *cancel_req = NULL, *req; |
8222 | DEFINE_WAIT(wait); | |
fcb323cc JA |
8223 | |
8224 | spin_lock_irq(&ctx->inflight_lock); | |
8225 | list_for_each_entry(req, &ctx->inflight_list, inflight_entry) { | |
768134d4 JA |
8226 | if (req->work.files != files) |
8227 | continue; | |
8228 | /* req is being completed, ignore */ | |
8229 | if (!refcount_inc_not_zero(&req->refs)) | |
8230 | continue; | |
8231 | cancel_req = req; | |
8232 | break; | |
fcb323cc | 8233 | } |
768134d4 | 8234 | if (cancel_req) |
fcb323cc | 8235 | prepare_to_wait(&ctx->inflight_wait, &wait, |
768134d4 | 8236 | TASK_UNINTERRUPTIBLE); |
fcb323cc JA |
8237 | spin_unlock_irq(&ctx->inflight_lock); |
8238 | ||
768134d4 JA |
8239 | /* We need to keep going until we don't find a matching req */ |
8240 | if (!cancel_req) | |
fcb323cc | 8241 | break; |
bb175342 PB |
8242 | /* cancel this request, or head link requests */ |
8243 | io_attempt_cancel(ctx, cancel_req); | |
8244 | io_put_req(cancel_req); | |
6200b0ae JA |
8245 | /* cancellations _may_ trigger task work */ |
8246 | io_run_task_work(); | |
fcb323cc | 8247 | schedule(); |
d8f1b971 | 8248 | finish_wait(&ctx->inflight_wait, &wait); |
fcb323cc JA |
8249 | } |
8250 | } | |
8251 | ||
801dd57b | 8252 | static bool io_cancel_task_cb(struct io_wq_work *work, void *data) |
44e728b8 | 8253 | { |
801dd57b PB |
8254 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
8255 | struct task_struct *task = data; | |
44e728b8 | 8256 | |
f3606e3a | 8257 | return io_task_match(req, task); |
44e728b8 PB |
8258 | } |
8259 | ||
fcb323cc JA |
8260 | static int io_uring_flush(struct file *file, void *data) |
8261 | { | |
8262 | struct io_ring_ctx *ctx = file->private_data; | |
8263 | ||
8264 | io_uring_cancel_files(ctx, data); | |
6ab23144 JA |
8265 | |
8266 | /* | |
8267 | * If the task is going away, cancel work it may have pending | |
8268 | */ | |
801dd57b PB |
8269 | if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) |
8270 | io_wq_cancel_cb(ctx->io_wq, io_cancel_task_cb, current, true); | |
6ab23144 | 8271 | |
fcb323cc JA |
8272 | return 0; |
8273 | } | |
8274 | ||
6c5c240e RP |
8275 | static void *io_uring_validate_mmap_request(struct file *file, |
8276 | loff_t pgoff, size_t sz) | |
2b188cc1 | 8277 | { |
2b188cc1 | 8278 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 8279 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
8280 | struct page *page; |
8281 | void *ptr; | |
8282 | ||
8283 | switch (offset) { | |
8284 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
8285 | case IORING_OFF_CQ_RING: |
8286 | ptr = ctx->rings; | |
2b188cc1 JA |
8287 | break; |
8288 | case IORING_OFF_SQES: | |
8289 | ptr = ctx->sq_sqes; | |
8290 | break; | |
2b188cc1 | 8291 | default: |
6c5c240e | 8292 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
8293 | } |
8294 | ||
8295 | page = virt_to_head_page(ptr); | |
a50b854e | 8296 | if (sz > page_size(page)) |
6c5c240e RP |
8297 | return ERR_PTR(-EINVAL); |
8298 | ||
8299 | return ptr; | |
8300 | } | |
8301 | ||
8302 | #ifdef CONFIG_MMU | |
8303 | ||
8304 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
8305 | { | |
8306 | size_t sz = vma->vm_end - vma->vm_start; | |
8307 | unsigned long pfn; | |
8308 | void *ptr; | |
8309 | ||
8310 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
8311 | if (IS_ERR(ptr)) | |
8312 | return PTR_ERR(ptr); | |
2b188cc1 JA |
8313 | |
8314 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
8315 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
8316 | } | |
8317 | ||
6c5c240e RP |
8318 | #else /* !CONFIG_MMU */ |
8319 | ||
8320 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
8321 | { | |
8322 | return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL; | |
8323 | } | |
8324 | ||
8325 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
8326 | { | |
8327 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
8328 | } | |
8329 | ||
8330 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
8331 | unsigned long addr, unsigned long len, | |
8332 | unsigned long pgoff, unsigned long flags) | |
8333 | { | |
8334 | void *ptr; | |
8335 | ||
8336 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
8337 | if (IS_ERR(ptr)) | |
8338 | return PTR_ERR(ptr); | |
8339 | ||
8340 | return (unsigned long) ptr; | |
8341 | } | |
8342 | ||
8343 | #endif /* !CONFIG_MMU */ | |
8344 | ||
2b188cc1 JA |
8345 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
8346 | u32, min_complete, u32, flags, const sigset_t __user *, sig, | |
8347 | size_t, sigsz) | |
8348 | { | |
8349 | struct io_ring_ctx *ctx; | |
8350 | long ret = -EBADF; | |
8351 | int submitted = 0; | |
8352 | struct fd f; | |
8353 | ||
4c6e277c | 8354 | io_run_task_work(); |
b41e9852 | 8355 | |
6c271ce2 | 8356 | if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP)) |
2b188cc1 JA |
8357 | return -EINVAL; |
8358 | ||
8359 | f = fdget(fd); | |
8360 | if (!f.file) | |
8361 | return -EBADF; | |
8362 | ||
8363 | ret = -EOPNOTSUPP; | |
8364 | if (f.file->f_op != &io_uring_fops) | |
8365 | goto out_fput; | |
8366 | ||
8367 | ret = -ENXIO; | |
8368 | ctx = f.file->private_data; | |
8369 | if (!percpu_ref_tryget(&ctx->refs)) | |
8370 | goto out_fput; | |
8371 | ||
6c271ce2 JA |
8372 | /* |
8373 | * For SQ polling, the thread will do all submissions and completions. | |
8374 | * Just return the requested submit count, and wake the thread if | |
8375 | * we were asked to. | |
8376 | */ | |
b2a9eada | 8377 | ret = 0; |
6c271ce2 | 8378 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
c1edbf5f JA |
8379 | if (!list_empty_careful(&ctx->cq_overflow_list)) |
8380 | io_cqring_overflow_flush(ctx, false); | |
6c271ce2 JA |
8381 | if (flags & IORING_ENTER_SQ_WAKEUP) |
8382 | wake_up(&ctx->sqo_wait); | |
8383 | submitted = to_submit; | |
b2a9eada | 8384 | } else if (to_submit) { |
2b188cc1 | 8385 | mutex_lock(&ctx->uring_lock); |
0cdaf760 | 8386 | submitted = io_submit_sqes(ctx, to_submit, f.file, fd); |
2b188cc1 | 8387 | mutex_unlock(&ctx->uring_lock); |
7c504e65 PB |
8388 | |
8389 | if (submitted != to_submit) | |
8390 | goto out; | |
2b188cc1 JA |
8391 | } |
8392 | if (flags & IORING_ENTER_GETEVENTS) { | |
8393 | min_complete = min(min_complete, ctx->cq_entries); | |
8394 | ||
32b2244a XW |
8395 | /* |
8396 | * When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, user | |
8397 | * space applications don't need to do io completion events | |
8398 | * polling again, they can rely on io_sq_thread to do polling | |
8399 | * work, which can reduce cpu usage and uring_lock contention. | |
8400 | */ | |
8401 | if (ctx->flags & IORING_SETUP_IOPOLL && | |
8402 | !(ctx->flags & IORING_SETUP_SQPOLL)) { | |
7668b92a | 8403 | ret = io_iopoll_check(ctx, min_complete); |
def596e9 JA |
8404 | } else { |
8405 | ret = io_cqring_wait(ctx, min_complete, sig, sigsz); | |
8406 | } | |
2b188cc1 JA |
8407 | } |
8408 | ||
7c504e65 | 8409 | out: |
6805b32e | 8410 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
8411 | out_fput: |
8412 | fdput(f); | |
8413 | return submitted ? submitted : ret; | |
8414 | } | |
8415 | ||
bebdb65e | 8416 | #ifdef CONFIG_PROC_FS |
87ce955b JA |
8417 | static int io_uring_show_cred(int id, void *p, void *data) |
8418 | { | |
8419 | const struct cred *cred = p; | |
8420 | struct seq_file *m = data; | |
8421 | struct user_namespace *uns = seq_user_ns(m); | |
8422 | struct group_info *gi; | |
8423 | kernel_cap_t cap; | |
8424 | unsigned __capi; | |
8425 | int g; | |
8426 | ||
8427 | seq_printf(m, "%5d\n", id); | |
8428 | seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid)); | |
8429 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid)); | |
8430 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid)); | |
8431 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid)); | |
8432 | seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid)); | |
8433 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid)); | |
8434 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid)); | |
8435 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid)); | |
8436 | seq_puts(m, "\n\tGroups:\t"); | |
8437 | gi = cred->group_info; | |
8438 | for (g = 0; g < gi->ngroups; g++) { | |
8439 | seq_put_decimal_ull(m, g ? " " : "", | |
8440 | from_kgid_munged(uns, gi->gid[g])); | |
8441 | } | |
8442 | seq_puts(m, "\n\tCapEff:\t"); | |
8443 | cap = cred->cap_effective; | |
8444 | CAP_FOR_EACH_U32(__capi) | |
8445 | seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8); | |
8446 | seq_putc(m, '\n'); | |
8447 | return 0; | |
8448 | } | |
8449 | ||
8450 | static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m) | |
8451 | { | |
fad8e0de | 8452 | bool has_lock; |
87ce955b JA |
8453 | int i; |
8454 | ||
fad8e0de JA |
8455 | /* |
8456 | * Avoid ABBA deadlock between the seq lock and the io_uring mutex, | |
8457 | * since fdinfo case grabs it in the opposite direction of normal use | |
8458 | * cases. If we fail to get the lock, we just don't iterate any | |
8459 | * structures that could be going away outside the io_uring mutex. | |
8460 | */ | |
8461 | has_lock = mutex_trylock(&ctx->uring_lock); | |
8462 | ||
87ce955b | 8463 | seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files); |
fad8e0de | 8464 | for (i = 0; has_lock && i < ctx->nr_user_files; i++) { |
87ce955b JA |
8465 | struct fixed_file_table *table; |
8466 | struct file *f; | |
8467 | ||
8468 | table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT]; | |
8469 | f = table->files[i & IORING_FILE_TABLE_MASK]; | |
8470 | if (f) | |
8471 | seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname); | |
8472 | else | |
8473 | seq_printf(m, "%5u: <none>\n", i); | |
8474 | } | |
8475 | seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs); | |
fad8e0de | 8476 | for (i = 0; has_lock && i < ctx->nr_user_bufs; i++) { |
87ce955b JA |
8477 | struct io_mapped_ubuf *buf = &ctx->user_bufs[i]; |
8478 | ||
8479 | seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, | |
8480 | (unsigned int) buf->len); | |
8481 | } | |
fad8e0de | 8482 | if (has_lock && !idr_is_empty(&ctx->personality_idr)) { |
87ce955b JA |
8483 | seq_printf(m, "Personalities:\n"); |
8484 | idr_for_each(&ctx->personality_idr, io_uring_show_cred, m); | |
8485 | } | |
d7718a9d JA |
8486 | seq_printf(m, "PollList:\n"); |
8487 | spin_lock_irq(&ctx->completion_lock); | |
8488 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { | |
8489 | struct hlist_head *list = &ctx->cancel_hash[i]; | |
8490 | struct io_kiocb *req; | |
8491 | ||
8492 | hlist_for_each_entry(req, list, hash_node) | |
8493 | seq_printf(m, " op=%d, task_works=%d\n", req->opcode, | |
8494 | req->task->task_works != NULL); | |
8495 | } | |
8496 | spin_unlock_irq(&ctx->completion_lock); | |
fad8e0de JA |
8497 | if (has_lock) |
8498 | mutex_unlock(&ctx->uring_lock); | |
87ce955b JA |
8499 | } |
8500 | ||
8501 | static void io_uring_show_fdinfo(struct seq_file *m, struct file *f) | |
8502 | { | |
8503 | struct io_ring_ctx *ctx = f->private_data; | |
8504 | ||
8505 | if (percpu_ref_tryget(&ctx->refs)) { | |
8506 | __io_uring_show_fdinfo(ctx, m); | |
8507 | percpu_ref_put(&ctx->refs); | |
8508 | } | |
8509 | } | |
bebdb65e | 8510 | #endif |
87ce955b | 8511 | |
2b188cc1 JA |
8512 | static const struct file_operations io_uring_fops = { |
8513 | .release = io_uring_release, | |
fcb323cc | 8514 | .flush = io_uring_flush, |
2b188cc1 | 8515 | .mmap = io_uring_mmap, |
6c5c240e RP |
8516 | #ifndef CONFIG_MMU |
8517 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
8518 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
8519 | #endif | |
2b188cc1 JA |
8520 | .poll = io_uring_poll, |
8521 | .fasync = io_uring_fasync, | |
bebdb65e | 8522 | #ifdef CONFIG_PROC_FS |
87ce955b | 8523 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 8524 | #endif |
2b188cc1 JA |
8525 | }; |
8526 | ||
8527 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
8528 | struct io_uring_params *p) | |
8529 | { | |
75b28aff HV |
8530 | struct io_rings *rings; |
8531 | size_t size, sq_array_offset; | |
2b188cc1 | 8532 | |
bd740481 JA |
8533 | /* make sure these are sane, as we already accounted them */ |
8534 | ctx->sq_entries = p->sq_entries; | |
8535 | ctx->cq_entries = p->cq_entries; | |
8536 | ||
75b28aff HV |
8537 | size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset); |
8538 | if (size == SIZE_MAX) | |
8539 | return -EOVERFLOW; | |
8540 | ||
8541 | rings = io_mem_alloc(size); | |
8542 | if (!rings) | |
2b188cc1 JA |
8543 | return -ENOMEM; |
8544 | ||
75b28aff HV |
8545 | ctx->rings = rings; |
8546 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
8547 | rings->sq_ring_mask = p->sq_entries - 1; | |
8548 | rings->cq_ring_mask = p->cq_entries - 1; | |
8549 | rings->sq_ring_entries = p->sq_entries; | |
8550 | rings->cq_ring_entries = p->cq_entries; | |
8551 | ctx->sq_mask = rings->sq_ring_mask; | |
8552 | ctx->cq_mask = rings->cq_ring_mask; | |
2b188cc1 JA |
8553 | |
8554 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 JA |
8555 | if (size == SIZE_MAX) { |
8556 | io_mem_free(ctx->rings); | |
8557 | ctx->rings = NULL; | |
2b188cc1 | 8558 | return -EOVERFLOW; |
eb065d30 | 8559 | } |
2b188cc1 JA |
8560 | |
8561 | ctx->sq_sqes = io_mem_alloc(size); | |
eb065d30 JA |
8562 | if (!ctx->sq_sqes) { |
8563 | io_mem_free(ctx->rings); | |
8564 | ctx->rings = NULL; | |
2b188cc1 | 8565 | return -ENOMEM; |
eb065d30 | 8566 | } |
2b188cc1 | 8567 | |
2b188cc1 JA |
8568 | return 0; |
8569 | } | |
8570 | ||
8571 | /* | |
8572 | * Allocate an anonymous fd, this is what constitutes the application | |
8573 | * visible backing of an io_uring instance. The application mmaps this | |
8574 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
8575 | * we have to tie this fd to a socket for file garbage collection purposes. | |
8576 | */ | |
8577 | static int io_uring_get_fd(struct io_ring_ctx *ctx) | |
8578 | { | |
8579 | struct file *file; | |
8580 | int ret; | |
8581 | ||
8582 | #if defined(CONFIG_UNIX) | |
8583 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, | |
8584 | &ctx->ring_sock); | |
8585 | if (ret) | |
8586 | return ret; | |
8587 | #endif | |
8588 | ||
8589 | ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
8590 | if (ret < 0) | |
8591 | goto err; | |
8592 | ||
8593 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, | |
8594 | O_RDWR | O_CLOEXEC); | |
8595 | if (IS_ERR(file)) { | |
8596 | put_unused_fd(ret); | |
8597 | ret = PTR_ERR(file); | |
8598 | goto err; | |
8599 | } | |
8600 | ||
8601 | #if defined(CONFIG_UNIX) | |
8602 | ctx->ring_sock->file = file; | |
8603 | #endif | |
8604 | fd_install(ret, file); | |
8605 | return ret; | |
8606 | err: | |
8607 | #if defined(CONFIG_UNIX) | |
8608 | sock_release(ctx->ring_sock); | |
8609 | ctx->ring_sock = NULL; | |
8610 | #endif | |
8611 | return ret; | |
8612 | } | |
8613 | ||
7f13657d XW |
8614 | static int io_uring_create(unsigned entries, struct io_uring_params *p, |
8615 | struct io_uring_params __user *params) | |
2b188cc1 JA |
8616 | { |
8617 | struct user_struct *user = NULL; | |
8618 | struct io_ring_ctx *ctx; | |
aad5d8da | 8619 | bool limit_mem; |
2b188cc1 JA |
8620 | int ret; |
8621 | ||
8110c1a6 | 8622 | if (!entries) |
2b188cc1 | 8623 | return -EINVAL; |
8110c1a6 JA |
8624 | if (entries > IORING_MAX_ENTRIES) { |
8625 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
8626 | return -EINVAL; | |
8627 | entries = IORING_MAX_ENTRIES; | |
8628 | } | |
2b188cc1 JA |
8629 | |
8630 | /* | |
8631 | * Use twice as many entries for the CQ ring. It's possible for the | |
8632 | * application to drive a higher depth than the size of the SQ ring, | |
8633 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
8634 | * some flexibility in overcommitting a bit. If the application has |
8635 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
8636 | * of CQ ring entries manually. | |
2b188cc1 JA |
8637 | */ |
8638 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
8639 | if (p->flags & IORING_SETUP_CQSIZE) { |
8640 | /* | |
8641 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
8642 | * to a power-of-two, if it isn't already. We do NOT impose | |
8643 | * any cq vs sq ring sizing. | |
8644 | */ | |
8110c1a6 | 8645 | if (p->cq_entries < p->sq_entries) |
33a107f0 | 8646 | return -EINVAL; |
8110c1a6 JA |
8647 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
8648 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
8649 | return -EINVAL; | |
8650 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
8651 | } | |
33a107f0 JA |
8652 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
8653 | } else { | |
8654 | p->cq_entries = 2 * p->sq_entries; | |
8655 | } | |
2b188cc1 JA |
8656 | |
8657 | user = get_uid(current_user()); | |
aad5d8da | 8658 | limit_mem = !capable(CAP_IPC_LOCK); |
2b188cc1 | 8659 | |
aad5d8da | 8660 | if (limit_mem) { |
a087e2b5 | 8661 | ret = __io_account_mem(user, |
2b188cc1 JA |
8662 | ring_pages(p->sq_entries, p->cq_entries)); |
8663 | if (ret) { | |
8664 | free_uid(user); | |
8665 | return ret; | |
8666 | } | |
8667 | } | |
8668 | ||
8669 | ctx = io_ring_ctx_alloc(p); | |
8670 | if (!ctx) { | |
aad5d8da | 8671 | if (limit_mem) |
a087e2b5 | 8672 | __io_unaccount_mem(user, ring_pages(p->sq_entries, |
2b188cc1 JA |
8673 | p->cq_entries)); |
8674 | free_uid(user); | |
8675 | return -ENOMEM; | |
8676 | } | |
8677 | ctx->compat = in_compat_syscall(); | |
2b188cc1 | 8678 | ctx->user = user; |
0b8c0ec7 | 8679 | ctx->creds = get_current_cred(); |
2b188cc1 | 8680 | |
2aede0e4 JA |
8681 | ctx->sqo_task = get_task_struct(current); |
8682 | ||
8683 | /* | |
8684 | * This is just grabbed for accounting purposes. When a process exits, | |
8685 | * the mm is exited and dropped before the files, hence we need to hang | |
8686 | * on to this mm purely for the purposes of being able to unaccount | |
8687 | * memory (locked/pinned vm). It's not used for anything else. | |
8688 | */ | |
6b7898eb | 8689 | mmgrab(current->mm); |
2aede0e4 | 8690 | ctx->mm_account = current->mm; |
6b7898eb | 8691 | |
f74441e6 JA |
8692 | /* |
8693 | * Account memory _before_ installing the file descriptor. Once | |
8694 | * the descriptor is installed, it can get closed at any time. Also | |
8695 | * do this before hitting the general error path, as ring freeing | |
8696 | * will un-account as well. | |
8697 | */ | |
8698 | io_account_mem(ctx, ring_pages(p->sq_entries, p->cq_entries), | |
8699 | ACCT_LOCKED); | |
8700 | ctx->limit_mem = limit_mem; | |
8701 | ||
2b188cc1 JA |
8702 | ret = io_allocate_scq_urings(ctx, p); |
8703 | if (ret) | |
8704 | goto err; | |
8705 | ||
6c271ce2 | 8706 | ret = io_sq_offload_start(ctx, p); |
2b188cc1 JA |
8707 | if (ret) |
8708 | goto err; | |
8709 | ||
2b188cc1 | 8710 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
8711 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
8712 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
8713 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
8714 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
8715 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
8716 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
8717 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
8718 | |
8719 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
8720 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
8721 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
8722 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
8723 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
8724 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
8725 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
0d9b5b3a | 8726 | p->cq_off.flags = offsetof(struct io_rings, cq_flags); |
ac90f249 | 8727 | |
7f13657d XW |
8728 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
8729 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | | |
5769a351 JX |
8730 | IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL | |
8731 | IORING_FEAT_POLL_32BITS; | |
7f13657d XW |
8732 | |
8733 | if (copy_to_user(params, p, sizeof(*p))) { | |
8734 | ret = -EFAULT; | |
8735 | goto err; | |
8736 | } | |
d1719f70 | 8737 | |
044c1ab3 JA |
8738 | /* |
8739 | * Install ring fd as the very last thing, so we don't risk someone | |
8740 | * having closed it before we finish setup | |
8741 | */ | |
8742 | ret = io_uring_get_fd(ctx); | |
8743 | if (ret < 0) | |
8744 | goto err; | |
8745 | ||
c826bd7a | 8746 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
8747 | return ret; |
8748 | err: | |
8749 | io_ring_ctx_wait_and_kill(ctx); | |
8750 | return ret; | |
8751 | } | |
8752 | ||
8753 | /* | |
8754 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
8755 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
8756 | * params structure passed in. | |
8757 | */ | |
8758 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
8759 | { | |
8760 | struct io_uring_params p; | |
2b188cc1 JA |
8761 | int i; |
8762 | ||
8763 | if (copy_from_user(&p, params, sizeof(p))) | |
8764 | return -EFAULT; | |
8765 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
8766 | if (p.resv[i]) | |
8767 | return -EINVAL; | |
8768 | } | |
8769 | ||
6c271ce2 | 8770 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 8771 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
24369c2e | 8772 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ)) |
2b188cc1 JA |
8773 | return -EINVAL; |
8774 | ||
7f13657d | 8775 | return io_uring_create(entries, &p, params); |
2b188cc1 JA |
8776 | } |
8777 | ||
8778 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
8779 | struct io_uring_params __user *, params) | |
8780 | { | |
8781 | return io_uring_setup(entries, params); | |
8782 | } | |
8783 | ||
66f4af93 JA |
8784 | static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args) |
8785 | { | |
8786 | struct io_uring_probe *p; | |
8787 | size_t size; | |
8788 | int i, ret; | |
8789 | ||
8790 | size = struct_size(p, ops, nr_args); | |
8791 | if (size == SIZE_MAX) | |
8792 | return -EOVERFLOW; | |
8793 | p = kzalloc(size, GFP_KERNEL); | |
8794 | if (!p) | |
8795 | return -ENOMEM; | |
8796 | ||
8797 | ret = -EFAULT; | |
8798 | if (copy_from_user(p, arg, size)) | |
8799 | goto out; | |
8800 | ret = -EINVAL; | |
8801 | if (memchr_inv(p, 0, size)) | |
8802 | goto out; | |
8803 | ||
8804 | p->last_op = IORING_OP_LAST - 1; | |
8805 | if (nr_args > IORING_OP_LAST) | |
8806 | nr_args = IORING_OP_LAST; | |
8807 | ||
8808 | for (i = 0; i < nr_args; i++) { | |
8809 | p->ops[i].op = i; | |
8810 | if (!io_op_defs[i].not_supported) | |
8811 | p->ops[i].flags = IO_URING_OP_SUPPORTED; | |
8812 | } | |
8813 | p->ops_len = i; | |
8814 | ||
8815 | ret = 0; | |
8816 | if (copy_to_user(arg, p, size)) | |
8817 | ret = -EFAULT; | |
8818 | out: | |
8819 | kfree(p); | |
8820 | return ret; | |
8821 | } | |
8822 | ||
071698e1 JA |
8823 | static int io_register_personality(struct io_ring_ctx *ctx) |
8824 | { | |
8825 | const struct cred *creds = get_current_cred(); | |
8826 | int id; | |
8827 | ||
8828 | id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1, | |
8829 | USHRT_MAX, GFP_KERNEL); | |
8830 | if (id < 0) | |
8831 | put_cred(creds); | |
8832 | return id; | |
8833 | } | |
8834 | ||
8835 | static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) | |
8836 | { | |
8837 | const struct cred *old_creds; | |
8838 | ||
8839 | old_creds = idr_remove(&ctx->personality_idr, id); | |
8840 | if (old_creds) { | |
8841 | put_cred(old_creds); | |
8842 | return 0; | |
8843 | } | |
8844 | ||
8845 | return -EINVAL; | |
8846 | } | |
8847 | ||
8848 | static bool io_register_op_must_quiesce(int op) | |
8849 | { | |
8850 | switch (op) { | |
8851 | case IORING_UNREGISTER_FILES: | |
8852 | case IORING_REGISTER_FILES_UPDATE: | |
8853 | case IORING_REGISTER_PROBE: | |
8854 | case IORING_REGISTER_PERSONALITY: | |
8855 | case IORING_UNREGISTER_PERSONALITY: | |
8856 | return false; | |
8857 | default: | |
8858 | return true; | |
8859 | } | |
8860 | } | |
8861 | ||
edafccee JA |
8862 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
8863 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
8864 | __releases(ctx->uring_lock) |
8865 | __acquires(ctx->uring_lock) | |
edafccee JA |
8866 | { |
8867 | int ret; | |
8868 | ||
35fa71a0 JA |
8869 | /* |
8870 | * We're inside the ring mutex, if the ref is already dying, then | |
8871 | * someone else killed the ctx or is already going through | |
8872 | * io_uring_register(). | |
8873 | */ | |
8874 | if (percpu_ref_is_dying(&ctx->refs)) | |
8875 | return -ENXIO; | |
8876 | ||
071698e1 | 8877 | if (io_register_op_must_quiesce(opcode)) { |
05f3fb3c | 8878 | percpu_ref_kill(&ctx->refs); |
b19062a5 | 8879 | |
05f3fb3c JA |
8880 | /* |
8881 | * Drop uring mutex before waiting for references to exit. If | |
8882 | * another thread is currently inside io_uring_enter() it might | |
8883 | * need to grab the uring_lock to make progress. If we hold it | |
8884 | * here across the drain wait, then we can deadlock. It's safe | |
8885 | * to drop the mutex here, since no new references will come in | |
8886 | * after we've killed the percpu ref. | |
8887 | */ | |
8888 | mutex_unlock(&ctx->uring_lock); | |
0f158b4c | 8889 | ret = wait_for_completion_interruptible(&ctx->ref_comp); |
05f3fb3c | 8890 | mutex_lock(&ctx->uring_lock); |
c150368b JA |
8891 | if (ret) { |
8892 | percpu_ref_resurrect(&ctx->refs); | |
8893 | ret = -EINTR; | |
8894 | goto out; | |
8895 | } | |
05f3fb3c | 8896 | } |
edafccee JA |
8897 | |
8898 | switch (opcode) { | |
8899 | case IORING_REGISTER_BUFFERS: | |
8900 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
8901 | break; | |
8902 | case IORING_UNREGISTER_BUFFERS: | |
8903 | ret = -EINVAL; | |
8904 | if (arg || nr_args) | |
8905 | break; | |
8906 | ret = io_sqe_buffer_unregister(ctx); | |
8907 | break; | |
6b06314c JA |
8908 | case IORING_REGISTER_FILES: |
8909 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
8910 | break; | |
8911 | case IORING_UNREGISTER_FILES: | |
8912 | ret = -EINVAL; | |
8913 | if (arg || nr_args) | |
8914 | break; | |
8915 | ret = io_sqe_files_unregister(ctx); | |
8916 | break; | |
c3a31e60 JA |
8917 | case IORING_REGISTER_FILES_UPDATE: |
8918 | ret = io_sqe_files_update(ctx, arg, nr_args); | |
8919 | break; | |
9b402849 | 8920 | case IORING_REGISTER_EVENTFD: |
f2842ab5 | 8921 | case IORING_REGISTER_EVENTFD_ASYNC: |
9b402849 JA |
8922 | ret = -EINVAL; |
8923 | if (nr_args != 1) | |
8924 | break; | |
8925 | ret = io_eventfd_register(ctx, arg); | |
f2842ab5 JA |
8926 | if (ret) |
8927 | break; | |
8928 | if (opcode == IORING_REGISTER_EVENTFD_ASYNC) | |
8929 | ctx->eventfd_async = 1; | |
8930 | else | |
8931 | ctx->eventfd_async = 0; | |
9b402849 JA |
8932 | break; |
8933 | case IORING_UNREGISTER_EVENTFD: | |
8934 | ret = -EINVAL; | |
8935 | if (arg || nr_args) | |
8936 | break; | |
8937 | ret = io_eventfd_unregister(ctx); | |
8938 | break; | |
66f4af93 JA |
8939 | case IORING_REGISTER_PROBE: |
8940 | ret = -EINVAL; | |
8941 | if (!arg || nr_args > 256) | |
8942 | break; | |
8943 | ret = io_probe(ctx, arg, nr_args); | |
8944 | break; | |
071698e1 JA |
8945 | case IORING_REGISTER_PERSONALITY: |
8946 | ret = -EINVAL; | |
8947 | if (arg || nr_args) | |
8948 | break; | |
8949 | ret = io_register_personality(ctx); | |
8950 | break; | |
8951 | case IORING_UNREGISTER_PERSONALITY: | |
8952 | ret = -EINVAL; | |
8953 | if (arg) | |
8954 | break; | |
8955 | ret = io_unregister_personality(ctx, nr_args); | |
8956 | break; | |
edafccee JA |
8957 | default: |
8958 | ret = -EINVAL; | |
8959 | break; | |
8960 | } | |
8961 | ||
071698e1 | 8962 | if (io_register_op_must_quiesce(opcode)) { |
05f3fb3c | 8963 | /* bring the ctx back to life */ |
05f3fb3c | 8964 | percpu_ref_reinit(&ctx->refs); |
c150368b | 8965 | out: |
0f158b4c | 8966 | reinit_completion(&ctx->ref_comp); |
05f3fb3c | 8967 | } |
edafccee JA |
8968 | return ret; |
8969 | } | |
8970 | ||
8971 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
8972 | void __user *, arg, unsigned int, nr_args) | |
8973 | { | |
8974 | struct io_ring_ctx *ctx; | |
8975 | long ret = -EBADF; | |
8976 | struct fd f; | |
8977 | ||
8978 | f = fdget(fd); | |
8979 | if (!f.file) | |
8980 | return -EBADF; | |
8981 | ||
8982 | ret = -EOPNOTSUPP; | |
8983 | if (f.file->f_op != &io_uring_fops) | |
8984 | goto out_fput; | |
8985 | ||
8986 | ctx = f.file->private_data; | |
8987 | ||
8988 | mutex_lock(&ctx->uring_lock); | |
8989 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
8990 | mutex_unlock(&ctx->uring_lock); | |
c826bd7a DD |
8991 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, |
8992 | ctx->cq_ev_fd != NULL, ret); | |
edafccee JA |
8993 | out_fput: |
8994 | fdput(f); | |
8995 | return ret; | |
8996 | } | |
8997 | ||
2b188cc1 JA |
8998 | static int __init io_uring_init(void) |
8999 | { | |
d7f62e82 SM |
9000 | #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \ |
9001 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ | |
9002 | BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \ | |
9003 | } while (0) | |
9004 | ||
9005 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
9006 | __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename) | |
9007 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); | |
9008 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
9009 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
9010 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
9011 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
9012 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
9013 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
9014 | BUILD_BUG_SQE_ELEM(16, __u64, addr); | |
7d67af2c | 9015 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
9016 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
9017 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
9018 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
9019 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
9020 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
5769a351 JX |
9021 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u16, poll_events); |
9022 | BUILD_BUG_SQE_ELEM(28, __u32, poll32_events); | |
d7f62e82 SM |
9023 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); |
9024 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
9025 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
9026 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
9027 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
9028 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
9029 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
9030 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 9031 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
d7f62e82 SM |
9032 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
9033 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
9034 | BUILD_BUG_SQE_ELEM(42, __u16, personality); | |
7d67af2c | 9035 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
d7f62e82 | 9036 | |
d3656344 | 9037 | BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST); |
84557871 | 9038 | BUILD_BUG_ON(__REQ_F_LAST_BIT >= 8 * sizeof(int)); |
2b188cc1 JA |
9039 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); |
9040 | return 0; | |
9041 | }; | |
9042 | __initcall(io_uring_init); |