]>
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> | |
47 | #include <linux/refcount.h> | |
48 | #include <linux/uio.h> | |
6b47ee6e | 49 | #include <linux/bits.h> |
2b188cc1 JA |
50 | |
51 | #include <linux/sched/signal.h> | |
52 | #include <linux/fs.h> | |
53 | #include <linux/file.h> | |
54 | #include <linux/fdtable.h> | |
55 | #include <linux/mm.h> | |
56 | #include <linux/mman.h> | |
57 | #include <linux/mmu_context.h> | |
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> |
2b188cc1 | 81 | |
c826bd7a DD |
82 | #define CREATE_TRACE_POINTS |
83 | #include <trace/events/io_uring.h> | |
84 | ||
2b188cc1 JA |
85 | #include <uapi/linux/io_uring.h> |
86 | ||
87 | #include "internal.h" | |
561fb04a | 88 | #include "io-wq.h" |
2b188cc1 | 89 | |
5277deaa | 90 | #define IORING_MAX_ENTRIES 32768 |
33a107f0 | 91 | #define IORING_MAX_CQ_ENTRIES (2 * IORING_MAX_ENTRIES) |
65e19f54 JA |
92 | |
93 | /* | |
94 | * Shift of 9 is 512 entries, or exactly one page on 64-bit archs | |
95 | */ | |
96 | #define IORING_FILE_TABLE_SHIFT 9 | |
97 | #define IORING_MAX_FILES_TABLE (1U << IORING_FILE_TABLE_SHIFT) | |
98 | #define IORING_FILE_TABLE_MASK (IORING_MAX_FILES_TABLE - 1) | |
99 | #define IORING_MAX_FIXED_FILES (64 * IORING_MAX_FILES_TABLE) | |
2b188cc1 JA |
100 | |
101 | struct io_uring { | |
102 | u32 head ____cacheline_aligned_in_smp; | |
103 | u32 tail ____cacheline_aligned_in_smp; | |
104 | }; | |
105 | ||
1e84b97b | 106 | /* |
75b28aff HV |
107 | * This data is shared with the application through the mmap at offsets |
108 | * IORING_OFF_SQ_RING and IORING_OFF_CQ_RING. | |
1e84b97b SB |
109 | * |
110 | * The offsets to the member fields are published through struct | |
111 | * io_sqring_offsets when calling io_uring_setup. | |
112 | */ | |
75b28aff | 113 | struct io_rings { |
1e84b97b SB |
114 | /* |
115 | * Head and tail offsets into the ring; the offsets need to be | |
116 | * masked to get valid indices. | |
117 | * | |
75b28aff HV |
118 | * The kernel controls head of the sq ring and the tail of the cq ring, |
119 | * and the application controls tail of the sq ring and the head of the | |
120 | * cq ring. | |
1e84b97b | 121 | */ |
75b28aff | 122 | struct io_uring sq, cq; |
1e84b97b | 123 | /* |
75b28aff | 124 | * Bitmasks to apply to head and tail offsets (constant, equals |
1e84b97b SB |
125 | * ring_entries - 1) |
126 | */ | |
75b28aff HV |
127 | u32 sq_ring_mask, cq_ring_mask; |
128 | /* Ring sizes (constant, power of 2) */ | |
129 | u32 sq_ring_entries, cq_ring_entries; | |
1e84b97b SB |
130 | /* |
131 | * Number of invalid entries dropped by the kernel due to | |
132 | * invalid index stored in array | |
133 | * | |
134 | * Written by the kernel, shouldn't be modified by the | |
135 | * application (i.e. get number of "new events" by comparing to | |
136 | * cached value). | |
137 | * | |
138 | * After a new SQ head value was read by the application this | |
139 | * counter includes all submissions that were dropped reaching | |
140 | * the new SQ head (and possibly more). | |
141 | */ | |
75b28aff | 142 | u32 sq_dropped; |
1e84b97b SB |
143 | /* |
144 | * Runtime flags | |
145 | * | |
146 | * Written by the kernel, shouldn't be modified by the | |
147 | * application. | |
148 | * | |
149 | * The application needs a full memory barrier before checking | |
150 | * for IORING_SQ_NEED_WAKEUP after updating the sq tail. | |
151 | */ | |
75b28aff | 152 | u32 sq_flags; |
1e84b97b SB |
153 | /* |
154 | * Number of completion events lost because the queue was full; | |
155 | * this should be avoided by the application by making sure | |
0b4295b5 | 156 | * there are not more requests pending than there is space in |
1e84b97b SB |
157 | * the completion queue. |
158 | * | |
159 | * Written by the kernel, shouldn't be modified by the | |
160 | * application (i.e. get number of "new events" by comparing to | |
161 | * cached value). | |
162 | * | |
163 | * As completion events come in out of order this counter is not | |
164 | * ordered with any other data. | |
165 | */ | |
75b28aff | 166 | u32 cq_overflow; |
1e84b97b SB |
167 | /* |
168 | * Ring buffer of completion events. | |
169 | * | |
170 | * The kernel writes completion events fresh every time they are | |
171 | * produced, so the application is allowed to modify pending | |
172 | * entries. | |
173 | */ | |
75b28aff | 174 | struct io_uring_cqe cqes[] ____cacheline_aligned_in_smp; |
2b188cc1 JA |
175 | }; |
176 | ||
edafccee JA |
177 | struct io_mapped_ubuf { |
178 | u64 ubuf; | |
179 | size_t len; | |
180 | struct bio_vec *bvec; | |
181 | unsigned int nr_bvecs; | |
182 | }; | |
183 | ||
65e19f54 JA |
184 | struct fixed_file_table { |
185 | struct file **files; | |
31b51510 JA |
186 | }; |
187 | ||
05f3fb3c JA |
188 | struct fixed_file_data { |
189 | struct fixed_file_table *table; | |
190 | struct io_ring_ctx *ctx; | |
191 | ||
192 | struct percpu_ref refs; | |
193 | struct llist_head put_llist; | |
05f3fb3c JA |
194 | struct work_struct ref_work; |
195 | struct completion done; | |
196 | }; | |
197 | ||
2b188cc1 JA |
198 | struct io_ring_ctx { |
199 | struct { | |
200 | struct percpu_ref refs; | |
201 | } ____cacheline_aligned_in_smp; | |
202 | ||
203 | struct { | |
204 | unsigned int flags; | |
e1d85334 RD |
205 | unsigned int compat: 1; |
206 | unsigned int account_mem: 1; | |
207 | unsigned int cq_overflow_flushed: 1; | |
208 | unsigned int drain_next: 1; | |
209 | unsigned int eventfd_async: 1; | |
2b188cc1 | 210 | |
75b28aff HV |
211 | /* |
212 | * Ring buffer of indices into array of io_uring_sqe, which is | |
213 | * mmapped by the application using the IORING_OFF_SQES offset. | |
214 | * | |
215 | * This indirection could e.g. be used to assign fixed | |
216 | * io_uring_sqe entries to operations and only submit them to | |
217 | * the queue when needed. | |
218 | * | |
219 | * The kernel modifies neither the indices array nor the entries | |
220 | * array. | |
221 | */ | |
222 | u32 *sq_array; | |
2b188cc1 JA |
223 | unsigned cached_sq_head; |
224 | unsigned sq_entries; | |
225 | unsigned sq_mask; | |
6c271ce2 | 226 | unsigned sq_thread_idle; |
498ccd9e | 227 | unsigned cached_sq_dropped; |
206aefde | 228 | atomic_t cached_cq_overflow; |
ad3eb2c8 | 229 | unsigned long sq_check_overflow; |
de0617e4 JA |
230 | |
231 | struct list_head defer_list; | |
5262f567 | 232 | struct list_head timeout_list; |
1d7bb1d5 | 233 | struct list_head cq_overflow_list; |
fcb323cc JA |
234 | |
235 | wait_queue_head_t inflight_wait; | |
ad3eb2c8 | 236 | struct io_uring_sqe *sq_sqes; |
2b188cc1 JA |
237 | } ____cacheline_aligned_in_smp; |
238 | ||
206aefde JA |
239 | struct io_rings *rings; |
240 | ||
2b188cc1 | 241 | /* IO offload */ |
561fb04a | 242 | struct io_wq *io_wq; |
6c271ce2 | 243 | struct task_struct *sqo_thread; /* if using sq thread polling */ |
2b188cc1 | 244 | struct mm_struct *sqo_mm; |
6c271ce2 | 245 | wait_queue_head_t sqo_wait; |
75b28aff | 246 | |
6b06314c JA |
247 | /* |
248 | * If used, fixed file set. Writers must ensure that ->refs is dead, | |
249 | * readers must ensure that ->refs is alive as long as the file* is | |
250 | * used. Only updated through io_uring_register(2). | |
251 | */ | |
05f3fb3c | 252 | struct fixed_file_data *file_data; |
6b06314c | 253 | unsigned nr_user_files; |
b14cca0c PB |
254 | int ring_fd; |
255 | struct file *ring_file; | |
6b06314c | 256 | |
edafccee JA |
257 | /* if used, fixed mapped user buffers */ |
258 | unsigned nr_user_bufs; | |
259 | struct io_mapped_ubuf *user_bufs; | |
260 | ||
2b188cc1 JA |
261 | struct user_struct *user; |
262 | ||
0b8c0ec7 | 263 | const struct cred *creds; |
181e448d | 264 | |
206aefde JA |
265 | /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */ |
266 | struct completion *completions; | |
267 | ||
0ddf92e8 JA |
268 | /* if all else fails... */ |
269 | struct io_kiocb *fallback_req; | |
270 | ||
206aefde JA |
271 | #if defined(CONFIG_UNIX) |
272 | struct socket *ring_sock; | |
273 | #endif | |
274 | ||
071698e1 JA |
275 | struct idr personality_idr; |
276 | ||
206aefde JA |
277 | struct { |
278 | unsigned cached_cq_tail; | |
279 | unsigned cq_entries; | |
280 | unsigned cq_mask; | |
281 | atomic_t cq_timeouts; | |
ad3eb2c8 | 282 | unsigned long cq_check_overflow; |
206aefde JA |
283 | struct wait_queue_head cq_wait; |
284 | struct fasync_struct *cq_fasync; | |
285 | struct eventfd_ctx *cq_ev_fd; | |
286 | } ____cacheline_aligned_in_smp; | |
2b188cc1 JA |
287 | |
288 | struct { | |
289 | struct mutex uring_lock; | |
290 | wait_queue_head_t wait; | |
291 | } ____cacheline_aligned_in_smp; | |
292 | ||
293 | struct { | |
294 | spinlock_t completion_lock; | |
e94f141b | 295 | |
def596e9 JA |
296 | /* |
297 | * ->poll_list is protected by the ctx->uring_lock for | |
298 | * io_uring instances that don't use IORING_SETUP_SQPOLL. | |
299 | * For SQPOLL, only the single threaded io_sq_thread() will | |
300 | * manipulate the list, hence no extra locking is needed there. | |
301 | */ | |
302 | struct list_head poll_list; | |
78076bb6 JA |
303 | struct hlist_head *cancel_hash; |
304 | unsigned cancel_hash_bits; | |
e94f141b | 305 | bool poll_multi_file; |
31b51510 | 306 | |
fcb323cc JA |
307 | spinlock_t inflight_lock; |
308 | struct list_head inflight_list; | |
2b188cc1 | 309 | } ____cacheline_aligned_in_smp; |
2b188cc1 JA |
310 | }; |
311 | ||
09bb8394 JA |
312 | /* |
313 | * First field must be the file pointer in all the | |
314 | * iocb unions! See also 'struct kiocb' in <linux/fs.h> | |
315 | */ | |
221c5eb2 JA |
316 | struct io_poll_iocb { |
317 | struct file *file; | |
0969e783 JA |
318 | union { |
319 | struct wait_queue_head *head; | |
320 | u64 addr; | |
321 | }; | |
221c5eb2 | 322 | __poll_t events; |
8c838788 | 323 | bool done; |
221c5eb2 | 324 | bool canceled; |
392edb45 | 325 | struct wait_queue_entry wait; |
221c5eb2 JA |
326 | }; |
327 | ||
b5dba59e JA |
328 | struct io_close { |
329 | struct file *file; | |
330 | struct file *put_file; | |
331 | int fd; | |
332 | }; | |
333 | ||
ad8a48ac JA |
334 | struct io_timeout_data { |
335 | struct io_kiocb *req; | |
336 | struct hrtimer timer; | |
337 | struct timespec64 ts; | |
338 | enum hrtimer_mode mode; | |
cc42e0ac | 339 | u32 seq_offset; |
ad8a48ac JA |
340 | }; |
341 | ||
8ed8d3c3 JA |
342 | struct io_accept { |
343 | struct file *file; | |
344 | struct sockaddr __user *addr; | |
345 | int __user *addr_len; | |
346 | int flags; | |
347 | }; | |
348 | ||
349 | struct io_sync { | |
350 | struct file *file; | |
351 | loff_t len; | |
352 | loff_t off; | |
353 | int flags; | |
d63d1b5e | 354 | int mode; |
8ed8d3c3 JA |
355 | }; |
356 | ||
fbf23849 JA |
357 | struct io_cancel { |
358 | struct file *file; | |
359 | u64 addr; | |
360 | }; | |
361 | ||
b29472ee JA |
362 | struct io_timeout { |
363 | struct file *file; | |
364 | u64 addr; | |
365 | int flags; | |
26a61679 | 366 | unsigned count; |
b29472ee JA |
367 | }; |
368 | ||
9adbd45d JA |
369 | struct io_rw { |
370 | /* NOTE: kiocb has the file as the first member, so don't do it here */ | |
371 | struct kiocb kiocb; | |
372 | u64 addr; | |
373 | u64 len; | |
374 | }; | |
375 | ||
3fbb51c1 JA |
376 | struct io_connect { |
377 | struct file *file; | |
378 | struct sockaddr __user *addr; | |
379 | int addr_len; | |
380 | }; | |
381 | ||
e47293fd JA |
382 | struct io_sr_msg { |
383 | struct file *file; | |
fddaface JA |
384 | union { |
385 | struct user_msghdr __user *msg; | |
386 | void __user *buf; | |
387 | }; | |
e47293fd | 388 | int msg_flags; |
fddaface | 389 | size_t len; |
e47293fd JA |
390 | }; |
391 | ||
15b71abe JA |
392 | struct io_open { |
393 | struct file *file; | |
394 | int dfd; | |
eddc7ef5 | 395 | union { |
eddc7ef5 JA |
396 | unsigned mask; |
397 | }; | |
15b71abe | 398 | struct filename *filename; |
eddc7ef5 | 399 | struct statx __user *buffer; |
c12cedf2 | 400 | struct open_how how; |
15b71abe JA |
401 | }; |
402 | ||
05f3fb3c JA |
403 | struct io_files_update { |
404 | struct file *file; | |
405 | u64 arg; | |
406 | u32 nr_args; | |
407 | u32 offset; | |
408 | }; | |
409 | ||
4840e418 JA |
410 | struct io_fadvise { |
411 | struct file *file; | |
412 | u64 offset; | |
413 | u32 len; | |
414 | u32 advice; | |
415 | }; | |
416 | ||
c1ca757b JA |
417 | struct io_madvise { |
418 | struct file *file; | |
419 | u64 addr; | |
420 | u32 len; | |
421 | u32 advice; | |
422 | }; | |
423 | ||
3e4827b0 JA |
424 | struct io_epoll { |
425 | struct file *file; | |
426 | int epfd; | |
427 | int op; | |
428 | int fd; | |
429 | struct epoll_event event; | |
e47293fd JA |
430 | }; |
431 | ||
7d67af2c PB |
432 | struct io_splice { |
433 | struct file *file_out; | |
434 | struct file *file_in; | |
435 | loff_t off_out; | |
436 | loff_t off_in; | |
437 | u64 len; | |
438 | unsigned int flags; | |
439 | }; | |
440 | ||
f499a021 JA |
441 | struct io_async_connect { |
442 | struct sockaddr_storage address; | |
443 | }; | |
444 | ||
03b1230c JA |
445 | struct io_async_msghdr { |
446 | struct iovec fast_iov[UIO_FASTIOV]; | |
447 | struct iovec *iov; | |
448 | struct sockaddr __user *uaddr; | |
449 | struct msghdr msg; | |
b537916c | 450 | struct sockaddr_storage addr; |
03b1230c JA |
451 | }; |
452 | ||
f67676d1 JA |
453 | struct io_async_rw { |
454 | struct iovec fast_iov[UIO_FASTIOV]; | |
455 | struct iovec *iov; | |
456 | ssize_t nr_segs; | |
457 | ssize_t size; | |
458 | }; | |
459 | ||
1a6b74fc | 460 | struct io_async_ctx { |
f67676d1 JA |
461 | union { |
462 | struct io_async_rw rw; | |
03b1230c | 463 | struct io_async_msghdr msg; |
f499a021 | 464 | struct io_async_connect connect; |
2d28390a | 465 | struct io_timeout_data timeout; |
f67676d1 | 466 | }; |
1a6b74fc JA |
467 | }; |
468 | ||
6b47ee6e PB |
469 | enum { |
470 | REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT, | |
471 | REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT, | |
472 | REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT, | |
473 | REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT, | |
474 | REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT, | |
475 | ||
476 | REQ_F_LINK_NEXT_BIT, | |
477 | REQ_F_FAIL_LINK_BIT, | |
478 | REQ_F_INFLIGHT_BIT, | |
479 | REQ_F_CUR_POS_BIT, | |
480 | REQ_F_NOWAIT_BIT, | |
481 | REQ_F_IOPOLL_COMPLETED_BIT, | |
482 | REQ_F_LINK_TIMEOUT_BIT, | |
483 | REQ_F_TIMEOUT_BIT, | |
484 | REQ_F_ISREG_BIT, | |
485 | REQ_F_MUST_PUNT_BIT, | |
486 | REQ_F_TIMEOUT_NOSEQ_BIT, | |
487 | REQ_F_COMP_LOCKED_BIT, | |
99bc4c38 | 488 | REQ_F_NEED_CLEANUP_BIT, |
2ca10259 | 489 | REQ_F_OVERFLOW_BIT, |
6b47ee6e PB |
490 | }; |
491 | ||
492 | enum { | |
493 | /* ctx owns file */ | |
494 | REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT), | |
495 | /* drain existing IO first */ | |
496 | REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT), | |
497 | /* linked sqes */ | |
498 | REQ_F_LINK = BIT(REQ_F_LINK_BIT), | |
499 | /* doesn't sever on completion < 0 */ | |
500 | REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT), | |
501 | /* IOSQE_ASYNC */ | |
502 | REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT), | |
503 | ||
504 | /* already grabbed next link */ | |
505 | REQ_F_LINK_NEXT = BIT(REQ_F_LINK_NEXT_BIT), | |
506 | /* fail rest of links */ | |
507 | REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT), | |
508 | /* on inflight list */ | |
509 | REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT), | |
510 | /* read/write uses file position */ | |
511 | REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT), | |
512 | /* must not punt to workers */ | |
513 | REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT), | |
514 | /* polled IO has completed */ | |
515 | REQ_F_IOPOLL_COMPLETED = BIT(REQ_F_IOPOLL_COMPLETED_BIT), | |
516 | /* has linked timeout */ | |
517 | REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT), | |
518 | /* timeout request */ | |
519 | REQ_F_TIMEOUT = BIT(REQ_F_TIMEOUT_BIT), | |
520 | /* regular file */ | |
521 | REQ_F_ISREG = BIT(REQ_F_ISREG_BIT), | |
522 | /* must be punted even for NONBLOCK */ | |
523 | REQ_F_MUST_PUNT = BIT(REQ_F_MUST_PUNT_BIT), | |
524 | /* no timeout sequence */ | |
525 | REQ_F_TIMEOUT_NOSEQ = BIT(REQ_F_TIMEOUT_NOSEQ_BIT), | |
526 | /* completion under lock */ | |
527 | REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT), | |
99bc4c38 PB |
528 | /* needs cleanup */ |
529 | REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT), | |
2ca10259 JA |
530 | /* in overflow list */ |
531 | REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT), | |
6b47ee6e PB |
532 | }; |
533 | ||
09bb8394 JA |
534 | /* |
535 | * NOTE! Each of the iocb union members has the file pointer | |
536 | * as the first entry in their struct definition. So you can | |
537 | * access the file pointer through any of the sub-structs, | |
538 | * or directly as just 'ki_filp' in this struct. | |
539 | */ | |
2b188cc1 | 540 | struct io_kiocb { |
221c5eb2 | 541 | union { |
09bb8394 | 542 | struct file *file; |
9adbd45d | 543 | struct io_rw rw; |
221c5eb2 | 544 | struct io_poll_iocb poll; |
8ed8d3c3 JA |
545 | struct io_accept accept; |
546 | struct io_sync sync; | |
fbf23849 | 547 | struct io_cancel cancel; |
b29472ee | 548 | struct io_timeout timeout; |
3fbb51c1 | 549 | struct io_connect connect; |
e47293fd | 550 | struct io_sr_msg sr_msg; |
15b71abe | 551 | struct io_open open; |
b5dba59e | 552 | struct io_close close; |
05f3fb3c | 553 | struct io_files_update files_update; |
4840e418 | 554 | struct io_fadvise fadvise; |
c1ca757b | 555 | struct io_madvise madvise; |
3e4827b0 | 556 | struct io_epoll epoll; |
7d67af2c | 557 | struct io_splice splice; |
221c5eb2 | 558 | }; |
2b188cc1 | 559 | |
1a6b74fc | 560 | struct io_async_ctx *io; |
cf6fd4bd | 561 | bool needs_fixed_file; |
d625c6ee | 562 | u8 opcode; |
2b188cc1 JA |
563 | |
564 | struct io_ring_ctx *ctx; | |
eac406c6 JA |
565 | union { |
566 | struct list_head list; | |
78076bb6 | 567 | struct hlist_node hash_node; |
eac406c6 | 568 | }; |
9e645e11 | 569 | struct list_head link_list; |
2b188cc1 | 570 | unsigned int flags; |
c16361c1 | 571 | refcount_t refs; |
2b188cc1 | 572 | u64 user_data; |
9e645e11 | 573 | u32 result; |
de0617e4 | 574 | u32 sequence; |
2b188cc1 | 575 | |
fcb323cc JA |
576 | struct list_head inflight_entry; |
577 | ||
b41e9852 JA |
578 | union { |
579 | /* | |
580 | * Only commands that never go async can use the below fields, | |
581 | * obviously. Right now only IORING_OP_POLL_ADD uses them. | |
582 | */ | |
583 | struct { | |
584 | struct task_struct *task; | |
585 | struct callback_head task_work; | |
586 | }; | |
587 | struct io_wq_work work; | |
588 | }; | |
2b188cc1 JA |
589 | }; |
590 | ||
591 | #define IO_PLUG_THRESHOLD 2 | |
def596e9 | 592 | #define IO_IOPOLL_BATCH 8 |
2b188cc1 | 593 | |
9a56a232 JA |
594 | struct io_submit_state { |
595 | struct blk_plug plug; | |
596 | ||
2579f913 JA |
597 | /* |
598 | * io_kiocb alloc cache | |
599 | */ | |
600 | void *reqs[IO_IOPOLL_BATCH]; | |
6c8a3134 | 601 | unsigned int free_reqs; |
2579f913 | 602 | |
9a56a232 JA |
603 | /* |
604 | * File reference cache | |
605 | */ | |
606 | struct file *file; | |
607 | unsigned int fd; | |
608 | unsigned int has_refs; | |
609 | unsigned int used_refs; | |
610 | unsigned int ios_left; | |
611 | }; | |
612 | ||
d3656344 JA |
613 | struct io_op_def { |
614 | /* needs req->io allocated for deferral/async */ | |
615 | unsigned async_ctx : 1; | |
616 | /* needs current->mm setup, does mm access */ | |
617 | unsigned needs_mm : 1; | |
618 | /* needs req->file assigned */ | |
619 | unsigned needs_file : 1; | |
620 | /* needs req->file assigned IFF fd is >= 0 */ | |
621 | unsigned fd_non_neg : 1; | |
622 | /* hash wq insertion if file is a regular file */ | |
623 | unsigned hash_reg_file : 1; | |
624 | /* unbound wq insertion if file is a non-regular file */ | |
625 | unsigned unbound_nonreg_file : 1; | |
66f4af93 JA |
626 | /* opcode is not supported by this kernel */ |
627 | unsigned not_supported : 1; | |
f86cd20c JA |
628 | /* needs file table */ |
629 | unsigned file_table : 1; | |
ff002b30 JA |
630 | /* needs ->fs */ |
631 | unsigned needs_fs : 1; | |
d3656344 JA |
632 | }; |
633 | ||
634 | static const struct io_op_def io_op_defs[] = { | |
0463b6c5 PB |
635 | [IORING_OP_NOP] = {}, |
636 | [IORING_OP_READV] = { | |
d3656344 JA |
637 | .async_ctx = 1, |
638 | .needs_mm = 1, | |
639 | .needs_file = 1, | |
640 | .unbound_nonreg_file = 1, | |
641 | }, | |
0463b6c5 | 642 | [IORING_OP_WRITEV] = { |
d3656344 JA |
643 | .async_ctx = 1, |
644 | .needs_mm = 1, | |
645 | .needs_file = 1, | |
646 | .hash_reg_file = 1, | |
647 | .unbound_nonreg_file = 1, | |
648 | }, | |
0463b6c5 | 649 | [IORING_OP_FSYNC] = { |
d3656344 JA |
650 | .needs_file = 1, |
651 | }, | |
0463b6c5 | 652 | [IORING_OP_READ_FIXED] = { |
d3656344 JA |
653 | .needs_file = 1, |
654 | .unbound_nonreg_file = 1, | |
655 | }, | |
0463b6c5 | 656 | [IORING_OP_WRITE_FIXED] = { |
d3656344 JA |
657 | .needs_file = 1, |
658 | .hash_reg_file = 1, | |
659 | .unbound_nonreg_file = 1, | |
660 | }, | |
0463b6c5 | 661 | [IORING_OP_POLL_ADD] = { |
d3656344 JA |
662 | .needs_file = 1, |
663 | .unbound_nonreg_file = 1, | |
664 | }, | |
0463b6c5 PB |
665 | [IORING_OP_POLL_REMOVE] = {}, |
666 | [IORING_OP_SYNC_FILE_RANGE] = { | |
d3656344 JA |
667 | .needs_file = 1, |
668 | }, | |
0463b6c5 | 669 | [IORING_OP_SENDMSG] = { |
d3656344 JA |
670 | .async_ctx = 1, |
671 | .needs_mm = 1, | |
672 | .needs_file = 1, | |
673 | .unbound_nonreg_file = 1, | |
ff002b30 | 674 | .needs_fs = 1, |
d3656344 | 675 | }, |
0463b6c5 | 676 | [IORING_OP_RECVMSG] = { |
d3656344 JA |
677 | .async_ctx = 1, |
678 | .needs_mm = 1, | |
679 | .needs_file = 1, | |
680 | .unbound_nonreg_file = 1, | |
ff002b30 | 681 | .needs_fs = 1, |
d3656344 | 682 | }, |
0463b6c5 | 683 | [IORING_OP_TIMEOUT] = { |
d3656344 JA |
684 | .async_ctx = 1, |
685 | .needs_mm = 1, | |
686 | }, | |
0463b6c5 PB |
687 | [IORING_OP_TIMEOUT_REMOVE] = {}, |
688 | [IORING_OP_ACCEPT] = { | |
d3656344 JA |
689 | .needs_mm = 1, |
690 | .needs_file = 1, | |
691 | .unbound_nonreg_file = 1, | |
f86cd20c | 692 | .file_table = 1, |
d3656344 | 693 | }, |
0463b6c5 PB |
694 | [IORING_OP_ASYNC_CANCEL] = {}, |
695 | [IORING_OP_LINK_TIMEOUT] = { | |
d3656344 JA |
696 | .async_ctx = 1, |
697 | .needs_mm = 1, | |
698 | }, | |
0463b6c5 | 699 | [IORING_OP_CONNECT] = { |
d3656344 JA |
700 | .async_ctx = 1, |
701 | .needs_mm = 1, | |
702 | .needs_file = 1, | |
703 | .unbound_nonreg_file = 1, | |
704 | }, | |
0463b6c5 | 705 | [IORING_OP_FALLOCATE] = { |
d3656344 JA |
706 | .needs_file = 1, |
707 | }, | |
0463b6c5 | 708 | [IORING_OP_OPENAT] = { |
d3656344 JA |
709 | .needs_file = 1, |
710 | .fd_non_neg = 1, | |
f86cd20c | 711 | .file_table = 1, |
ff002b30 | 712 | .needs_fs = 1, |
d3656344 | 713 | }, |
0463b6c5 | 714 | [IORING_OP_CLOSE] = { |
d3656344 | 715 | .needs_file = 1, |
f86cd20c | 716 | .file_table = 1, |
d3656344 | 717 | }, |
0463b6c5 | 718 | [IORING_OP_FILES_UPDATE] = { |
d3656344 | 719 | .needs_mm = 1, |
f86cd20c | 720 | .file_table = 1, |
d3656344 | 721 | }, |
0463b6c5 | 722 | [IORING_OP_STATX] = { |
d3656344 JA |
723 | .needs_mm = 1, |
724 | .needs_file = 1, | |
725 | .fd_non_neg = 1, | |
ff002b30 | 726 | .needs_fs = 1, |
d3656344 | 727 | }, |
0463b6c5 | 728 | [IORING_OP_READ] = { |
3a6820f2 JA |
729 | .needs_mm = 1, |
730 | .needs_file = 1, | |
731 | .unbound_nonreg_file = 1, | |
732 | }, | |
0463b6c5 | 733 | [IORING_OP_WRITE] = { |
3a6820f2 JA |
734 | .needs_mm = 1, |
735 | .needs_file = 1, | |
736 | .unbound_nonreg_file = 1, | |
737 | }, | |
0463b6c5 | 738 | [IORING_OP_FADVISE] = { |
4840e418 JA |
739 | .needs_file = 1, |
740 | }, | |
0463b6c5 | 741 | [IORING_OP_MADVISE] = { |
c1ca757b JA |
742 | .needs_mm = 1, |
743 | }, | |
0463b6c5 | 744 | [IORING_OP_SEND] = { |
fddaface JA |
745 | .needs_mm = 1, |
746 | .needs_file = 1, | |
747 | .unbound_nonreg_file = 1, | |
748 | }, | |
0463b6c5 | 749 | [IORING_OP_RECV] = { |
fddaface JA |
750 | .needs_mm = 1, |
751 | .needs_file = 1, | |
752 | .unbound_nonreg_file = 1, | |
753 | }, | |
0463b6c5 | 754 | [IORING_OP_OPENAT2] = { |
cebdb986 JA |
755 | .needs_file = 1, |
756 | .fd_non_neg = 1, | |
f86cd20c | 757 | .file_table = 1, |
ff002b30 | 758 | .needs_fs = 1, |
cebdb986 | 759 | }, |
3e4827b0 JA |
760 | [IORING_OP_EPOLL_CTL] = { |
761 | .unbound_nonreg_file = 1, | |
762 | .file_table = 1, | |
763 | }, | |
7d67af2c PB |
764 | [IORING_OP_SPLICE] = { |
765 | .needs_file = 1, | |
766 | .hash_reg_file = 1, | |
767 | .unbound_nonreg_file = 1, | |
768 | } | |
d3656344 JA |
769 | }; |
770 | ||
561fb04a | 771 | static void io_wq_submit_work(struct io_wq_work **workptr); |
78e19bbe | 772 | static void io_cqring_fill_event(struct io_kiocb *req, long res); |
ec9c02ad | 773 | static void io_put_req(struct io_kiocb *req); |
978db57e | 774 | static void __io_double_put_req(struct io_kiocb *req); |
94ae5e77 JA |
775 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req); |
776 | static void io_queue_linked_timeout(struct io_kiocb *req); | |
05f3fb3c JA |
777 | static int __io_sqe_files_update(struct io_ring_ctx *ctx, |
778 | struct io_uring_files_update *ip, | |
779 | unsigned nr_args); | |
f86cd20c | 780 | static int io_grab_files(struct io_kiocb *req); |
2faf852d | 781 | static void io_ring_file_ref_flush(struct fixed_file_data *data); |
99bc4c38 | 782 | static void io_cleanup_req(struct io_kiocb *req); |
b41e9852 JA |
783 | static int io_file_get(struct io_submit_state *state, struct io_kiocb *req, |
784 | int fd, struct file **out_file, bool fixed); | |
785 | static void __io_queue_sqe(struct io_kiocb *req, | |
786 | const struct io_uring_sqe *sqe); | |
de0617e4 | 787 | |
2b188cc1 JA |
788 | static struct kmem_cache *req_cachep; |
789 | ||
790 | static const struct file_operations io_uring_fops; | |
791 | ||
792 | struct sock *io_uring_get_socket(struct file *file) | |
793 | { | |
794 | #if defined(CONFIG_UNIX) | |
795 | if (file->f_op == &io_uring_fops) { | |
796 | struct io_ring_ctx *ctx = file->private_data; | |
797 | ||
798 | return ctx->ring_sock->sk; | |
799 | } | |
800 | #endif | |
801 | return NULL; | |
802 | } | |
803 | EXPORT_SYMBOL(io_uring_get_socket); | |
804 | ||
805 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) | |
806 | { | |
807 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
808 | ||
206aefde | 809 | complete(&ctx->completions[0]); |
2b188cc1 JA |
810 | } |
811 | ||
812 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) | |
813 | { | |
814 | struct io_ring_ctx *ctx; | |
78076bb6 | 815 | int hash_bits; |
2b188cc1 JA |
816 | |
817 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
818 | if (!ctx) | |
819 | return NULL; | |
820 | ||
0ddf92e8 JA |
821 | ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL); |
822 | if (!ctx->fallback_req) | |
823 | goto err; | |
824 | ||
206aefde JA |
825 | ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL); |
826 | if (!ctx->completions) | |
827 | goto err; | |
828 | ||
78076bb6 JA |
829 | /* |
830 | * Use 5 bits less than the max cq entries, that should give us around | |
831 | * 32 entries per hash list if totally full and uniformly spread. | |
832 | */ | |
833 | hash_bits = ilog2(p->cq_entries); | |
834 | hash_bits -= 5; | |
835 | if (hash_bits <= 0) | |
836 | hash_bits = 1; | |
837 | ctx->cancel_hash_bits = hash_bits; | |
838 | ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head), | |
839 | GFP_KERNEL); | |
840 | if (!ctx->cancel_hash) | |
841 | goto err; | |
842 | __hash_init(ctx->cancel_hash, 1U << hash_bits); | |
843 | ||
21482896 | 844 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
206aefde JA |
845 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) |
846 | goto err; | |
2b188cc1 JA |
847 | |
848 | ctx->flags = p->flags; | |
849 | init_waitqueue_head(&ctx->cq_wait); | |
1d7bb1d5 | 850 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
206aefde JA |
851 | init_completion(&ctx->completions[0]); |
852 | init_completion(&ctx->completions[1]); | |
071698e1 | 853 | idr_init(&ctx->personality_idr); |
2b188cc1 JA |
854 | mutex_init(&ctx->uring_lock); |
855 | init_waitqueue_head(&ctx->wait); | |
856 | spin_lock_init(&ctx->completion_lock); | |
def596e9 | 857 | INIT_LIST_HEAD(&ctx->poll_list); |
de0617e4 | 858 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 859 | INIT_LIST_HEAD(&ctx->timeout_list); |
fcb323cc JA |
860 | init_waitqueue_head(&ctx->inflight_wait); |
861 | spin_lock_init(&ctx->inflight_lock); | |
862 | INIT_LIST_HEAD(&ctx->inflight_list); | |
2b188cc1 | 863 | return ctx; |
206aefde | 864 | err: |
0ddf92e8 JA |
865 | if (ctx->fallback_req) |
866 | kmem_cache_free(req_cachep, ctx->fallback_req); | |
206aefde | 867 | kfree(ctx->completions); |
78076bb6 | 868 | kfree(ctx->cancel_hash); |
206aefde JA |
869 | kfree(ctx); |
870 | return NULL; | |
2b188cc1 JA |
871 | } |
872 | ||
9d858b21 | 873 | static inline bool __req_need_defer(struct io_kiocb *req) |
7adf4eaf | 874 | { |
a197f664 JL |
875 | struct io_ring_ctx *ctx = req->ctx; |
876 | ||
498ccd9e JA |
877 | return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped |
878 | + atomic_read(&ctx->cached_cq_overflow); | |
7adf4eaf JA |
879 | } |
880 | ||
9d858b21 | 881 | static inline bool req_need_defer(struct io_kiocb *req) |
de0617e4 | 882 | { |
87987898 | 883 | if (unlikely(req->flags & REQ_F_IO_DRAIN)) |
9d858b21 | 884 | return __req_need_defer(req); |
de0617e4 | 885 | |
9d858b21 | 886 | return false; |
de0617e4 JA |
887 | } |
888 | ||
7adf4eaf | 889 | static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx) |
de0617e4 JA |
890 | { |
891 | struct io_kiocb *req; | |
892 | ||
7adf4eaf | 893 | req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list); |
9d858b21 | 894 | if (req && !req_need_defer(req)) { |
de0617e4 JA |
895 | list_del_init(&req->list); |
896 | return req; | |
897 | } | |
898 | ||
899 | return NULL; | |
900 | } | |
901 | ||
5262f567 JA |
902 | static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx) |
903 | { | |
7adf4eaf JA |
904 | struct io_kiocb *req; |
905 | ||
906 | req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list); | |
93bd25bb JA |
907 | if (req) { |
908 | if (req->flags & REQ_F_TIMEOUT_NOSEQ) | |
909 | return NULL; | |
fb4b3d3f | 910 | if (!__req_need_defer(req)) { |
93bd25bb JA |
911 | list_del_init(&req->list); |
912 | return req; | |
913 | } | |
7adf4eaf JA |
914 | } |
915 | ||
916 | return NULL; | |
5262f567 JA |
917 | } |
918 | ||
de0617e4 | 919 | static void __io_commit_cqring(struct io_ring_ctx *ctx) |
2b188cc1 | 920 | { |
75b28aff | 921 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 922 | |
07910158 PB |
923 | /* order cqe stores with ring update */ |
924 | smp_store_release(&rings->cq.tail, ctx->cached_cq_tail); | |
2b188cc1 | 925 | |
07910158 PB |
926 | if (wq_has_sleeper(&ctx->cq_wait)) { |
927 | wake_up_interruptible(&ctx->cq_wait); | |
928 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
2b188cc1 JA |
929 | } |
930 | } | |
931 | ||
cccf0ee8 JA |
932 | static inline void io_req_work_grab_env(struct io_kiocb *req, |
933 | const struct io_op_def *def) | |
934 | { | |
935 | if (!req->work.mm && def->needs_mm) { | |
936 | mmgrab(current->mm); | |
937 | req->work.mm = current->mm; | |
2b188cc1 | 938 | } |
cccf0ee8 JA |
939 | if (!req->work.creds) |
940 | req->work.creds = get_current_cred(); | |
ff002b30 JA |
941 | if (!req->work.fs && def->needs_fs) { |
942 | spin_lock(¤t->fs->lock); | |
943 | if (!current->fs->in_exec) { | |
944 | req->work.fs = current->fs; | |
945 | req->work.fs->users++; | |
946 | } else { | |
947 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
948 | } | |
949 | spin_unlock(¤t->fs->lock); | |
950 | } | |
6ab23144 JA |
951 | if (!req->work.task_pid) |
952 | req->work.task_pid = task_pid_vnr(current); | |
2b188cc1 JA |
953 | } |
954 | ||
cccf0ee8 | 955 | static inline void io_req_work_drop_env(struct io_kiocb *req) |
18d9be1a | 956 | { |
cccf0ee8 JA |
957 | if (req->work.mm) { |
958 | mmdrop(req->work.mm); | |
959 | req->work.mm = NULL; | |
960 | } | |
961 | if (req->work.creds) { | |
962 | put_cred(req->work.creds); | |
963 | req->work.creds = NULL; | |
964 | } | |
ff002b30 JA |
965 | if (req->work.fs) { |
966 | struct fs_struct *fs = req->work.fs; | |
967 | ||
968 | spin_lock(&req->work.fs->lock); | |
969 | if (--fs->users) | |
970 | fs = NULL; | |
971 | spin_unlock(&req->work.fs->lock); | |
972 | if (fs) | |
973 | free_fs_struct(fs); | |
974 | } | |
561fb04a JA |
975 | } |
976 | ||
deb6dc05 PB |
977 | static inline void io_prep_next_work(struct io_kiocb *req, |
978 | struct io_kiocb **link) | |
979 | { | |
980 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
981 | ||
982 | if (!(req->flags & REQ_F_ISREG) && def->unbound_nonreg_file) | |
983 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
984 | ||
985 | *link = io_prep_linked_timeout(req); | |
986 | } | |
987 | ||
94ae5e77 JA |
988 | static inline bool io_prep_async_work(struct io_kiocb *req, |
989 | struct io_kiocb **link) | |
18d9be1a | 990 | { |
d3656344 | 991 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
561fb04a | 992 | bool do_hashed = false; |
54a91f3b | 993 | |
d3656344 JA |
994 | if (req->flags & REQ_F_ISREG) { |
995 | if (def->hash_reg_file) | |
3529d8c2 | 996 | do_hashed = true; |
d3656344 JA |
997 | } else { |
998 | if (def->unbound_nonreg_file) | |
3529d8c2 | 999 | req->work.flags |= IO_WQ_WORK_UNBOUND; |
54a91f3b | 1000 | } |
cccf0ee8 JA |
1001 | |
1002 | io_req_work_grab_env(req, def); | |
54a91f3b | 1003 | |
94ae5e77 | 1004 | *link = io_prep_linked_timeout(req); |
561fb04a JA |
1005 | return do_hashed; |
1006 | } | |
1007 | ||
a197f664 | 1008 | static inline void io_queue_async_work(struct io_kiocb *req) |
561fb04a | 1009 | { |
a197f664 | 1010 | struct io_ring_ctx *ctx = req->ctx; |
94ae5e77 JA |
1011 | struct io_kiocb *link; |
1012 | bool do_hashed; | |
1013 | ||
1014 | do_hashed = io_prep_async_work(req, &link); | |
561fb04a JA |
1015 | |
1016 | trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work, | |
1017 | req->flags); | |
1018 | if (!do_hashed) { | |
1019 | io_wq_enqueue(ctx->io_wq, &req->work); | |
1020 | } else { | |
1021 | io_wq_enqueue_hashed(ctx->io_wq, &req->work, | |
1022 | file_inode(req->file)); | |
1023 | } | |
94ae5e77 JA |
1024 | |
1025 | if (link) | |
1026 | io_queue_linked_timeout(link); | |
18d9be1a JA |
1027 | } |
1028 | ||
5262f567 JA |
1029 | static void io_kill_timeout(struct io_kiocb *req) |
1030 | { | |
1031 | int ret; | |
1032 | ||
2d28390a | 1033 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
5262f567 JA |
1034 | if (ret != -1) { |
1035 | atomic_inc(&req->ctx->cq_timeouts); | |
842f9612 | 1036 | list_del_init(&req->list); |
78e19bbe | 1037 | io_cqring_fill_event(req, 0); |
ec9c02ad | 1038 | io_put_req(req); |
5262f567 JA |
1039 | } |
1040 | } | |
1041 | ||
1042 | static void io_kill_timeouts(struct io_ring_ctx *ctx) | |
1043 | { | |
1044 | struct io_kiocb *req, *tmp; | |
1045 | ||
1046 | spin_lock_irq(&ctx->completion_lock); | |
1047 | list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list) | |
1048 | io_kill_timeout(req); | |
1049 | spin_unlock_irq(&ctx->completion_lock); | |
1050 | } | |
1051 | ||
de0617e4 JA |
1052 | static void io_commit_cqring(struct io_ring_ctx *ctx) |
1053 | { | |
1054 | struct io_kiocb *req; | |
1055 | ||
5262f567 JA |
1056 | while ((req = io_get_timeout_req(ctx)) != NULL) |
1057 | io_kill_timeout(req); | |
1058 | ||
de0617e4 JA |
1059 | __io_commit_cqring(ctx); |
1060 | ||
87987898 | 1061 | while ((req = io_get_deferred_req(ctx)) != NULL) |
a197f664 | 1062 | io_queue_async_work(req); |
de0617e4 JA |
1063 | } |
1064 | ||
2b188cc1 JA |
1065 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) |
1066 | { | |
75b28aff | 1067 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
1068 | unsigned tail; |
1069 | ||
1070 | tail = ctx->cached_cq_tail; | |
115e12e5 SB |
1071 | /* |
1072 | * writes to the cq entry need to come after reading head; the | |
1073 | * control dependency is enough as we're using WRITE_ONCE to | |
1074 | * fill the cq entry | |
1075 | */ | |
75b28aff | 1076 | if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries) |
2b188cc1 JA |
1077 | return NULL; |
1078 | ||
1079 | ctx->cached_cq_tail++; | |
75b28aff | 1080 | return &rings->cqes[tail & ctx->cq_mask]; |
2b188cc1 JA |
1081 | } |
1082 | ||
f2842ab5 JA |
1083 | static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx) |
1084 | { | |
f0b493e6 JA |
1085 | if (!ctx->cq_ev_fd) |
1086 | return false; | |
f2842ab5 JA |
1087 | if (!ctx->eventfd_async) |
1088 | return true; | |
b41e9852 | 1089 | return io_wq_current_is_worker(); |
f2842ab5 JA |
1090 | } |
1091 | ||
b41e9852 | 1092 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
1d7bb1d5 JA |
1093 | { |
1094 | if (waitqueue_active(&ctx->wait)) | |
1095 | wake_up(&ctx->wait); | |
1096 | if (waitqueue_active(&ctx->sqo_wait)) | |
1097 | wake_up(&ctx->sqo_wait); | |
b41e9852 | 1098 | if (io_should_trigger_evfd(ctx)) |
1d7bb1d5 JA |
1099 | eventfd_signal(ctx->cq_ev_fd, 1); |
1100 | } | |
1101 | ||
c4a2ed72 JA |
1102 | /* Returns true if there are no backlogged entries after the flush */ |
1103 | static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force) | |
1d7bb1d5 JA |
1104 | { |
1105 | struct io_rings *rings = ctx->rings; | |
1106 | struct io_uring_cqe *cqe; | |
1107 | struct io_kiocb *req; | |
1108 | unsigned long flags; | |
1109 | LIST_HEAD(list); | |
1110 | ||
1111 | if (!force) { | |
1112 | if (list_empty_careful(&ctx->cq_overflow_list)) | |
c4a2ed72 | 1113 | return true; |
1d7bb1d5 JA |
1114 | if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) == |
1115 | rings->cq_ring_entries)) | |
c4a2ed72 | 1116 | return false; |
1d7bb1d5 JA |
1117 | } |
1118 | ||
1119 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
1120 | ||
1121 | /* if force is set, the ring is going away. always drop after that */ | |
1122 | if (force) | |
69b3e546 | 1123 | ctx->cq_overflow_flushed = 1; |
1d7bb1d5 | 1124 | |
c4a2ed72 | 1125 | cqe = NULL; |
1d7bb1d5 JA |
1126 | while (!list_empty(&ctx->cq_overflow_list)) { |
1127 | cqe = io_get_cqring(ctx); | |
1128 | if (!cqe && !force) | |
1129 | break; | |
1130 | ||
1131 | req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb, | |
1132 | list); | |
1133 | list_move(&req->list, &list); | |
2ca10259 | 1134 | req->flags &= ~REQ_F_OVERFLOW; |
1d7bb1d5 JA |
1135 | if (cqe) { |
1136 | WRITE_ONCE(cqe->user_data, req->user_data); | |
1137 | WRITE_ONCE(cqe->res, req->result); | |
1138 | WRITE_ONCE(cqe->flags, 0); | |
1139 | } else { | |
1140 | WRITE_ONCE(ctx->rings->cq_overflow, | |
1141 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
1142 | } | |
1143 | } | |
1144 | ||
1145 | io_commit_cqring(ctx); | |
ad3eb2c8 JA |
1146 | if (cqe) { |
1147 | clear_bit(0, &ctx->sq_check_overflow); | |
1148 | clear_bit(0, &ctx->cq_check_overflow); | |
1149 | } | |
1d7bb1d5 JA |
1150 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
1151 | io_cqring_ev_posted(ctx); | |
1152 | ||
1153 | while (!list_empty(&list)) { | |
1154 | req = list_first_entry(&list, struct io_kiocb, list); | |
1155 | list_del(&req->list); | |
ec9c02ad | 1156 | io_put_req(req); |
1d7bb1d5 | 1157 | } |
c4a2ed72 JA |
1158 | |
1159 | return cqe != NULL; | |
1d7bb1d5 JA |
1160 | } |
1161 | ||
78e19bbe | 1162 | static void io_cqring_fill_event(struct io_kiocb *req, long res) |
2b188cc1 | 1163 | { |
78e19bbe | 1164 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
1165 | struct io_uring_cqe *cqe; |
1166 | ||
78e19bbe | 1167 | trace_io_uring_complete(ctx, req->user_data, res); |
51c3ff62 | 1168 | |
2b188cc1 JA |
1169 | /* |
1170 | * If we can't get a cq entry, userspace overflowed the | |
1171 | * submission (by quite a lot). Increment the overflow count in | |
1172 | * the ring. | |
1173 | */ | |
1174 | cqe = io_get_cqring(ctx); | |
1d7bb1d5 | 1175 | if (likely(cqe)) { |
78e19bbe | 1176 | WRITE_ONCE(cqe->user_data, req->user_data); |
2b188cc1 | 1177 | WRITE_ONCE(cqe->res, res); |
c71ffb67 | 1178 | WRITE_ONCE(cqe->flags, 0); |
1d7bb1d5 | 1179 | } else if (ctx->cq_overflow_flushed) { |
498ccd9e JA |
1180 | WRITE_ONCE(ctx->rings->cq_overflow, |
1181 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
1d7bb1d5 | 1182 | } else { |
ad3eb2c8 JA |
1183 | if (list_empty(&ctx->cq_overflow_list)) { |
1184 | set_bit(0, &ctx->sq_check_overflow); | |
1185 | set_bit(0, &ctx->cq_check_overflow); | |
1186 | } | |
2ca10259 | 1187 | req->flags |= REQ_F_OVERFLOW; |
1d7bb1d5 JA |
1188 | refcount_inc(&req->refs); |
1189 | req->result = res; | |
1190 | list_add_tail(&req->list, &ctx->cq_overflow_list); | |
2b188cc1 JA |
1191 | } |
1192 | } | |
1193 | ||
78e19bbe | 1194 | static void io_cqring_add_event(struct io_kiocb *req, long res) |
2b188cc1 | 1195 | { |
78e19bbe | 1196 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
1197 | unsigned long flags; |
1198 | ||
1199 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
78e19bbe | 1200 | io_cqring_fill_event(req, res); |
2b188cc1 JA |
1201 | io_commit_cqring(ctx); |
1202 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1203 | ||
8c838788 | 1204 | io_cqring_ev_posted(ctx); |
2b188cc1 JA |
1205 | } |
1206 | ||
0ddf92e8 JA |
1207 | static inline bool io_is_fallback_req(struct io_kiocb *req) |
1208 | { | |
1209 | return req == (struct io_kiocb *) | |
1210 | ((unsigned long) req->ctx->fallback_req & ~1UL); | |
1211 | } | |
1212 | ||
1213 | static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx) | |
1214 | { | |
1215 | struct io_kiocb *req; | |
1216 | ||
1217 | req = ctx->fallback_req; | |
1218 | if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req)) | |
1219 | return req; | |
1220 | ||
1221 | return NULL; | |
1222 | } | |
1223 | ||
2579f913 JA |
1224 | static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx, |
1225 | struct io_submit_state *state) | |
2b188cc1 | 1226 | { |
fd6fab2c | 1227 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
2b188cc1 JA |
1228 | struct io_kiocb *req; |
1229 | ||
2579f913 | 1230 | if (!state) { |
fd6fab2c | 1231 | req = kmem_cache_alloc(req_cachep, gfp); |
2579f913 | 1232 | if (unlikely(!req)) |
0ddf92e8 | 1233 | goto fallback; |
2579f913 JA |
1234 | } else if (!state->free_reqs) { |
1235 | size_t sz; | |
1236 | int ret; | |
1237 | ||
1238 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
fd6fab2c JA |
1239 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs); |
1240 | ||
1241 | /* | |
1242 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
1243 | * retry single alloc to be on the safe side. | |
1244 | */ | |
1245 | if (unlikely(ret <= 0)) { | |
1246 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
1247 | if (!state->reqs[0]) | |
0ddf92e8 | 1248 | goto fallback; |
fd6fab2c JA |
1249 | ret = 1; |
1250 | } | |
2579f913 | 1251 | state->free_reqs = ret - 1; |
6c8a3134 | 1252 | req = state->reqs[ret - 1]; |
2579f913 | 1253 | } else { |
2579f913 | 1254 | state->free_reqs--; |
6c8a3134 | 1255 | req = state->reqs[state->free_reqs]; |
2b188cc1 JA |
1256 | } |
1257 | ||
0ddf92e8 | 1258 | got_it: |
1a6b74fc | 1259 | req->io = NULL; |
60c112b0 | 1260 | req->file = NULL; |
2579f913 JA |
1261 | req->ctx = ctx; |
1262 | req->flags = 0; | |
e65ef56d JA |
1263 | /* one is dropped after submission, the other at completion */ |
1264 | refcount_set(&req->refs, 2); | |
9e645e11 | 1265 | req->result = 0; |
561fb04a | 1266 | INIT_IO_WORK(&req->work, io_wq_submit_work); |
2579f913 | 1267 | return req; |
0ddf92e8 JA |
1268 | fallback: |
1269 | req = io_get_fallback_req(ctx); | |
1270 | if (req) | |
1271 | goto got_it; | |
6805b32e | 1272 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
1273 | return NULL; |
1274 | } | |
1275 | ||
8da11c19 PB |
1276 | static inline void io_put_file(struct io_kiocb *req, struct file *file, |
1277 | bool fixed) | |
1278 | { | |
1279 | if (fixed) | |
1280 | percpu_ref_put(&req->ctx->file_data->refs); | |
1281 | else | |
1282 | fput(file); | |
1283 | } | |
1284 | ||
2b85edfc | 1285 | static void __io_req_do_free(struct io_kiocb *req) |
def596e9 | 1286 | { |
2b85edfc PB |
1287 | if (likely(!io_is_fallback_req(req))) |
1288 | kmem_cache_free(req_cachep, req); | |
1289 | else | |
1290 | clear_bit_unlock(0, (unsigned long *) req->ctx->fallback_req); | |
1291 | } | |
1292 | ||
c6ca97b3 | 1293 | static void __io_req_aux_free(struct io_kiocb *req) |
2b188cc1 | 1294 | { |
929a3af9 PB |
1295 | if (req->flags & REQ_F_NEED_CLEANUP) |
1296 | io_cleanup_req(req); | |
1297 | ||
96fd84d8 | 1298 | kfree(req->io); |
8da11c19 PB |
1299 | if (req->file) |
1300 | io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE)); | |
cccf0ee8 JA |
1301 | |
1302 | io_req_work_drop_env(req); | |
def596e9 JA |
1303 | } |
1304 | ||
9e645e11 | 1305 | static void __io_free_req(struct io_kiocb *req) |
2b188cc1 | 1306 | { |
c6ca97b3 | 1307 | __io_req_aux_free(req); |
fcb323cc | 1308 | |
fcb323cc | 1309 | if (req->flags & REQ_F_INFLIGHT) { |
c6ca97b3 | 1310 | struct io_ring_ctx *ctx = req->ctx; |
fcb323cc JA |
1311 | unsigned long flags; |
1312 | ||
1313 | spin_lock_irqsave(&ctx->inflight_lock, flags); | |
1314 | list_del(&req->inflight_entry); | |
1315 | if (waitqueue_active(&ctx->inflight_wait)) | |
1316 | wake_up(&ctx->inflight_wait); | |
1317 | spin_unlock_irqrestore(&ctx->inflight_lock, flags); | |
1318 | } | |
2b85edfc PB |
1319 | |
1320 | percpu_ref_put(&req->ctx->refs); | |
1321 | __io_req_do_free(req); | |
e65ef56d JA |
1322 | } |
1323 | ||
c6ca97b3 JA |
1324 | struct req_batch { |
1325 | void *reqs[IO_IOPOLL_BATCH]; | |
1326 | int to_free; | |
1327 | int need_iter; | |
1328 | }; | |
1329 | ||
1330 | static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb) | |
1331 | { | |
10fef4be JA |
1332 | int fixed_refs = rb->to_free; |
1333 | ||
c6ca97b3 JA |
1334 | if (!rb->to_free) |
1335 | return; | |
1336 | if (rb->need_iter) { | |
1337 | int i, inflight = 0; | |
1338 | unsigned long flags; | |
1339 | ||
10fef4be | 1340 | fixed_refs = 0; |
c6ca97b3 JA |
1341 | for (i = 0; i < rb->to_free; i++) { |
1342 | struct io_kiocb *req = rb->reqs[i]; | |
1343 | ||
10fef4be | 1344 | if (req->flags & REQ_F_FIXED_FILE) { |
c6ca97b3 | 1345 | req->file = NULL; |
10fef4be JA |
1346 | fixed_refs++; |
1347 | } | |
c6ca97b3 JA |
1348 | if (req->flags & REQ_F_INFLIGHT) |
1349 | inflight++; | |
c6ca97b3 JA |
1350 | __io_req_aux_free(req); |
1351 | } | |
1352 | if (!inflight) | |
1353 | goto do_free; | |
1354 | ||
1355 | spin_lock_irqsave(&ctx->inflight_lock, flags); | |
1356 | for (i = 0; i < rb->to_free; i++) { | |
1357 | struct io_kiocb *req = rb->reqs[i]; | |
1358 | ||
10fef4be | 1359 | if (req->flags & REQ_F_INFLIGHT) { |
c6ca97b3 JA |
1360 | list_del(&req->inflight_entry); |
1361 | if (!--inflight) | |
1362 | break; | |
1363 | } | |
1364 | } | |
1365 | spin_unlock_irqrestore(&ctx->inflight_lock, flags); | |
1366 | ||
1367 | if (waitqueue_active(&ctx->inflight_wait)) | |
1368 | wake_up(&ctx->inflight_wait); | |
1369 | } | |
1370 | do_free: | |
1371 | kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs); | |
10fef4be JA |
1372 | if (fixed_refs) |
1373 | percpu_ref_put_many(&ctx->file_data->refs, fixed_refs); | |
c6ca97b3 | 1374 | percpu_ref_put_many(&ctx->refs, rb->to_free); |
c6ca97b3 | 1375 | rb->to_free = rb->need_iter = 0; |
e65ef56d JA |
1376 | } |
1377 | ||
a197f664 | 1378 | static bool io_link_cancel_timeout(struct io_kiocb *req) |
2665abfd | 1379 | { |
a197f664 | 1380 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd JA |
1381 | int ret; |
1382 | ||
2d28390a | 1383 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
2665abfd | 1384 | if (ret != -1) { |
78e19bbe | 1385 | io_cqring_fill_event(req, -ECANCELED); |
2665abfd JA |
1386 | io_commit_cqring(ctx); |
1387 | req->flags &= ~REQ_F_LINK; | |
ec9c02ad | 1388 | io_put_req(req); |
2665abfd JA |
1389 | return true; |
1390 | } | |
1391 | ||
1392 | return false; | |
e65ef56d JA |
1393 | } |
1394 | ||
ba816ad6 | 1395 | static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr) |
9e645e11 | 1396 | { |
2665abfd | 1397 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd | 1398 | bool wake_ev = false; |
9e645e11 | 1399 | |
4d7dd462 JA |
1400 | /* Already got next link */ |
1401 | if (req->flags & REQ_F_LINK_NEXT) | |
1402 | return; | |
1403 | ||
9e645e11 JA |
1404 | /* |
1405 | * The list should never be empty when we are called here. But could | |
1406 | * potentially happen if the chain is messed up, check to be on the | |
1407 | * safe side. | |
1408 | */ | |
4493233e PB |
1409 | while (!list_empty(&req->link_list)) { |
1410 | struct io_kiocb *nxt = list_first_entry(&req->link_list, | |
1411 | struct io_kiocb, link_list); | |
94ae5e77 | 1412 | |
4493233e PB |
1413 | if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) && |
1414 | (nxt->flags & REQ_F_TIMEOUT))) { | |
1415 | list_del_init(&nxt->link_list); | |
94ae5e77 | 1416 | wake_ev |= io_link_cancel_timeout(nxt); |
94ae5e77 JA |
1417 | req->flags &= ~REQ_F_LINK_TIMEOUT; |
1418 | continue; | |
1419 | } | |
9e645e11 | 1420 | |
4493233e PB |
1421 | list_del_init(&req->link_list); |
1422 | if (!list_empty(&nxt->link_list)) | |
1423 | nxt->flags |= REQ_F_LINK; | |
b18fdf71 | 1424 | *nxtptr = nxt; |
94ae5e77 | 1425 | break; |
9e645e11 | 1426 | } |
2665abfd | 1427 | |
4d7dd462 | 1428 | req->flags |= REQ_F_LINK_NEXT; |
2665abfd JA |
1429 | if (wake_ev) |
1430 | io_cqring_ev_posted(ctx); | |
9e645e11 JA |
1431 | } |
1432 | ||
1433 | /* | |
1434 | * Called if REQ_F_LINK is set, and we fail the head request | |
1435 | */ | |
1436 | static void io_fail_links(struct io_kiocb *req) | |
1437 | { | |
2665abfd | 1438 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd JA |
1439 | unsigned long flags; |
1440 | ||
1441 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
9e645e11 JA |
1442 | |
1443 | while (!list_empty(&req->link_list)) { | |
4493233e PB |
1444 | struct io_kiocb *link = list_first_entry(&req->link_list, |
1445 | struct io_kiocb, link_list); | |
9e645e11 | 1446 | |
4493233e | 1447 | list_del_init(&link->link_list); |
c826bd7a | 1448 | trace_io_uring_fail_link(req, link); |
2665abfd JA |
1449 | |
1450 | if ((req->flags & REQ_F_LINK_TIMEOUT) && | |
d625c6ee | 1451 | link->opcode == IORING_OP_LINK_TIMEOUT) { |
a197f664 | 1452 | io_link_cancel_timeout(link); |
2665abfd | 1453 | } else { |
78e19bbe | 1454 | io_cqring_fill_event(link, -ECANCELED); |
978db57e | 1455 | __io_double_put_req(link); |
2665abfd | 1456 | } |
5d960724 | 1457 | req->flags &= ~REQ_F_LINK_TIMEOUT; |
9e645e11 | 1458 | } |
2665abfd JA |
1459 | |
1460 | io_commit_cqring(ctx); | |
1461 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1462 | io_cqring_ev_posted(ctx); | |
9e645e11 JA |
1463 | } |
1464 | ||
4d7dd462 | 1465 | static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt) |
9e645e11 | 1466 | { |
4d7dd462 | 1467 | if (likely(!(req->flags & REQ_F_LINK))) |
2665abfd | 1468 | return; |
2665abfd | 1469 | |
9e645e11 JA |
1470 | /* |
1471 | * If LINK is set, we have dependent requests in this chain. If we | |
1472 | * didn't fail this request, queue the first one up, moving any other | |
1473 | * dependencies to the next request. In case of failure, fail the rest | |
1474 | * of the chain. | |
1475 | */ | |
2665abfd JA |
1476 | if (req->flags & REQ_F_FAIL_LINK) { |
1477 | io_fail_links(req); | |
7c9e7f0f JA |
1478 | } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) == |
1479 | REQ_F_LINK_TIMEOUT) { | |
2665abfd JA |
1480 | struct io_ring_ctx *ctx = req->ctx; |
1481 | unsigned long flags; | |
1482 | ||
1483 | /* | |
1484 | * If this is a timeout link, we could be racing with the | |
1485 | * timeout timer. Grab the completion lock for this case to | |
7c9e7f0f | 1486 | * protect against that. |
2665abfd JA |
1487 | */ |
1488 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
1489 | io_req_link_next(req, nxt); | |
1490 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1491 | } else { | |
1492 | io_req_link_next(req, nxt); | |
9e645e11 | 1493 | } |
4d7dd462 | 1494 | } |
9e645e11 | 1495 | |
c69f8dbe JL |
1496 | static void io_free_req(struct io_kiocb *req) |
1497 | { | |
944e58bf PB |
1498 | struct io_kiocb *nxt = NULL; |
1499 | ||
1500 | io_req_find_next(req, &nxt); | |
70cf9f32 | 1501 | __io_free_req(req); |
944e58bf PB |
1502 | |
1503 | if (nxt) | |
1504 | io_queue_async_work(nxt); | |
c69f8dbe JL |
1505 | } |
1506 | ||
ba816ad6 JA |
1507 | /* |
1508 | * Drop reference to request, return next in chain (if there is one) if this | |
1509 | * was the last reference to this request. | |
1510 | */ | |
f9bd67f6 | 1511 | __attribute__((nonnull)) |
ec9c02ad | 1512 | static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr) |
e65ef56d | 1513 | { |
2a44f467 JA |
1514 | if (refcount_dec_and_test(&req->refs)) { |
1515 | io_req_find_next(req, nxtptr); | |
4d7dd462 | 1516 | __io_free_req(req); |
2a44f467 | 1517 | } |
2b188cc1 JA |
1518 | } |
1519 | ||
e65ef56d JA |
1520 | static void io_put_req(struct io_kiocb *req) |
1521 | { | |
1522 | if (refcount_dec_and_test(&req->refs)) | |
1523 | io_free_req(req); | |
2b188cc1 JA |
1524 | } |
1525 | ||
978db57e JA |
1526 | /* |
1527 | * Must only be used if we don't need to care about links, usually from | |
1528 | * within the completion handling itself. | |
1529 | */ | |
1530 | static void __io_double_put_req(struct io_kiocb *req) | |
78e19bbe JA |
1531 | { |
1532 | /* drop both submit and complete references */ | |
1533 | if (refcount_sub_and_test(2, &req->refs)) | |
1534 | __io_free_req(req); | |
1535 | } | |
1536 | ||
978db57e JA |
1537 | static void io_double_put_req(struct io_kiocb *req) |
1538 | { | |
1539 | /* drop both submit and complete references */ | |
1540 | if (refcount_sub_and_test(2, &req->refs)) | |
1541 | io_free_req(req); | |
1542 | } | |
1543 | ||
1d7bb1d5 | 1544 | static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush) |
a3a0e43f | 1545 | { |
84f97dc2 JA |
1546 | struct io_rings *rings = ctx->rings; |
1547 | ||
ad3eb2c8 JA |
1548 | if (test_bit(0, &ctx->cq_check_overflow)) { |
1549 | /* | |
1550 | * noflush == true is from the waitqueue handler, just ensure | |
1551 | * we wake up the task, and the next invocation will flush the | |
1552 | * entries. We cannot safely to it from here. | |
1553 | */ | |
1554 | if (noflush && !list_empty(&ctx->cq_overflow_list)) | |
1555 | return -1U; | |
1d7bb1d5 | 1556 | |
ad3eb2c8 JA |
1557 | io_cqring_overflow_flush(ctx, false); |
1558 | } | |
1d7bb1d5 | 1559 | |
a3a0e43f JA |
1560 | /* See comment at the top of this file */ |
1561 | smp_rmb(); | |
ad3eb2c8 | 1562 | return ctx->cached_cq_tail - READ_ONCE(rings->cq.head); |
a3a0e43f JA |
1563 | } |
1564 | ||
fb5ccc98 PB |
1565 | static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) |
1566 | { | |
1567 | struct io_rings *rings = ctx->rings; | |
1568 | ||
1569 | /* make sure SQ entry isn't read before tail */ | |
1570 | return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; | |
1571 | } | |
1572 | ||
8237e045 | 1573 | static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req) |
e94f141b | 1574 | { |
c6ca97b3 JA |
1575 | if ((req->flags & REQ_F_LINK) || io_is_fallback_req(req)) |
1576 | return false; | |
e94f141b | 1577 | |
c6ca97b3 JA |
1578 | if (!(req->flags & REQ_F_FIXED_FILE) || req->io) |
1579 | rb->need_iter++; | |
1580 | ||
1581 | rb->reqs[rb->to_free++] = req; | |
1582 | if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs))) | |
1583 | io_free_req_many(req->ctx, rb); | |
1584 | return true; | |
e94f141b JA |
1585 | } |
1586 | ||
def596e9 JA |
1587 | /* |
1588 | * Find and free completed poll iocbs | |
1589 | */ | |
1590 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1591 | struct list_head *done) | |
1592 | { | |
8237e045 | 1593 | struct req_batch rb; |
def596e9 | 1594 | struct io_kiocb *req; |
def596e9 | 1595 | |
c6ca97b3 | 1596 | rb.to_free = rb.need_iter = 0; |
def596e9 JA |
1597 | while (!list_empty(done)) { |
1598 | req = list_first_entry(done, struct io_kiocb, list); | |
1599 | list_del(&req->list); | |
1600 | ||
78e19bbe | 1601 | io_cqring_fill_event(req, req->result); |
def596e9 JA |
1602 | (*nr_events)++; |
1603 | ||
8237e045 JA |
1604 | if (refcount_dec_and_test(&req->refs) && |
1605 | !io_req_multi_free(&rb, req)) | |
1606 | io_free_req(req); | |
def596e9 | 1607 | } |
def596e9 | 1608 | |
09bb8394 | 1609 | io_commit_cqring(ctx); |
8237e045 | 1610 | io_free_req_many(ctx, &rb); |
def596e9 JA |
1611 | } |
1612 | ||
1613 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1614 | long min) | |
1615 | { | |
1616 | struct io_kiocb *req, *tmp; | |
1617 | LIST_HEAD(done); | |
1618 | bool spin; | |
1619 | int ret; | |
1620 | ||
1621 | /* | |
1622 | * Only spin for completions if we don't have multiple devices hanging | |
1623 | * off our complete list, and we're under the requested amount. | |
1624 | */ | |
1625 | spin = !ctx->poll_multi_file && *nr_events < min; | |
1626 | ||
1627 | ret = 0; | |
1628 | list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) { | |
9adbd45d | 1629 | struct kiocb *kiocb = &req->rw.kiocb; |
def596e9 JA |
1630 | |
1631 | /* | |
1632 | * Move completed entries to our local list. If we find a | |
1633 | * request that requires polling, break out and complete | |
1634 | * the done list first, if we have entries there. | |
1635 | */ | |
1636 | if (req->flags & REQ_F_IOPOLL_COMPLETED) { | |
1637 | list_move_tail(&req->list, &done); | |
1638 | continue; | |
1639 | } | |
1640 | if (!list_empty(&done)) | |
1641 | break; | |
1642 | ||
1643 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
1644 | if (ret < 0) | |
1645 | break; | |
1646 | ||
1647 | if (ret && spin) | |
1648 | spin = false; | |
1649 | ret = 0; | |
1650 | } | |
1651 | ||
1652 | if (!list_empty(&done)) | |
1653 | io_iopoll_complete(ctx, nr_events, &done); | |
1654 | ||
1655 | return ret; | |
1656 | } | |
1657 | ||
1658 | /* | |
d195a66e | 1659 | * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a |
def596e9 JA |
1660 | * non-spinning poll check - we'll still enter the driver poll loop, but only |
1661 | * as a non-spinning completion check. | |
1662 | */ | |
1663 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1664 | long min) | |
1665 | { | |
08f5439f | 1666 | while (!list_empty(&ctx->poll_list) && !need_resched()) { |
def596e9 JA |
1667 | int ret; |
1668 | ||
1669 | ret = io_do_iopoll(ctx, nr_events, min); | |
1670 | if (ret < 0) | |
1671 | return ret; | |
1672 | if (!min || *nr_events >= min) | |
1673 | return 0; | |
1674 | } | |
1675 | ||
1676 | return 1; | |
1677 | } | |
1678 | ||
1679 | /* | |
1680 | * We can't just wait for polled events to come to us, we have to actively | |
1681 | * find and complete them. | |
1682 | */ | |
1683 | static void io_iopoll_reap_events(struct io_ring_ctx *ctx) | |
1684 | { | |
1685 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1686 | return; | |
1687 | ||
1688 | mutex_lock(&ctx->uring_lock); | |
1689 | while (!list_empty(&ctx->poll_list)) { | |
1690 | unsigned int nr_events = 0; | |
1691 | ||
1692 | io_iopoll_getevents(ctx, &nr_events, 1); | |
08f5439f JA |
1693 | |
1694 | /* | |
1695 | * Ensure we allow local-to-the-cpu processing to take place, | |
1696 | * in this case we need to ensure that we reap all events. | |
1697 | */ | |
1698 | cond_resched(); | |
def596e9 JA |
1699 | } |
1700 | mutex_unlock(&ctx->uring_lock); | |
1701 | } | |
1702 | ||
c7849be9 XW |
1703 | static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, |
1704 | long min) | |
def596e9 | 1705 | { |
2b2ed975 | 1706 | int iters = 0, ret = 0; |
500f9fba | 1707 | |
c7849be9 XW |
1708 | /* |
1709 | * We disallow the app entering submit/complete with polling, but we | |
1710 | * still need to lock the ring to prevent racing with polled issue | |
1711 | * that got punted to a workqueue. | |
1712 | */ | |
1713 | mutex_lock(&ctx->uring_lock); | |
def596e9 JA |
1714 | do { |
1715 | int tmin = 0; | |
1716 | ||
a3a0e43f JA |
1717 | /* |
1718 | * Don't enter poll loop if we already have events pending. | |
1719 | * If we do, we can potentially be spinning for commands that | |
1720 | * already triggered a CQE (eg in error). | |
1721 | */ | |
1d7bb1d5 | 1722 | if (io_cqring_events(ctx, false)) |
a3a0e43f JA |
1723 | break; |
1724 | ||
500f9fba JA |
1725 | /* |
1726 | * If a submit got punted to a workqueue, we can have the | |
1727 | * application entering polling for a command before it gets | |
1728 | * issued. That app will hold the uring_lock for the duration | |
1729 | * of the poll right here, so we need to take a breather every | |
1730 | * now and then to ensure that the issue has a chance to add | |
1731 | * the poll to the issued list. Otherwise we can spin here | |
1732 | * forever, while the workqueue is stuck trying to acquire the | |
1733 | * very same mutex. | |
1734 | */ | |
1735 | if (!(++iters & 7)) { | |
1736 | mutex_unlock(&ctx->uring_lock); | |
1737 | mutex_lock(&ctx->uring_lock); | |
1738 | } | |
1739 | ||
def596e9 JA |
1740 | if (*nr_events < min) |
1741 | tmin = min - *nr_events; | |
1742 | ||
1743 | ret = io_iopoll_getevents(ctx, nr_events, tmin); | |
1744 | if (ret <= 0) | |
1745 | break; | |
1746 | ret = 0; | |
1747 | } while (min && !*nr_events && !need_resched()); | |
1748 | ||
500f9fba | 1749 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
1750 | return ret; |
1751 | } | |
1752 | ||
491381ce | 1753 | static void kiocb_end_write(struct io_kiocb *req) |
2b188cc1 | 1754 | { |
491381ce JA |
1755 | /* |
1756 | * Tell lockdep we inherited freeze protection from submission | |
1757 | * thread. | |
1758 | */ | |
1759 | if (req->flags & REQ_F_ISREG) { | |
1760 | struct inode *inode = file_inode(req->file); | |
2b188cc1 | 1761 | |
491381ce | 1762 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); |
2b188cc1 | 1763 | } |
491381ce | 1764 | file_end_write(req->file); |
2b188cc1 JA |
1765 | } |
1766 | ||
4e88d6e7 JA |
1767 | static inline void req_set_fail_links(struct io_kiocb *req) |
1768 | { | |
1769 | if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK) | |
1770 | req->flags |= REQ_F_FAIL_LINK; | |
1771 | } | |
1772 | ||
ba816ad6 | 1773 | static void io_complete_rw_common(struct kiocb *kiocb, long res) |
2b188cc1 | 1774 | { |
9adbd45d | 1775 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
2b188cc1 | 1776 | |
491381ce JA |
1777 | if (kiocb->ki_flags & IOCB_WRITE) |
1778 | kiocb_end_write(req); | |
2b188cc1 | 1779 | |
4e88d6e7 JA |
1780 | if (res != req->result) |
1781 | req_set_fail_links(req); | |
78e19bbe | 1782 | io_cqring_add_event(req, res); |
ba816ad6 JA |
1783 | } |
1784 | ||
1785 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
1786 | { | |
9adbd45d | 1787 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
ba816ad6 JA |
1788 | |
1789 | io_complete_rw_common(kiocb, res); | |
e65ef56d | 1790 | io_put_req(req); |
2b188cc1 JA |
1791 | } |
1792 | ||
ba816ad6 JA |
1793 | static struct io_kiocb *__io_complete_rw(struct kiocb *kiocb, long res) |
1794 | { | |
9adbd45d | 1795 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
ec9c02ad | 1796 | struct io_kiocb *nxt = NULL; |
ba816ad6 JA |
1797 | |
1798 | io_complete_rw_common(kiocb, res); | |
ec9c02ad JL |
1799 | io_put_req_find_next(req, &nxt); |
1800 | ||
1801 | return nxt; | |
2b188cc1 JA |
1802 | } |
1803 | ||
def596e9 JA |
1804 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
1805 | { | |
9adbd45d | 1806 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
def596e9 | 1807 | |
491381ce JA |
1808 | if (kiocb->ki_flags & IOCB_WRITE) |
1809 | kiocb_end_write(req); | |
def596e9 | 1810 | |
4e88d6e7 JA |
1811 | if (res != req->result) |
1812 | req_set_fail_links(req); | |
9e645e11 | 1813 | req->result = res; |
def596e9 JA |
1814 | if (res != -EAGAIN) |
1815 | req->flags |= REQ_F_IOPOLL_COMPLETED; | |
1816 | } | |
1817 | ||
1818 | /* | |
1819 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1820 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
1821 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
1822 | * accessing the kiocb cookie. | |
1823 | */ | |
1824 | static void io_iopoll_req_issued(struct io_kiocb *req) | |
1825 | { | |
1826 | struct io_ring_ctx *ctx = req->ctx; | |
1827 | ||
1828 | /* | |
1829 | * Track whether we have multiple files in our lists. This will impact | |
1830 | * how we do polling eventually, not spinning if we're on potentially | |
1831 | * different devices. | |
1832 | */ | |
1833 | if (list_empty(&ctx->poll_list)) { | |
1834 | ctx->poll_multi_file = false; | |
1835 | } else if (!ctx->poll_multi_file) { | |
1836 | struct io_kiocb *list_req; | |
1837 | ||
1838 | list_req = list_first_entry(&ctx->poll_list, struct io_kiocb, | |
1839 | list); | |
9adbd45d | 1840 | if (list_req->file != req->file) |
def596e9 JA |
1841 | ctx->poll_multi_file = true; |
1842 | } | |
1843 | ||
1844 | /* | |
1845 | * For fast devices, IO may have already completed. If it has, add | |
1846 | * it to the front so we find it first. | |
1847 | */ | |
1848 | if (req->flags & REQ_F_IOPOLL_COMPLETED) | |
1849 | list_add(&req->list, &ctx->poll_list); | |
1850 | else | |
1851 | list_add_tail(&req->list, &ctx->poll_list); | |
bdcd3eab XW |
1852 | |
1853 | if ((ctx->flags & IORING_SETUP_SQPOLL) && | |
1854 | wq_has_sleeper(&ctx->sqo_wait)) | |
1855 | wake_up(&ctx->sqo_wait); | |
def596e9 JA |
1856 | } |
1857 | ||
3d6770fb | 1858 | static void io_file_put(struct io_submit_state *state) |
9a56a232 | 1859 | { |
3d6770fb | 1860 | if (state->file) { |
9a56a232 JA |
1861 | int diff = state->has_refs - state->used_refs; |
1862 | ||
1863 | if (diff) | |
1864 | fput_many(state->file, diff); | |
1865 | state->file = NULL; | |
1866 | } | |
1867 | } | |
1868 | ||
1869 | /* | |
1870 | * Get as many references to a file as we have IOs left in this submission, | |
1871 | * assuming most submissions are for one file, or at least that each file | |
1872 | * has more than one submission. | |
1873 | */ | |
8da11c19 | 1874 | static struct file *__io_file_get(struct io_submit_state *state, int fd) |
9a56a232 JA |
1875 | { |
1876 | if (!state) | |
1877 | return fget(fd); | |
1878 | ||
1879 | if (state->file) { | |
1880 | if (state->fd == fd) { | |
1881 | state->used_refs++; | |
1882 | state->ios_left--; | |
1883 | return state->file; | |
1884 | } | |
3d6770fb | 1885 | io_file_put(state); |
9a56a232 JA |
1886 | } |
1887 | state->file = fget_many(fd, state->ios_left); | |
1888 | if (!state->file) | |
1889 | return NULL; | |
1890 | ||
1891 | state->fd = fd; | |
1892 | state->has_refs = state->ios_left; | |
1893 | state->used_refs = 1; | |
1894 | state->ios_left--; | |
1895 | return state->file; | |
1896 | } | |
1897 | ||
2b188cc1 JA |
1898 | /* |
1899 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1900 | * any file. For now, just ensure that anything potentially problematic is done | |
1901 | * inline. | |
1902 | */ | |
1903 | static bool io_file_supports_async(struct file *file) | |
1904 | { | |
1905 | umode_t mode = file_inode(file)->i_mode; | |
1906 | ||
10d59345 | 1907 | if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode)) |
2b188cc1 JA |
1908 | return true; |
1909 | if (S_ISREG(mode) && file->f_op != &io_uring_fops) | |
1910 | return true; | |
1911 | ||
1912 | return false; | |
1913 | } | |
1914 | ||
3529d8c2 JA |
1915 | static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
1916 | bool force_nonblock) | |
2b188cc1 | 1917 | { |
def596e9 | 1918 | struct io_ring_ctx *ctx = req->ctx; |
9adbd45d | 1919 | struct kiocb *kiocb = &req->rw.kiocb; |
09bb8394 JA |
1920 | unsigned ioprio; |
1921 | int ret; | |
2b188cc1 | 1922 | |
491381ce JA |
1923 | if (S_ISREG(file_inode(req->file)->i_mode)) |
1924 | req->flags |= REQ_F_ISREG; | |
1925 | ||
2b188cc1 | 1926 | kiocb->ki_pos = READ_ONCE(sqe->off); |
ba04291e JA |
1927 | if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) { |
1928 | req->flags |= REQ_F_CUR_POS; | |
1929 | kiocb->ki_pos = req->file->f_pos; | |
1930 | } | |
2b188cc1 | 1931 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); |
3e577dcd PB |
1932 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); |
1933 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
1934 | if (unlikely(ret)) | |
1935 | return ret; | |
2b188cc1 JA |
1936 | |
1937 | ioprio = READ_ONCE(sqe->ioprio); | |
1938 | if (ioprio) { | |
1939 | ret = ioprio_check_cap(ioprio); | |
1940 | if (ret) | |
09bb8394 | 1941 | return ret; |
2b188cc1 JA |
1942 | |
1943 | kiocb->ki_ioprio = ioprio; | |
1944 | } else | |
1945 | kiocb->ki_ioprio = get_current_ioprio(); | |
1946 | ||
8449eeda | 1947 | /* don't allow async punt if RWF_NOWAIT was requested */ |
491381ce JA |
1948 | if ((kiocb->ki_flags & IOCB_NOWAIT) || |
1949 | (req->file->f_flags & O_NONBLOCK)) | |
8449eeda SB |
1950 | req->flags |= REQ_F_NOWAIT; |
1951 | ||
1952 | if (force_nonblock) | |
2b188cc1 | 1953 | kiocb->ki_flags |= IOCB_NOWAIT; |
8449eeda | 1954 | |
def596e9 | 1955 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
1956 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
1957 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 1958 | return -EOPNOTSUPP; |
2b188cc1 | 1959 | |
def596e9 JA |
1960 | kiocb->ki_flags |= IOCB_HIPRI; |
1961 | kiocb->ki_complete = io_complete_rw_iopoll; | |
6873e0bd | 1962 | req->result = 0; |
def596e9 | 1963 | } else { |
09bb8394 JA |
1964 | if (kiocb->ki_flags & IOCB_HIPRI) |
1965 | return -EINVAL; | |
def596e9 JA |
1966 | kiocb->ki_complete = io_complete_rw; |
1967 | } | |
9adbd45d | 1968 | |
3529d8c2 JA |
1969 | req->rw.addr = READ_ONCE(sqe->addr); |
1970 | req->rw.len = READ_ONCE(sqe->len); | |
9adbd45d JA |
1971 | /* we own ->private, reuse it for the buffer index */ |
1972 | req->rw.kiocb.private = (void *) (unsigned long) | |
3529d8c2 | 1973 | READ_ONCE(sqe->buf_index); |
2b188cc1 | 1974 | return 0; |
2b188cc1 JA |
1975 | } |
1976 | ||
1977 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
1978 | { | |
1979 | switch (ret) { | |
1980 | case -EIOCBQUEUED: | |
1981 | break; | |
1982 | case -ERESTARTSYS: | |
1983 | case -ERESTARTNOINTR: | |
1984 | case -ERESTARTNOHAND: | |
1985 | case -ERESTART_RESTARTBLOCK: | |
1986 | /* | |
1987 | * We can't just restart the syscall, since previously | |
1988 | * submitted sqes may already be in progress. Just fail this | |
1989 | * IO with EINTR. | |
1990 | */ | |
1991 | ret = -EINTR; | |
1992 | /* fall through */ | |
1993 | default: | |
1994 | kiocb->ki_complete(kiocb, ret, 0); | |
1995 | } | |
1996 | } | |
1997 | ||
bcaec089 | 1998 | static void kiocb_done(struct kiocb *kiocb, ssize_t ret, struct io_kiocb **nxt) |
ba816ad6 | 1999 | { |
ba04291e JA |
2000 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
2001 | ||
2002 | if (req->flags & REQ_F_CUR_POS) | |
2003 | req->file->f_pos = kiocb->ki_pos; | |
bcaec089 | 2004 | if (ret >= 0 && kiocb->ki_complete == io_complete_rw) |
ba816ad6 JA |
2005 | *nxt = __io_complete_rw(kiocb, ret); |
2006 | else | |
2007 | io_rw_done(kiocb, ret); | |
2008 | } | |
2009 | ||
9adbd45d | 2010 | static ssize_t io_import_fixed(struct io_kiocb *req, int rw, |
7d009165 | 2011 | struct iov_iter *iter) |
edafccee | 2012 | { |
9adbd45d JA |
2013 | struct io_ring_ctx *ctx = req->ctx; |
2014 | size_t len = req->rw.len; | |
edafccee JA |
2015 | struct io_mapped_ubuf *imu; |
2016 | unsigned index, buf_index; | |
2017 | size_t offset; | |
2018 | u64 buf_addr; | |
2019 | ||
2020 | /* attempt to use fixed buffers without having provided iovecs */ | |
2021 | if (unlikely(!ctx->user_bufs)) | |
2022 | return -EFAULT; | |
2023 | ||
9adbd45d | 2024 | buf_index = (unsigned long) req->rw.kiocb.private; |
edafccee JA |
2025 | if (unlikely(buf_index >= ctx->nr_user_bufs)) |
2026 | return -EFAULT; | |
2027 | ||
2028 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
2029 | imu = &ctx->user_bufs[index]; | |
9adbd45d | 2030 | buf_addr = req->rw.addr; |
edafccee JA |
2031 | |
2032 | /* overflow */ | |
2033 | if (buf_addr + len < buf_addr) | |
2034 | return -EFAULT; | |
2035 | /* not inside the mapped region */ | |
2036 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
2037 | return -EFAULT; | |
2038 | ||
2039 | /* | |
2040 | * May not be a start of buffer, set size appropriately | |
2041 | * and advance us to the beginning. | |
2042 | */ | |
2043 | offset = buf_addr - imu->ubuf; | |
2044 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
2045 | |
2046 | if (offset) { | |
2047 | /* | |
2048 | * Don't use iov_iter_advance() here, as it's really slow for | |
2049 | * using the latter parts of a big fixed buffer - it iterates | |
2050 | * over each segment manually. We can cheat a bit here, because | |
2051 | * we know that: | |
2052 | * | |
2053 | * 1) it's a BVEC iter, we set it up | |
2054 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
2055 | * first and last bvec | |
2056 | * | |
2057 | * So just find our index, and adjust the iterator afterwards. | |
2058 | * If the offset is within the first bvec (or the whole first | |
2059 | * bvec, just use iov_iter_advance(). This makes it easier | |
2060 | * since we can just skip the first segment, which may not | |
2061 | * be PAGE_SIZE aligned. | |
2062 | */ | |
2063 | const struct bio_vec *bvec = imu->bvec; | |
2064 | ||
2065 | if (offset <= bvec->bv_len) { | |
2066 | iov_iter_advance(iter, offset); | |
2067 | } else { | |
2068 | unsigned long seg_skip; | |
2069 | ||
2070 | /* skip first vec */ | |
2071 | offset -= bvec->bv_len; | |
2072 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
2073 | ||
2074 | iter->bvec = bvec + seg_skip; | |
2075 | iter->nr_segs -= seg_skip; | |
99c79f66 | 2076 | iter->count -= bvec->bv_len + offset; |
bd11b3a3 | 2077 | iter->iov_offset = offset & ~PAGE_MASK; |
bd11b3a3 JA |
2078 | } |
2079 | } | |
2080 | ||
5e559561 | 2081 | return len; |
edafccee JA |
2082 | } |
2083 | ||
cf6fd4bd PB |
2084 | static ssize_t io_import_iovec(int rw, struct io_kiocb *req, |
2085 | struct iovec **iovec, struct iov_iter *iter) | |
2b188cc1 | 2086 | { |
9adbd45d JA |
2087 | void __user *buf = u64_to_user_ptr(req->rw.addr); |
2088 | size_t sqe_len = req->rw.len; | |
edafccee JA |
2089 | u8 opcode; |
2090 | ||
d625c6ee | 2091 | opcode = req->opcode; |
7d009165 | 2092 | if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) { |
edafccee | 2093 | *iovec = NULL; |
9adbd45d | 2094 | return io_import_fixed(req, rw, iter); |
edafccee | 2095 | } |
2b188cc1 | 2096 | |
9adbd45d JA |
2097 | /* buffer index only valid with fixed read/write */ |
2098 | if (req->rw.kiocb.private) | |
2099 | return -EINVAL; | |
2100 | ||
3a6820f2 JA |
2101 | if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) { |
2102 | ssize_t ret; | |
2103 | ret = import_single_range(rw, buf, sqe_len, *iovec, iter); | |
2104 | *iovec = NULL; | |
3a901598 | 2105 | return ret < 0 ? ret : sqe_len; |
3a6820f2 JA |
2106 | } |
2107 | ||
f67676d1 JA |
2108 | if (req->io) { |
2109 | struct io_async_rw *iorw = &req->io->rw; | |
2110 | ||
2111 | *iovec = iorw->iov; | |
2112 | iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size); | |
2113 | if (iorw->iov == iorw->fast_iov) | |
2114 | *iovec = NULL; | |
2115 | return iorw->size; | |
2116 | } | |
2117 | ||
2b188cc1 | 2118 | #ifdef CONFIG_COMPAT |
cf6fd4bd | 2119 | if (req->ctx->compat) |
2b188cc1 JA |
2120 | return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV, |
2121 | iovec, iter); | |
2122 | #endif | |
2123 | ||
2124 | return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter); | |
2125 | } | |
2126 | ||
31b51510 | 2127 | /* |
32960613 JA |
2128 | * For files that don't have ->read_iter() and ->write_iter(), handle them |
2129 | * by looping over ->read() or ->write() manually. | |
31b51510 | 2130 | */ |
32960613 JA |
2131 | static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb, |
2132 | struct iov_iter *iter) | |
2133 | { | |
2134 | ssize_t ret = 0; | |
2135 | ||
2136 | /* | |
2137 | * Don't support polled IO through this interface, and we can't | |
2138 | * support non-blocking either. For the latter, this just causes | |
2139 | * the kiocb to be handled from an async context. | |
2140 | */ | |
2141 | if (kiocb->ki_flags & IOCB_HIPRI) | |
2142 | return -EOPNOTSUPP; | |
2143 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
2144 | return -EAGAIN; | |
2145 | ||
2146 | while (iov_iter_count(iter)) { | |
311ae9e1 | 2147 | struct iovec iovec; |
32960613 JA |
2148 | ssize_t nr; |
2149 | ||
311ae9e1 PB |
2150 | if (!iov_iter_is_bvec(iter)) { |
2151 | iovec = iov_iter_iovec(iter); | |
2152 | } else { | |
2153 | /* fixed buffers import bvec */ | |
2154 | iovec.iov_base = kmap(iter->bvec->bv_page) | |
2155 | + iter->iov_offset; | |
2156 | iovec.iov_len = min(iter->count, | |
2157 | iter->bvec->bv_len - iter->iov_offset); | |
2158 | } | |
2159 | ||
32960613 JA |
2160 | if (rw == READ) { |
2161 | nr = file->f_op->read(file, iovec.iov_base, | |
2162 | iovec.iov_len, &kiocb->ki_pos); | |
2163 | } else { | |
2164 | nr = file->f_op->write(file, iovec.iov_base, | |
2165 | iovec.iov_len, &kiocb->ki_pos); | |
2166 | } | |
2167 | ||
311ae9e1 PB |
2168 | if (iov_iter_is_bvec(iter)) |
2169 | kunmap(iter->bvec->bv_page); | |
2170 | ||
32960613 JA |
2171 | if (nr < 0) { |
2172 | if (!ret) | |
2173 | ret = nr; | |
2174 | break; | |
2175 | } | |
2176 | ret += nr; | |
2177 | if (nr != iovec.iov_len) | |
2178 | break; | |
2179 | iov_iter_advance(iter, nr); | |
2180 | } | |
2181 | ||
2182 | return ret; | |
2183 | } | |
2184 | ||
b7bb4f7d | 2185 | static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size, |
f67676d1 JA |
2186 | struct iovec *iovec, struct iovec *fast_iov, |
2187 | struct iov_iter *iter) | |
2188 | { | |
2189 | req->io->rw.nr_segs = iter->nr_segs; | |
2190 | req->io->rw.size = io_size; | |
2191 | req->io->rw.iov = iovec; | |
2192 | if (!req->io->rw.iov) { | |
2193 | req->io->rw.iov = req->io->rw.fast_iov; | |
2194 | memcpy(req->io->rw.iov, fast_iov, | |
2195 | sizeof(struct iovec) * iter->nr_segs); | |
99bc4c38 PB |
2196 | } else { |
2197 | req->flags |= REQ_F_NEED_CLEANUP; | |
f67676d1 JA |
2198 | } |
2199 | } | |
2200 | ||
b7bb4f7d | 2201 | static int io_alloc_async_ctx(struct io_kiocb *req) |
f67676d1 | 2202 | { |
d3656344 JA |
2203 | if (!io_op_defs[req->opcode].async_ctx) |
2204 | return 0; | |
f67676d1 | 2205 | req->io = kmalloc(sizeof(*req->io), GFP_KERNEL); |
06b76d44 | 2206 | return req->io == NULL; |
b7bb4f7d JA |
2207 | } |
2208 | ||
b7bb4f7d JA |
2209 | static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size, |
2210 | struct iovec *iovec, struct iovec *fast_iov, | |
2211 | struct iov_iter *iter) | |
2212 | { | |
980ad263 | 2213 | if (!io_op_defs[req->opcode].async_ctx) |
74566df3 | 2214 | return 0; |
5d204bcf JA |
2215 | if (!req->io) { |
2216 | if (io_alloc_async_ctx(req)) | |
2217 | return -ENOMEM; | |
b7bb4f7d | 2218 | |
5d204bcf JA |
2219 | io_req_map_rw(req, io_size, iovec, fast_iov, iter); |
2220 | } | |
b7bb4f7d | 2221 | return 0; |
f67676d1 JA |
2222 | } |
2223 | ||
3529d8c2 JA |
2224 | static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2225 | bool force_nonblock) | |
f67676d1 | 2226 | { |
3529d8c2 JA |
2227 | struct io_async_ctx *io; |
2228 | struct iov_iter iter; | |
f67676d1 JA |
2229 | ssize_t ret; |
2230 | ||
3529d8c2 JA |
2231 | ret = io_prep_rw(req, sqe, force_nonblock); |
2232 | if (ret) | |
2233 | return ret; | |
f67676d1 | 2234 | |
3529d8c2 JA |
2235 | if (unlikely(!(req->file->f_mode & FMODE_READ))) |
2236 | return -EBADF; | |
f67676d1 | 2237 | |
5f798bea PB |
2238 | /* either don't need iovec imported or already have it */ |
2239 | if (!req->io || req->flags & REQ_F_NEED_CLEANUP) | |
3529d8c2 JA |
2240 | return 0; |
2241 | ||
2242 | io = req->io; | |
2243 | io->rw.iov = io->rw.fast_iov; | |
2244 | req->io = NULL; | |
2245 | ret = io_import_iovec(READ, req, &io->rw.iov, &iter); | |
2246 | req->io = io; | |
2247 | if (ret < 0) | |
2248 | return ret; | |
2249 | ||
2250 | io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter); | |
2251 | return 0; | |
f67676d1 JA |
2252 | } |
2253 | ||
267bc904 | 2254 | static int io_read(struct io_kiocb *req, struct io_kiocb **nxt, |
8358e3a8 | 2255 | bool force_nonblock) |
2b188cc1 JA |
2256 | { |
2257 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
9adbd45d | 2258 | struct kiocb *kiocb = &req->rw.kiocb; |
2b188cc1 | 2259 | struct iov_iter iter; |
31b51510 | 2260 | size_t iov_count; |
f67676d1 | 2261 | ssize_t io_size, ret; |
2b188cc1 | 2262 | |
3529d8c2 | 2263 | ret = io_import_iovec(READ, req, &iovec, &iter); |
06b76d44 JA |
2264 | if (ret < 0) |
2265 | return ret; | |
2b188cc1 | 2266 | |
fd6c2e4c JA |
2267 | /* Ensure we clear previously set non-block flag */ |
2268 | if (!force_nonblock) | |
29de5f6a | 2269 | kiocb->ki_flags &= ~IOCB_NOWAIT; |
fd6c2e4c | 2270 | |
797f3f53 | 2271 | req->result = 0; |
f67676d1 | 2272 | io_size = ret; |
9e645e11 | 2273 | if (req->flags & REQ_F_LINK) |
f67676d1 JA |
2274 | req->result = io_size; |
2275 | ||
2276 | /* | |
2277 | * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so | |
2278 | * we know to async punt it even if it was opened O_NONBLOCK | |
2279 | */ | |
29de5f6a | 2280 | if (force_nonblock && !io_file_supports_async(req->file)) |
f67676d1 | 2281 | goto copy_iov; |
9e645e11 | 2282 | |
31b51510 | 2283 | iov_count = iov_iter_count(&iter); |
9adbd45d | 2284 | ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count); |
2b188cc1 JA |
2285 | if (!ret) { |
2286 | ssize_t ret2; | |
2287 | ||
9adbd45d JA |
2288 | if (req->file->f_op->read_iter) |
2289 | ret2 = call_read_iter(req->file, kiocb, &iter); | |
32960613 | 2290 | else |
9adbd45d | 2291 | ret2 = loop_rw_iter(READ, req->file, kiocb, &iter); |
32960613 | 2292 | |
9d93a3f5 | 2293 | /* Catch -EAGAIN return for forced non-blocking submission */ |
f67676d1 | 2294 | if (!force_nonblock || ret2 != -EAGAIN) { |
bcaec089 | 2295 | kiocb_done(kiocb, ret2, nxt); |
f67676d1 JA |
2296 | } else { |
2297 | copy_iov: | |
b7bb4f7d | 2298 | ret = io_setup_async_rw(req, io_size, iovec, |
f67676d1 JA |
2299 | inline_vecs, &iter); |
2300 | if (ret) | |
2301 | goto out_free; | |
29de5f6a JA |
2302 | /* any defer here is final, must blocking retry */ |
2303 | if (!(req->flags & REQ_F_NOWAIT)) | |
2304 | req->flags |= REQ_F_MUST_PUNT; | |
f67676d1 JA |
2305 | return -EAGAIN; |
2306 | } | |
2b188cc1 | 2307 | } |
f67676d1 | 2308 | out_free: |
1e95081c | 2309 | kfree(iovec); |
99bc4c38 | 2310 | req->flags &= ~REQ_F_NEED_CLEANUP; |
2b188cc1 JA |
2311 | return ret; |
2312 | } | |
2313 | ||
3529d8c2 JA |
2314 | static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2315 | bool force_nonblock) | |
f67676d1 | 2316 | { |
3529d8c2 JA |
2317 | struct io_async_ctx *io; |
2318 | struct iov_iter iter; | |
f67676d1 JA |
2319 | ssize_t ret; |
2320 | ||
3529d8c2 JA |
2321 | ret = io_prep_rw(req, sqe, force_nonblock); |
2322 | if (ret) | |
2323 | return ret; | |
f67676d1 | 2324 | |
3529d8c2 JA |
2325 | if (unlikely(!(req->file->f_mode & FMODE_WRITE))) |
2326 | return -EBADF; | |
f67676d1 | 2327 | |
5f798bea PB |
2328 | /* either don't need iovec imported or already have it */ |
2329 | if (!req->io || req->flags & REQ_F_NEED_CLEANUP) | |
3529d8c2 JA |
2330 | return 0; |
2331 | ||
2332 | io = req->io; | |
2333 | io->rw.iov = io->rw.fast_iov; | |
2334 | req->io = NULL; | |
2335 | ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter); | |
2336 | req->io = io; | |
2337 | if (ret < 0) | |
2338 | return ret; | |
2339 | ||
2340 | io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter); | |
2341 | return 0; | |
f67676d1 JA |
2342 | } |
2343 | ||
267bc904 | 2344 | static int io_write(struct io_kiocb *req, struct io_kiocb **nxt, |
8358e3a8 | 2345 | bool force_nonblock) |
2b188cc1 JA |
2346 | { |
2347 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
9adbd45d | 2348 | struct kiocb *kiocb = &req->rw.kiocb; |
2b188cc1 | 2349 | struct iov_iter iter; |
31b51510 | 2350 | size_t iov_count; |
f67676d1 | 2351 | ssize_t ret, io_size; |
2b188cc1 | 2352 | |
3529d8c2 | 2353 | ret = io_import_iovec(WRITE, req, &iovec, &iter); |
06b76d44 JA |
2354 | if (ret < 0) |
2355 | return ret; | |
2b188cc1 | 2356 | |
fd6c2e4c JA |
2357 | /* Ensure we clear previously set non-block flag */ |
2358 | if (!force_nonblock) | |
9adbd45d | 2359 | req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT; |
fd6c2e4c | 2360 | |
797f3f53 | 2361 | req->result = 0; |
f67676d1 | 2362 | io_size = ret; |
9e645e11 | 2363 | if (req->flags & REQ_F_LINK) |
f67676d1 | 2364 | req->result = io_size; |
9e645e11 | 2365 | |
f67676d1 JA |
2366 | /* |
2367 | * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so | |
2368 | * we know to async punt it even if it was opened O_NONBLOCK | |
2369 | */ | |
29de5f6a | 2370 | if (force_nonblock && !io_file_supports_async(req->file)) |
f67676d1 | 2371 | goto copy_iov; |
31b51510 | 2372 | |
10d59345 JA |
2373 | /* file path doesn't support NOWAIT for non-direct_IO */ |
2374 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) && | |
2375 | (req->flags & REQ_F_ISREG)) | |
f67676d1 | 2376 | goto copy_iov; |
31b51510 | 2377 | |
f67676d1 | 2378 | iov_count = iov_iter_count(&iter); |
9adbd45d | 2379 | ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count); |
2b188cc1 | 2380 | if (!ret) { |
9bf7933f RP |
2381 | ssize_t ret2; |
2382 | ||
2b188cc1 JA |
2383 | /* |
2384 | * Open-code file_start_write here to grab freeze protection, | |
2385 | * which will be released by another thread in | |
2386 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
2387 | * released so that it doesn't complain about the held lock when | |
2388 | * we return to userspace. | |
2389 | */ | |
491381ce | 2390 | if (req->flags & REQ_F_ISREG) { |
9adbd45d | 2391 | __sb_start_write(file_inode(req->file)->i_sb, |
2b188cc1 | 2392 | SB_FREEZE_WRITE, true); |
9adbd45d | 2393 | __sb_writers_release(file_inode(req->file)->i_sb, |
2b188cc1 JA |
2394 | SB_FREEZE_WRITE); |
2395 | } | |
2396 | kiocb->ki_flags |= IOCB_WRITE; | |
9bf7933f | 2397 | |
9adbd45d JA |
2398 | if (req->file->f_op->write_iter) |
2399 | ret2 = call_write_iter(req->file, kiocb, &iter); | |
32960613 | 2400 | else |
9adbd45d | 2401 | ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter); |
faac996c JA |
2402 | /* |
2403 | * Raw bdev writes will -EOPNOTSUPP for IOCB_NOWAIT. Just | |
2404 | * retry them without IOCB_NOWAIT. | |
2405 | */ | |
2406 | if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT)) | |
2407 | ret2 = -EAGAIN; | |
f67676d1 | 2408 | if (!force_nonblock || ret2 != -EAGAIN) { |
bcaec089 | 2409 | kiocb_done(kiocb, ret2, nxt); |
f67676d1 JA |
2410 | } else { |
2411 | copy_iov: | |
b7bb4f7d | 2412 | ret = io_setup_async_rw(req, io_size, iovec, |
f67676d1 JA |
2413 | inline_vecs, &iter); |
2414 | if (ret) | |
2415 | goto out_free; | |
29de5f6a JA |
2416 | /* any defer here is final, must blocking retry */ |
2417 | req->flags |= REQ_F_MUST_PUNT; | |
f67676d1 JA |
2418 | return -EAGAIN; |
2419 | } | |
2b188cc1 | 2420 | } |
31b51510 | 2421 | out_free: |
99bc4c38 | 2422 | req->flags &= ~REQ_F_NEED_CLEANUP; |
1e95081c | 2423 | kfree(iovec); |
2b188cc1 JA |
2424 | return ret; |
2425 | } | |
2426 | ||
7d67af2c PB |
2427 | static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2428 | { | |
2429 | struct io_splice* sp = &req->splice; | |
2430 | unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL; | |
2431 | int ret; | |
2432 | ||
2433 | if (req->flags & REQ_F_NEED_CLEANUP) | |
2434 | return 0; | |
2435 | ||
2436 | sp->file_in = NULL; | |
2437 | sp->off_in = READ_ONCE(sqe->splice_off_in); | |
2438 | sp->off_out = READ_ONCE(sqe->off); | |
2439 | sp->len = READ_ONCE(sqe->len); | |
2440 | sp->flags = READ_ONCE(sqe->splice_flags); | |
2441 | ||
2442 | if (unlikely(sp->flags & ~valid_flags)) | |
2443 | return -EINVAL; | |
2444 | ||
2445 | ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in, | |
2446 | (sp->flags & SPLICE_F_FD_IN_FIXED)); | |
2447 | if (ret) | |
2448 | return ret; | |
2449 | req->flags |= REQ_F_NEED_CLEANUP; | |
2450 | ||
2451 | if (!S_ISREG(file_inode(sp->file_in)->i_mode)) | |
2452 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
2453 | ||
2454 | return 0; | |
2455 | } | |
2456 | ||
2457 | static bool io_splice_punt(struct file *file) | |
2458 | { | |
2459 | if (get_pipe_info(file)) | |
2460 | return false; | |
2461 | if (!io_file_supports_async(file)) | |
2462 | return true; | |
2463 | return !(file->f_mode & O_NONBLOCK); | |
2464 | } | |
2465 | ||
2466 | static int io_splice(struct io_kiocb *req, struct io_kiocb **nxt, | |
2467 | bool force_nonblock) | |
2468 | { | |
2469 | struct io_splice *sp = &req->splice; | |
2470 | struct file *in = sp->file_in; | |
2471 | struct file *out = sp->file_out; | |
2472 | unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED; | |
2473 | loff_t *poff_in, *poff_out; | |
2474 | long ret; | |
2475 | ||
2476 | if (force_nonblock) { | |
2477 | if (io_splice_punt(in) || io_splice_punt(out)) | |
2478 | return -EAGAIN; | |
2479 | flags |= SPLICE_F_NONBLOCK; | |
2480 | } | |
2481 | ||
2482 | poff_in = (sp->off_in == -1) ? NULL : &sp->off_in; | |
2483 | poff_out = (sp->off_out == -1) ? NULL : &sp->off_out; | |
2484 | ret = do_splice(in, poff_in, out, poff_out, sp->len, flags); | |
2485 | if (force_nonblock && ret == -EAGAIN) | |
2486 | return -EAGAIN; | |
2487 | ||
2488 | io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED)); | |
2489 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
2490 | ||
2491 | io_cqring_add_event(req, ret); | |
2492 | if (ret != sp->len) | |
2493 | req_set_fail_links(req); | |
2494 | io_put_req_find_next(req, nxt); | |
2495 | return 0; | |
2496 | } | |
2497 | ||
2b188cc1 JA |
2498 | /* |
2499 | * IORING_OP_NOP just posts a completion event, nothing else. | |
2500 | */ | |
78e19bbe | 2501 | static int io_nop(struct io_kiocb *req) |
2b188cc1 JA |
2502 | { |
2503 | struct io_ring_ctx *ctx = req->ctx; | |
2b188cc1 | 2504 | |
def596e9 JA |
2505 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
2506 | return -EINVAL; | |
2507 | ||
78e19bbe | 2508 | io_cqring_add_event(req, 0); |
e65ef56d | 2509 | io_put_req(req); |
2b188cc1 JA |
2510 | return 0; |
2511 | } | |
2512 | ||
3529d8c2 | 2513 | static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
c992fe29 | 2514 | { |
6b06314c | 2515 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 2516 | |
09bb8394 JA |
2517 | if (!req->file) |
2518 | return -EBADF; | |
c992fe29 | 2519 | |
6b06314c | 2520 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 2521 | return -EINVAL; |
edafccee | 2522 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
2523 | return -EINVAL; |
2524 | ||
8ed8d3c3 JA |
2525 | req->sync.flags = READ_ONCE(sqe->fsync_flags); |
2526 | if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC)) | |
2527 | return -EINVAL; | |
2528 | ||
2529 | req->sync.off = READ_ONCE(sqe->off); | |
2530 | req->sync.len = READ_ONCE(sqe->len); | |
c992fe29 CH |
2531 | return 0; |
2532 | } | |
2533 | ||
8ed8d3c3 JA |
2534 | static bool io_req_cancelled(struct io_kiocb *req) |
2535 | { | |
2536 | if (req->work.flags & IO_WQ_WORK_CANCEL) { | |
2537 | req_set_fail_links(req); | |
2538 | io_cqring_add_event(req, -ECANCELED); | |
2539 | io_put_req(req); | |
2540 | return true; | |
2541 | } | |
2542 | ||
2543 | return false; | |
2544 | } | |
2545 | ||
78912934 JA |
2546 | static void io_link_work_cb(struct io_wq_work **workptr) |
2547 | { | |
2548 | struct io_wq_work *work = *workptr; | |
2549 | struct io_kiocb *link = work->data; | |
2550 | ||
2551 | io_queue_linked_timeout(link); | |
5eae8619 | 2552 | io_wq_submit_work(workptr); |
78912934 JA |
2553 | } |
2554 | ||
2555 | static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt) | |
2556 | { | |
2557 | struct io_kiocb *link; | |
2558 | ||
deb6dc05 | 2559 | io_prep_next_work(nxt, &link); |
78912934 JA |
2560 | *workptr = &nxt->work; |
2561 | if (link) { | |
78912934 JA |
2562 | nxt->work.func = io_link_work_cb; |
2563 | nxt->work.data = link; | |
2564 | } | |
2565 | } | |
2566 | ||
5ea62161 | 2567 | static void __io_fsync(struct io_kiocb *req, struct io_kiocb **nxt) |
8ed8d3c3 | 2568 | { |
8ed8d3c3 | 2569 | loff_t end = req->sync.off + req->sync.len; |
8ed8d3c3 JA |
2570 | int ret; |
2571 | ||
9adbd45d | 2572 | ret = vfs_fsync_range(req->file, req->sync.off, |
8ed8d3c3 JA |
2573 | end > 0 ? end : LLONG_MAX, |
2574 | req->sync.flags & IORING_FSYNC_DATASYNC); | |
2575 | if (ret < 0) | |
2576 | req_set_fail_links(req); | |
2577 | io_cqring_add_event(req, ret); | |
5ea62161 PB |
2578 | io_put_req_find_next(req, nxt); |
2579 | } | |
2580 | ||
2581 | static void io_fsync_finish(struct io_wq_work **workptr) | |
2582 | { | |
2583 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
2584 | struct io_kiocb *nxt = NULL; | |
2585 | ||
2586 | if (io_req_cancelled(req)) | |
2587 | return; | |
2588 | __io_fsync(req, &nxt); | |
8ed8d3c3 | 2589 | if (nxt) |
78912934 | 2590 | io_wq_assign_next(workptr, nxt); |
8ed8d3c3 JA |
2591 | } |
2592 | ||
fc4df999 JA |
2593 | static int io_fsync(struct io_kiocb *req, struct io_kiocb **nxt, |
2594 | bool force_nonblock) | |
c992fe29 | 2595 | { |
c992fe29 | 2596 | /* fsync always requires a blocking context */ |
8ed8d3c3 JA |
2597 | if (force_nonblock) { |
2598 | io_put_req(req); | |
2599 | req->work.func = io_fsync_finish; | |
c992fe29 | 2600 | return -EAGAIN; |
8ed8d3c3 | 2601 | } |
5ea62161 | 2602 | __io_fsync(req, nxt); |
c992fe29 CH |
2603 | return 0; |
2604 | } | |
2605 | ||
5ea62161 | 2606 | static void __io_fallocate(struct io_kiocb *req, struct io_kiocb **nxt) |
8ed8d3c3 | 2607 | { |
8ed8d3c3 JA |
2608 | int ret; |
2609 | ||
7fbeb95d PB |
2610 | if (io_req_cancelled(req)) |
2611 | return; | |
2612 | ||
d63d1b5e JA |
2613 | ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off, |
2614 | req->sync.len); | |
8ed8d3c3 JA |
2615 | if (ret < 0) |
2616 | req_set_fail_links(req); | |
2617 | io_cqring_add_event(req, ret); | |
5ea62161 PB |
2618 | io_put_req_find_next(req, nxt); |
2619 | } | |
2620 | ||
2621 | static void io_fallocate_finish(struct io_wq_work **workptr) | |
2622 | { | |
2623 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
2624 | struct io_kiocb *nxt = NULL; | |
2625 | ||
2626 | __io_fallocate(req, &nxt); | |
8ed8d3c3 | 2627 | if (nxt) |
78912934 | 2628 | io_wq_assign_next(workptr, nxt); |
5d17b4a4 JA |
2629 | } |
2630 | ||
d63d1b5e JA |
2631 | static int io_fallocate_prep(struct io_kiocb *req, |
2632 | const struct io_uring_sqe *sqe) | |
2633 | { | |
2634 | if (sqe->ioprio || sqe->buf_index || sqe->rw_flags) | |
2635 | return -EINVAL; | |
2636 | ||
2637 | req->sync.off = READ_ONCE(sqe->off); | |
2638 | req->sync.len = READ_ONCE(sqe->addr); | |
2639 | req->sync.mode = READ_ONCE(sqe->len); | |
2640 | return 0; | |
2641 | } | |
2642 | ||
2643 | static int io_fallocate(struct io_kiocb *req, struct io_kiocb **nxt, | |
2644 | bool force_nonblock) | |
5d17b4a4 | 2645 | { |
d63d1b5e | 2646 | /* fallocate always requiring blocking context */ |
8ed8d3c3 JA |
2647 | if (force_nonblock) { |
2648 | io_put_req(req); | |
d63d1b5e | 2649 | req->work.func = io_fallocate_finish; |
5d17b4a4 | 2650 | return -EAGAIN; |
8ed8d3c3 | 2651 | } |
5d17b4a4 | 2652 | |
5ea62161 | 2653 | __io_fallocate(req, nxt); |
5d17b4a4 JA |
2654 | return 0; |
2655 | } | |
2656 | ||
15b71abe | 2657 | static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
b7bb4f7d | 2658 | { |
f8748881 | 2659 | const char __user *fname; |
15b71abe | 2660 | int ret; |
b7bb4f7d | 2661 | |
15b71abe JA |
2662 | if (sqe->ioprio || sqe->buf_index) |
2663 | return -EINVAL; | |
cf3040ca JA |
2664 | if (sqe->flags & IOSQE_FIXED_FILE) |
2665 | return -EBADF; | |
0bdbdd08 PB |
2666 | if (req->flags & REQ_F_NEED_CLEANUP) |
2667 | return 0; | |
03b1230c | 2668 | |
15b71abe | 2669 | req->open.dfd = READ_ONCE(sqe->fd); |
c12cedf2 | 2670 | req->open.how.mode = READ_ONCE(sqe->len); |
f8748881 | 2671 | fname = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
c12cedf2 | 2672 | req->open.how.flags = READ_ONCE(sqe->open_flags); |
3529d8c2 | 2673 | |
f8748881 | 2674 | req->open.filename = getname(fname); |
15b71abe JA |
2675 | if (IS_ERR(req->open.filename)) { |
2676 | ret = PTR_ERR(req->open.filename); | |
2677 | req->open.filename = NULL; | |
2678 | return ret; | |
2679 | } | |
3529d8c2 | 2680 | |
8fef80bf | 2681 | req->flags |= REQ_F_NEED_CLEANUP; |
15b71abe | 2682 | return 0; |
03b1230c JA |
2683 | } |
2684 | ||
cebdb986 | 2685 | static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
aa1fa28f | 2686 | { |
cebdb986 JA |
2687 | struct open_how __user *how; |
2688 | const char __user *fname; | |
2689 | size_t len; | |
0fa03c62 JA |
2690 | int ret; |
2691 | ||
cebdb986 | 2692 | if (sqe->ioprio || sqe->buf_index) |
0fa03c62 | 2693 | return -EINVAL; |
cf3040ca JA |
2694 | if (sqe->flags & IOSQE_FIXED_FILE) |
2695 | return -EBADF; | |
0bdbdd08 PB |
2696 | if (req->flags & REQ_F_NEED_CLEANUP) |
2697 | return 0; | |
0fa03c62 | 2698 | |
cebdb986 JA |
2699 | req->open.dfd = READ_ONCE(sqe->fd); |
2700 | fname = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
2701 | how = u64_to_user_ptr(READ_ONCE(sqe->addr2)); | |
2702 | len = READ_ONCE(sqe->len); | |
0fa03c62 | 2703 | |
cebdb986 JA |
2704 | if (len < OPEN_HOW_SIZE_VER0) |
2705 | return -EINVAL; | |
3529d8c2 | 2706 | |
cebdb986 JA |
2707 | ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how, |
2708 | len); | |
2709 | if (ret) | |
2710 | return ret; | |
3529d8c2 | 2711 | |
cebdb986 JA |
2712 | if (!(req->open.how.flags & O_PATH) && force_o_largefile()) |
2713 | req->open.how.flags |= O_LARGEFILE; | |
0fa03c62 | 2714 | |
cebdb986 JA |
2715 | req->open.filename = getname(fname); |
2716 | if (IS_ERR(req->open.filename)) { | |
2717 | ret = PTR_ERR(req->open.filename); | |
2718 | req->open.filename = NULL; | |
2719 | return ret; | |
2720 | } | |
2721 | ||
8fef80bf | 2722 | req->flags |= REQ_F_NEED_CLEANUP; |
cebdb986 JA |
2723 | return 0; |
2724 | } | |
2725 | ||
2726 | static int io_openat2(struct io_kiocb *req, struct io_kiocb **nxt, | |
2727 | bool force_nonblock) | |
15b71abe JA |
2728 | { |
2729 | struct open_flags op; | |
15b71abe JA |
2730 | struct file *file; |
2731 | int ret; | |
2732 | ||
f86cd20c | 2733 | if (force_nonblock) |
15b71abe | 2734 | return -EAGAIN; |
15b71abe | 2735 | |
cebdb986 | 2736 | ret = build_open_flags(&req->open.how, &op); |
15b71abe JA |
2737 | if (ret) |
2738 | goto err; | |
2739 | ||
cebdb986 | 2740 | ret = get_unused_fd_flags(req->open.how.flags); |
15b71abe JA |
2741 | if (ret < 0) |
2742 | goto err; | |
2743 | ||
2744 | file = do_filp_open(req->open.dfd, req->open.filename, &op); | |
2745 | if (IS_ERR(file)) { | |
2746 | put_unused_fd(ret); | |
2747 | ret = PTR_ERR(file); | |
2748 | } else { | |
2749 | fsnotify_open(file); | |
2750 | fd_install(ret, file); | |
2751 | } | |
2752 | err: | |
2753 | putname(req->open.filename); | |
8fef80bf | 2754 | req->flags &= ~REQ_F_NEED_CLEANUP; |
15b71abe JA |
2755 | if (ret < 0) |
2756 | req_set_fail_links(req); | |
2757 | io_cqring_add_event(req, ret); | |
2758 | io_put_req_find_next(req, nxt); | |
2759 | return 0; | |
2760 | } | |
2761 | ||
cebdb986 JA |
2762 | static int io_openat(struct io_kiocb *req, struct io_kiocb **nxt, |
2763 | bool force_nonblock) | |
2764 | { | |
2765 | req->open.how = build_open_how(req->open.how.flags, req->open.how.mode); | |
2766 | return io_openat2(req, nxt, force_nonblock); | |
2767 | } | |
2768 | ||
3e4827b0 JA |
2769 | static int io_epoll_ctl_prep(struct io_kiocb *req, |
2770 | const struct io_uring_sqe *sqe) | |
2771 | { | |
2772 | #if defined(CONFIG_EPOLL) | |
2773 | if (sqe->ioprio || sqe->buf_index) | |
2774 | return -EINVAL; | |
2775 | ||
2776 | req->epoll.epfd = READ_ONCE(sqe->fd); | |
2777 | req->epoll.op = READ_ONCE(sqe->len); | |
2778 | req->epoll.fd = READ_ONCE(sqe->off); | |
2779 | ||
2780 | if (ep_op_has_event(req->epoll.op)) { | |
2781 | struct epoll_event __user *ev; | |
2782 | ||
2783 | ev = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
2784 | if (copy_from_user(&req->epoll.event, ev, sizeof(*ev))) | |
2785 | return -EFAULT; | |
2786 | } | |
2787 | ||
2788 | return 0; | |
2789 | #else | |
2790 | return -EOPNOTSUPP; | |
2791 | #endif | |
2792 | } | |
2793 | ||
2794 | static int io_epoll_ctl(struct io_kiocb *req, struct io_kiocb **nxt, | |
2795 | bool force_nonblock) | |
2796 | { | |
2797 | #if defined(CONFIG_EPOLL) | |
2798 | struct io_epoll *ie = &req->epoll; | |
2799 | int ret; | |
2800 | ||
2801 | ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock); | |
2802 | if (force_nonblock && ret == -EAGAIN) | |
2803 | return -EAGAIN; | |
2804 | ||
2805 | if (ret < 0) | |
2806 | req_set_fail_links(req); | |
2807 | io_cqring_add_event(req, ret); | |
2808 | io_put_req_find_next(req, nxt); | |
2809 | return 0; | |
2810 | #else | |
2811 | return -EOPNOTSUPP; | |
2812 | #endif | |
2813 | } | |
2814 | ||
c1ca757b JA |
2815 | static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2816 | { | |
2817 | #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU) | |
2818 | if (sqe->ioprio || sqe->buf_index || sqe->off) | |
2819 | return -EINVAL; | |
2820 | ||
2821 | req->madvise.addr = READ_ONCE(sqe->addr); | |
2822 | req->madvise.len = READ_ONCE(sqe->len); | |
2823 | req->madvise.advice = READ_ONCE(sqe->fadvise_advice); | |
2824 | return 0; | |
2825 | #else | |
2826 | return -EOPNOTSUPP; | |
2827 | #endif | |
2828 | } | |
2829 | ||
2830 | static int io_madvise(struct io_kiocb *req, struct io_kiocb **nxt, | |
2831 | bool force_nonblock) | |
2832 | { | |
2833 | #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU) | |
2834 | struct io_madvise *ma = &req->madvise; | |
2835 | int ret; | |
2836 | ||
2837 | if (force_nonblock) | |
2838 | return -EAGAIN; | |
2839 | ||
2840 | ret = do_madvise(ma->addr, ma->len, ma->advice); | |
2841 | if (ret < 0) | |
2842 | req_set_fail_links(req); | |
2843 | io_cqring_add_event(req, ret); | |
2844 | io_put_req_find_next(req, nxt); | |
2845 | return 0; | |
2846 | #else | |
2847 | return -EOPNOTSUPP; | |
2848 | #endif | |
2849 | } | |
2850 | ||
4840e418 JA |
2851 | static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2852 | { | |
2853 | if (sqe->ioprio || sqe->buf_index || sqe->addr) | |
2854 | return -EINVAL; | |
2855 | ||
2856 | req->fadvise.offset = READ_ONCE(sqe->off); | |
2857 | req->fadvise.len = READ_ONCE(sqe->len); | |
2858 | req->fadvise.advice = READ_ONCE(sqe->fadvise_advice); | |
2859 | return 0; | |
2860 | } | |
2861 | ||
2862 | static int io_fadvise(struct io_kiocb *req, struct io_kiocb **nxt, | |
2863 | bool force_nonblock) | |
2864 | { | |
2865 | struct io_fadvise *fa = &req->fadvise; | |
2866 | int ret; | |
2867 | ||
3e69426d JA |
2868 | if (force_nonblock) { |
2869 | switch (fa->advice) { | |
2870 | case POSIX_FADV_NORMAL: | |
2871 | case POSIX_FADV_RANDOM: | |
2872 | case POSIX_FADV_SEQUENTIAL: | |
2873 | break; | |
2874 | default: | |
2875 | return -EAGAIN; | |
2876 | } | |
2877 | } | |
4840e418 JA |
2878 | |
2879 | ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice); | |
2880 | if (ret < 0) | |
2881 | req_set_fail_links(req); | |
2882 | io_cqring_add_event(req, ret); | |
2883 | io_put_req_find_next(req, nxt); | |
2884 | return 0; | |
2885 | } | |
2886 | ||
eddc7ef5 JA |
2887 | static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2888 | { | |
f8748881 | 2889 | const char __user *fname; |
eddc7ef5 JA |
2890 | unsigned lookup_flags; |
2891 | int ret; | |
2892 | ||
2893 | if (sqe->ioprio || sqe->buf_index) | |
2894 | return -EINVAL; | |
cf3040ca JA |
2895 | if (sqe->flags & IOSQE_FIXED_FILE) |
2896 | return -EBADF; | |
0bdbdd08 PB |
2897 | if (req->flags & REQ_F_NEED_CLEANUP) |
2898 | return 0; | |
eddc7ef5 JA |
2899 | |
2900 | req->open.dfd = READ_ONCE(sqe->fd); | |
2901 | req->open.mask = READ_ONCE(sqe->len); | |
f8748881 | 2902 | fname = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
eddc7ef5 | 2903 | req->open.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2)); |
c12cedf2 | 2904 | req->open.how.flags = READ_ONCE(sqe->statx_flags); |
eddc7ef5 | 2905 | |
c12cedf2 | 2906 | if (vfs_stat_set_lookup_flags(&lookup_flags, req->open.how.flags)) |
eddc7ef5 JA |
2907 | return -EINVAL; |
2908 | ||
f8748881 | 2909 | req->open.filename = getname_flags(fname, lookup_flags, NULL); |
eddc7ef5 JA |
2910 | if (IS_ERR(req->open.filename)) { |
2911 | ret = PTR_ERR(req->open.filename); | |
2912 | req->open.filename = NULL; | |
2913 | return ret; | |
2914 | } | |
2915 | ||
8fef80bf | 2916 | req->flags |= REQ_F_NEED_CLEANUP; |
eddc7ef5 JA |
2917 | return 0; |
2918 | } | |
2919 | ||
2920 | static int io_statx(struct io_kiocb *req, struct io_kiocb **nxt, | |
2921 | bool force_nonblock) | |
2922 | { | |
2923 | struct io_open *ctx = &req->open; | |
2924 | unsigned lookup_flags; | |
2925 | struct path path; | |
2926 | struct kstat stat; | |
2927 | int ret; | |
2928 | ||
2929 | if (force_nonblock) | |
2930 | return -EAGAIN; | |
2931 | ||
c12cedf2 | 2932 | if (vfs_stat_set_lookup_flags(&lookup_flags, ctx->how.flags)) |
eddc7ef5 JA |
2933 | return -EINVAL; |
2934 | ||
2935 | retry: | |
2936 | /* filename_lookup() drops it, keep a reference */ | |
2937 | ctx->filename->refcnt++; | |
2938 | ||
2939 | ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags, &path, | |
2940 | NULL); | |
2941 | if (ret) | |
2942 | goto err; | |
2943 | ||
c12cedf2 | 2944 | ret = vfs_getattr(&path, &stat, ctx->mask, ctx->how.flags); |
eddc7ef5 JA |
2945 | path_put(&path); |
2946 | if (retry_estale(ret, lookup_flags)) { | |
2947 | lookup_flags |= LOOKUP_REVAL; | |
2948 | goto retry; | |
2949 | } | |
2950 | if (!ret) | |
2951 | ret = cp_statx(&stat, ctx->buffer); | |
2952 | err: | |
2953 | putname(ctx->filename); | |
8fef80bf | 2954 | req->flags &= ~REQ_F_NEED_CLEANUP; |
eddc7ef5 JA |
2955 | if (ret < 0) |
2956 | req_set_fail_links(req); | |
2957 | io_cqring_add_event(req, ret); | |
2958 | io_put_req_find_next(req, nxt); | |
2959 | return 0; | |
2960 | } | |
2961 | ||
b5dba59e JA |
2962 | static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2963 | { | |
2964 | /* | |
2965 | * If we queue this for async, it must not be cancellable. That would | |
2966 | * leave the 'file' in an undeterminate state. | |
2967 | */ | |
2968 | req->work.flags |= IO_WQ_WORK_NO_CANCEL; | |
2969 | ||
2970 | if (sqe->ioprio || sqe->off || sqe->addr || sqe->len || | |
2971 | sqe->rw_flags || sqe->buf_index) | |
2972 | return -EINVAL; | |
2973 | if (sqe->flags & IOSQE_FIXED_FILE) | |
cf3040ca | 2974 | return -EBADF; |
b5dba59e JA |
2975 | |
2976 | req->close.fd = READ_ONCE(sqe->fd); | |
2977 | if (req->file->f_op == &io_uring_fops || | |
b14cca0c | 2978 | req->close.fd == req->ctx->ring_fd) |
b5dba59e JA |
2979 | return -EBADF; |
2980 | ||
2981 | return 0; | |
2982 | } | |
2983 | ||
a93b3331 PB |
2984 | /* only called when __close_fd_get_file() is done */ |
2985 | static void __io_close_finish(struct io_kiocb *req, struct io_kiocb **nxt) | |
2986 | { | |
2987 | int ret; | |
2988 | ||
2989 | ret = filp_close(req->close.put_file, req->work.files); | |
2990 | if (ret < 0) | |
2991 | req_set_fail_links(req); | |
2992 | io_cqring_add_event(req, ret); | |
2993 | fput(req->close.put_file); | |
2994 | io_put_req_find_next(req, nxt); | |
2995 | } | |
2996 | ||
b5dba59e JA |
2997 | static void io_close_finish(struct io_wq_work **workptr) |
2998 | { | |
2999 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
3000 | struct io_kiocb *nxt = NULL; | |
3001 | ||
7fbeb95d | 3002 | /* not cancellable, don't do io_req_cancelled() */ |
a93b3331 | 3003 | __io_close_finish(req, &nxt); |
b5dba59e JA |
3004 | if (nxt) |
3005 | io_wq_assign_next(workptr, nxt); | |
3006 | } | |
3007 | ||
3008 | static int io_close(struct io_kiocb *req, struct io_kiocb **nxt, | |
3009 | bool force_nonblock) | |
3010 | { | |
3011 | int ret; | |
3012 | ||
3013 | req->close.put_file = NULL; | |
3014 | ret = __close_fd_get_file(req->close.fd, &req->close.put_file); | |
3015 | if (ret < 0) | |
3016 | return ret; | |
3017 | ||
3018 | /* if the file has a flush method, be safe and punt to async */ | |
f86cd20c | 3019 | if (req->close.put_file->f_op->flush && !io_wq_current_is_worker()) |
b5dba59e | 3020 | goto eagain; |
b5dba59e JA |
3021 | |
3022 | /* | |
3023 | * No ->flush(), safely close from here and just punt the | |
3024 | * fput() to async context. | |
3025 | */ | |
a93b3331 PB |
3026 | __io_close_finish(req, nxt); |
3027 | return 0; | |
b5dba59e JA |
3028 | eagain: |
3029 | req->work.func = io_close_finish; | |
1a417f4e JA |
3030 | /* |
3031 | * Do manual async queue here to avoid grabbing files - we don't | |
3032 | * need the files, and it'll cause io_close_finish() to close | |
3033 | * the file again and cause a double CQE entry for this request | |
3034 | */ | |
3035 | io_queue_async_work(req); | |
3036 | return 0; | |
b5dba59e JA |
3037 | } |
3038 | ||
3529d8c2 | 3039 | static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
5d17b4a4 JA |
3040 | { |
3041 | struct io_ring_ctx *ctx = req->ctx; | |
5d17b4a4 JA |
3042 | |
3043 | if (!req->file) | |
3044 | return -EBADF; | |
5d17b4a4 JA |
3045 | |
3046 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
3047 | return -EINVAL; | |
3048 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
3049 | return -EINVAL; | |
3050 | ||
8ed8d3c3 JA |
3051 | req->sync.off = READ_ONCE(sqe->off); |
3052 | req->sync.len = READ_ONCE(sqe->len); | |
3053 | req->sync.flags = READ_ONCE(sqe->sync_range_flags); | |
8ed8d3c3 JA |
3054 | return 0; |
3055 | } | |
3056 | ||
5ea62161 | 3057 | static void __io_sync_file_range(struct io_kiocb *req, struct io_kiocb **nxt) |
8ed8d3c3 | 3058 | { |
8ed8d3c3 JA |
3059 | int ret; |
3060 | ||
9adbd45d | 3061 | ret = sync_file_range(req->file, req->sync.off, req->sync.len, |
8ed8d3c3 JA |
3062 | req->sync.flags); |
3063 | if (ret < 0) | |
3064 | req_set_fail_links(req); | |
3065 | io_cqring_add_event(req, ret); | |
5ea62161 PB |
3066 | io_put_req_find_next(req, nxt); |
3067 | } | |
3068 | ||
3069 | ||
3070 | static void io_sync_file_range_finish(struct io_wq_work **workptr) | |
3071 | { | |
3072 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
3073 | struct io_kiocb *nxt = NULL; | |
3074 | ||
3075 | if (io_req_cancelled(req)) | |
3076 | return; | |
3077 | __io_sync_file_range(req, &nxt); | |
8ed8d3c3 | 3078 | if (nxt) |
78912934 | 3079 | io_wq_assign_next(workptr, nxt); |
5d17b4a4 JA |
3080 | } |
3081 | ||
fc4df999 | 3082 | static int io_sync_file_range(struct io_kiocb *req, struct io_kiocb **nxt, |
5d17b4a4 JA |
3083 | bool force_nonblock) |
3084 | { | |
5d17b4a4 | 3085 | /* sync_file_range always requires a blocking context */ |
8ed8d3c3 JA |
3086 | if (force_nonblock) { |
3087 | io_put_req(req); | |
3088 | req->work.func = io_sync_file_range_finish; | |
5d17b4a4 | 3089 | return -EAGAIN; |
8ed8d3c3 | 3090 | } |
5d17b4a4 | 3091 | |
5ea62161 | 3092 | __io_sync_file_range(req, nxt); |
5d17b4a4 JA |
3093 | return 0; |
3094 | } | |
3095 | ||
02d27d89 PB |
3096 | static int io_setup_async_msg(struct io_kiocb *req, |
3097 | struct io_async_msghdr *kmsg) | |
3098 | { | |
3099 | if (req->io) | |
3100 | return -EAGAIN; | |
3101 | if (io_alloc_async_ctx(req)) { | |
3102 | if (kmsg->iov != kmsg->fast_iov) | |
3103 | kfree(kmsg->iov); | |
3104 | return -ENOMEM; | |
3105 | } | |
3106 | req->flags |= REQ_F_NEED_CLEANUP; | |
3107 | memcpy(&req->io->msg, kmsg, sizeof(*kmsg)); | |
3108 | return -EAGAIN; | |
3109 | } | |
3110 | ||
3529d8c2 | 3111 | static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
03b1230c | 3112 | { |
0fa03c62 | 3113 | #if defined(CONFIG_NET) |
e47293fd | 3114 | struct io_sr_msg *sr = &req->sr_msg; |
3529d8c2 | 3115 | struct io_async_ctx *io = req->io; |
99bc4c38 | 3116 | int ret; |
03b1230c | 3117 | |
e47293fd JA |
3118 | sr->msg_flags = READ_ONCE(sqe->msg_flags); |
3119 | sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
fddaface | 3120 | sr->len = READ_ONCE(sqe->len); |
3529d8c2 | 3121 | |
d8768362 JA |
3122 | #ifdef CONFIG_COMPAT |
3123 | if (req->ctx->compat) | |
3124 | sr->msg_flags |= MSG_CMSG_COMPAT; | |
3125 | #endif | |
3126 | ||
fddaface | 3127 | if (!io || req->opcode == IORING_OP_SEND) |
3529d8c2 | 3128 | return 0; |
5f798bea PB |
3129 | /* iovec is already imported */ |
3130 | if (req->flags & REQ_F_NEED_CLEANUP) | |
3131 | return 0; | |
3529d8c2 | 3132 | |
d9688565 | 3133 | io->msg.iov = io->msg.fast_iov; |
99bc4c38 | 3134 | ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags, |
e47293fd | 3135 | &io->msg.iov); |
99bc4c38 PB |
3136 | if (!ret) |
3137 | req->flags |= REQ_F_NEED_CLEANUP; | |
3138 | return ret; | |
03b1230c | 3139 | #else |
e47293fd | 3140 | return -EOPNOTSUPP; |
03b1230c JA |
3141 | #endif |
3142 | } | |
3143 | ||
fc4df999 JA |
3144 | static int io_sendmsg(struct io_kiocb *req, struct io_kiocb **nxt, |
3145 | bool force_nonblock) | |
aa1fa28f | 3146 | { |
03b1230c | 3147 | #if defined(CONFIG_NET) |
0b416c3e | 3148 | struct io_async_msghdr *kmsg = NULL; |
0fa03c62 JA |
3149 | struct socket *sock; |
3150 | int ret; | |
3151 | ||
3152 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
3153 | return -EINVAL; | |
3154 | ||
3155 | sock = sock_from_file(req->file, &ret); | |
3156 | if (sock) { | |
b7bb4f7d | 3157 | struct io_async_ctx io; |
0fa03c62 JA |
3158 | unsigned flags; |
3159 | ||
03b1230c | 3160 | if (req->io) { |
0b416c3e | 3161 | kmsg = &req->io->msg; |
b537916c | 3162 | kmsg->msg.msg_name = &req->io->msg.addr; |
0b416c3e JA |
3163 | /* if iov is set, it's allocated already */ |
3164 | if (!kmsg->iov) | |
3165 | kmsg->iov = kmsg->fast_iov; | |
3166 | kmsg->msg.msg_iter.iov = kmsg->iov; | |
03b1230c | 3167 | } else { |
3529d8c2 JA |
3168 | struct io_sr_msg *sr = &req->sr_msg; |
3169 | ||
0b416c3e | 3170 | kmsg = &io.msg; |
b537916c | 3171 | kmsg->msg.msg_name = &io.msg.addr; |
3529d8c2 JA |
3172 | |
3173 | io.msg.iov = io.msg.fast_iov; | |
3174 | ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg, | |
3175 | sr->msg_flags, &io.msg.iov); | |
03b1230c | 3176 | if (ret) |
3529d8c2 | 3177 | return ret; |
03b1230c | 3178 | } |
0fa03c62 | 3179 | |
e47293fd JA |
3180 | flags = req->sr_msg.msg_flags; |
3181 | if (flags & MSG_DONTWAIT) | |
3182 | req->flags |= REQ_F_NOWAIT; | |
3183 | else if (force_nonblock) | |
3184 | flags |= MSG_DONTWAIT; | |
3185 | ||
0b416c3e | 3186 | ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags); |
02d27d89 PB |
3187 | if (force_nonblock && ret == -EAGAIN) |
3188 | return io_setup_async_msg(req, kmsg); | |
441cdbd5 JA |
3189 | if (ret == -ERESTARTSYS) |
3190 | ret = -EINTR; | |
0fa03c62 JA |
3191 | } |
3192 | ||
1e95081c | 3193 | if (kmsg && kmsg->iov != kmsg->fast_iov) |
0b416c3e | 3194 | kfree(kmsg->iov); |
99bc4c38 | 3195 | req->flags &= ~REQ_F_NEED_CLEANUP; |
78e19bbe | 3196 | io_cqring_add_event(req, ret); |
4e88d6e7 JA |
3197 | if (ret < 0) |
3198 | req_set_fail_links(req); | |
ec9c02ad | 3199 | io_put_req_find_next(req, nxt); |
5d17b4a4 | 3200 | return 0; |
03b1230c JA |
3201 | #else |
3202 | return -EOPNOTSUPP; | |
aa1fa28f | 3203 | #endif |
03b1230c | 3204 | } |
aa1fa28f | 3205 | |
fddaface JA |
3206 | static int io_send(struct io_kiocb *req, struct io_kiocb **nxt, |
3207 | bool force_nonblock) | |
3208 | { | |
3209 | #if defined(CONFIG_NET) | |
3210 | struct socket *sock; | |
3211 | int ret; | |
3212 | ||
3213 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
3214 | return -EINVAL; | |
3215 | ||
3216 | sock = sock_from_file(req->file, &ret); | |
3217 | if (sock) { | |
3218 | struct io_sr_msg *sr = &req->sr_msg; | |
3219 | struct msghdr msg; | |
3220 | struct iovec iov; | |
3221 | unsigned flags; | |
3222 | ||
3223 | ret = import_single_range(WRITE, sr->buf, sr->len, &iov, | |
3224 | &msg.msg_iter); | |
3225 | if (ret) | |
3226 | return ret; | |
3227 | ||
3228 | msg.msg_name = NULL; | |
3229 | msg.msg_control = NULL; | |
3230 | msg.msg_controllen = 0; | |
3231 | msg.msg_namelen = 0; | |
3232 | ||
3233 | flags = req->sr_msg.msg_flags; | |
3234 | if (flags & MSG_DONTWAIT) | |
3235 | req->flags |= REQ_F_NOWAIT; | |
3236 | else if (force_nonblock) | |
3237 | flags |= MSG_DONTWAIT; | |
3238 | ||
0b7b21e4 JA |
3239 | msg.msg_flags = flags; |
3240 | ret = sock_sendmsg(sock, &msg); | |
fddaface JA |
3241 | if (force_nonblock && ret == -EAGAIN) |
3242 | return -EAGAIN; | |
3243 | if (ret == -ERESTARTSYS) | |
3244 | ret = -EINTR; | |
3245 | } | |
3246 | ||
3247 | io_cqring_add_event(req, ret); | |
3248 | if (ret < 0) | |
3249 | req_set_fail_links(req); | |
3250 | io_put_req_find_next(req, nxt); | |
3251 | return 0; | |
3252 | #else | |
3253 | return -EOPNOTSUPP; | |
3254 | #endif | |
3255 | } | |
3256 | ||
3529d8c2 JA |
3257 | static int io_recvmsg_prep(struct io_kiocb *req, |
3258 | const struct io_uring_sqe *sqe) | |
aa1fa28f JA |
3259 | { |
3260 | #if defined(CONFIG_NET) | |
e47293fd | 3261 | struct io_sr_msg *sr = &req->sr_msg; |
3529d8c2 | 3262 | struct io_async_ctx *io = req->io; |
99bc4c38 | 3263 | int ret; |
3529d8c2 JA |
3264 | |
3265 | sr->msg_flags = READ_ONCE(sqe->msg_flags); | |
3266 | sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
0b7b21e4 | 3267 | sr->len = READ_ONCE(sqe->len); |
06b76d44 | 3268 | |
d8768362 JA |
3269 | #ifdef CONFIG_COMPAT |
3270 | if (req->ctx->compat) | |
3271 | sr->msg_flags |= MSG_CMSG_COMPAT; | |
3272 | #endif | |
3273 | ||
fddaface | 3274 | if (!io || req->opcode == IORING_OP_RECV) |
06b76d44 | 3275 | return 0; |
5f798bea PB |
3276 | /* iovec is already imported */ |
3277 | if (req->flags & REQ_F_NEED_CLEANUP) | |
3278 | return 0; | |
03b1230c | 3279 | |
d9688565 | 3280 | io->msg.iov = io->msg.fast_iov; |
99bc4c38 | 3281 | ret = recvmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags, |
e47293fd | 3282 | &io->msg.uaddr, &io->msg.iov); |
99bc4c38 PB |
3283 | if (!ret) |
3284 | req->flags |= REQ_F_NEED_CLEANUP; | |
3285 | return ret; | |
aa1fa28f | 3286 | #else |
e47293fd | 3287 | return -EOPNOTSUPP; |
aa1fa28f JA |
3288 | #endif |
3289 | } | |
3290 | ||
fc4df999 JA |
3291 | static int io_recvmsg(struct io_kiocb *req, struct io_kiocb **nxt, |
3292 | bool force_nonblock) | |
aa1fa28f JA |
3293 | { |
3294 | #if defined(CONFIG_NET) | |
0b416c3e | 3295 | struct io_async_msghdr *kmsg = NULL; |
03b1230c JA |
3296 | struct socket *sock; |
3297 | int ret; | |
3298 | ||
3299 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
3300 | return -EINVAL; | |
3301 | ||
3302 | sock = sock_from_file(req->file, &ret); | |
3303 | if (sock) { | |
b7bb4f7d | 3304 | struct io_async_ctx io; |
03b1230c JA |
3305 | unsigned flags; |
3306 | ||
03b1230c | 3307 | if (req->io) { |
0b416c3e | 3308 | kmsg = &req->io->msg; |
b537916c | 3309 | kmsg->msg.msg_name = &req->io->msg.addr; |
0b416c3e JA |
3310 | /* if iov is set, it's allocated already */ |
3311 | if (!kmsg->iov) | |
3312 | kmsg->iov = kmsg->fast_iov; | |
3313 | kmsg->msg.msg_iter.iov = kmsg->iov; | |
03b1230c | 3314 | } else { |
3529d8c2 JA |
3315 | struct io_sr_msg *sr = &req->sr_msg; |
3316 | ||
0b416c3e | 3317 | kmsg = &io.msg; |
b537916c | 3318 | kmsg->msg.msg_name = &io.msg.addr; |
3529d8c2 JA |
3319 | |
3320 | io.msg.iov = io.msg.fast_iov; | |
3321 | ret = recvmsg_copy_msghdr(&io.msg.msg, sr->msg, | |
3322 | sr->msg_flags, &io.msg.uaddr, | |
3323 | &io.msg.iov); | |
03b1230c | 3324 | if (ret) |
3529d8c2 | 3325 | return ret; |
03b1230c JA |
3326 | } |
3327 | ||
e47293fd JA |
3328 | flags = req->sr_msg.msg_flags; |
3329 | if (flags & MSG_DONTWAIT) | |
3330 | req->flags |= REQ_F_NOWAIT; | |
3331 | else if (force_nonblock) | |
3332 | flags |= MSG_DONTWAIT; | |
3333 | ||
3334 | ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg, | |
3335 | kmsg->uaddr, flags); | |
02d27d89 PB |
3336 | if (force_nonblock && ret == -EAGAIN) |
3337 | return io_setup_async_msg(req, kmsg); | |
03b1230c JA |
3338 | if (ret == -ERESTARTSYS) |
3339 | ret = -EINTR; | |
3340 | } | |
3341 | ||
1e95081c | 3342 | if (kmsg && kmsg->iov != kmsg->fast_iov) |
0b416c3e | 3343 | kfree(kmsg->iov); |
99bc4c38 | 3344 | req->flags &= ~REQ_F_NEED_CLEANUP; |
03b1230c | 3345 | io_cqring_add_event(req, ret); |
4e88d6e7 JA |
3346 | if (ret < 0) |
3347 | req_set_fail_links(req); | |
03b1230c JA |
3348 | io_put_req_find_next(req, nxt); |
3349 | return 0; | |
0fa03c62 JA |
3350 | #else |
3351 | return -EOPNOTSUPP; | |
3352 | #endif | |
3353 | } | |
5d17b4a4 | 3354 | |
fddaface JA |
3355 | static int io_recv(struct io_kiocb *req, struct io_kiocb **nxt, |
3356 | bool force_nonblock) | |
3357 | { | |
3358 | #if defined(CONFIG_NET) | |
3359 | struct socket *sock; | |
3360 | int ret; | |
3361 | ||
3362 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
3363 | return -EINVAL; | |
3364 | ||
3365 | sock = sock_from_file(req->file, &ret); | |
3366 | if (sock) { | |
3367 | struct io_sr_msg *sr = &req->sr_msg; | |
3368 | struct msghdr msg; | |
3369 | struct iovec iov; | |
3370 | unsigned flags; | |
3371 | ||
3372 | ret = import_single_range(READ, sr->buf, sr->len, &iov, | |
3373 | &msg.msg_iter); | |
3374 | if (ret) | |
3375 | return ret; | |
3376 | ||
3377 | msg.msg_name = NULL; | |
3378 | msg.msg_control = NULL; | |
3379 | msg.msg_controllen = 0; | |
3380 | msg.msg_namelen = 0; | |
3381 | msg.msg_iocb = NULL; | |
3382 | msg.msg_flags = 0; | |
3383 | ||
3384 | flags = req->sr_msg.msg_flags; | |
3385 | if (flags & MSG_DONTWAIT) | |
3386 | req->flags |= REQ_F_NOWAIT; | |
3387 | else if (force_nonblock) | |
3388 | flags |= MSG_DONTWAIT; | |
3389 | ||
0b7b21e4 | 3390 | ret = sock_recvmsg(sock, &msg, flags); |
fddaface JA |
3391 | if (force_nonblock && ret == -EAGAIN) |
3392 | return -EAGAIN; | |
3393 | if (ret == -ERESTARTSYS) | |
3394 | ret = -EINTR; | |
3395 | } | |
3396 | ||
3397 | io_cqring_add_event(req, ret); | |
3398 | if (ret < 0) | |
3399 | req_set_fail_links(req); | |
3400 | io_put_req_find_next(req, nxt); | |
3401 | return 0; | |
3402 | #else | |
3403 | return -EOPNOTSUPP; | |
3404 | #endif | |
3405 | } | |
3406 | ||
3407 | ||
3529d8c2 | 3408 | static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
17f2fe35 JA |
3409 | { |
3410 | #if defined(CONFIG_NET) | |
8ed8d3c3 JA |
3411 | struct io_accept *accept = &req->accept; |
3412 | ||
17f2fe35 JA |
3413 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) |
3414 | return -EINVAL; | |
8042d6ce | 3415 | if (sqe->ioprio || sqe->len || sqe->buf_index) |
17f2fe35 JA |
3416 | return -EINVAL; |
3417 | ||
d55e5f5b JA |
3418 | accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
3419 | accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2)); | |
8ed8d3c3 | 3420 | accept->flags = READ_ONCE(sqe->accept_flags); |
8ed8d3c3 JA |
3421 | return 0; |
3422 | #else | |
3423 | return -EOPNOTSUPP; | |
3424 | #endif | |
3425 | } | |
17f2fe35 | 3426 | |
8ed8d3c3 JA |
3427 | #if defined(CONFIG_NET) |
3428 | static int __io_accept(struct io_kiocb *req, struct io_kiocb **nxt, | |
3429 | bool force_nonblock) | |
3430 | { | |
3431 | struct io_accept *accept = &req->accept; | |
3432 | unsigned file_flags; | |
3433 | int ret; | |
3434 | ||
3435 | file_flags = force_nonblock ? O_NONBLOCK : 0; | |
3436 | ret = __sys_accept4_file(req->file, file_flags, accept->addr, | |
3437 | accept->addr_len, accept->flags); | |
3438 | if (ret == -EAGAIN && force_nonblock) | |
17f2fe35 | 3439 | return -EAGAIN; |
8e3cca12 JA |
3440 | if (ret == -ERESTARTSYS) |
3441 | ret = -EINTR; | |
4e88d6e7 JA |
3442 | if (ret < 0) |
3443 | req_set_fail_links(req); | |
78e19bbe | 3444 | io_cqring_add_event(req, ret); |
ec9c02ad | 3445 | io_put_req_find_next(req, nxt); |
17f2fe35 | 3446 | return 0; |
8ed8d3c3 JA |
3447 | } |
3448 | ||
3449 | static void io_accept_finish(struct io_wq_work **workptr) | |
3450 | { | |
3451 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
3452 | struct io_kiocb *nxt = NULL; | |
3453 | ||
e441d1cf JA |
3454 | io_put_req(req); |
3455 | ||
8ed8d3c3 JA |
3456 | if (io_req_cancelled(req)) |
3457 | return; | |
3458 | __io_accept(req, &nxt, false); | |
3459 | if (nxt) | |
78912934 | 3460 | io_wq_assign_next(workptr, nxt); |
8ed8d3c3 JA |
3461 | } |
3462 | #endif | |
3463 | ||
3464 | static int io_accept(struct io_kiocb *req, struct io_kiocb **nxt, | |
3465 | bool force_nonblock) | |
3466 | { | |
3467 | #if defined(CONFIG_NET) | |
3468 | int ret; | |
3469 | ||
8ed8d3c3 JA |
3470 | ret = __io_accept(req, nxt, force_nonblock); |
3471 | if (ret == -EAGAIN && force_nonblock) { | |
3472 | req->work.func = io_accept_finish; | |
8ed8d3c3 JA |
3473 | return -EAGAIN; |
3474 | } | |
3475 | return 0; | |
0fa03c62 JA |
3476 | #else |
3477 | return -EOPNOTSUPP; | |
3478 | #endif | |
3479 | } | |
5d17b4a4 | 3480 | |
3529d8c2 | 3481 | static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f499a021 JA |
3482 | { |
3483 | #if defined(CONFIG_NET) | |
3529d8c2 JA |
3484 | struct io_connect *conn = &req->connect; |
3485 | struct io_async_ctx *io = req->io; | |
f499a021 | 3486 | |
3fbb51c1 JA |
3487 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) |
3488 | return -EINVAL; | |
3489 | if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags) | |
3490 | return -EINVAL; | |
3491 | ||
3529d8c2 JA |
3492 | conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
3493 | conn->addr_len = READ_ONCE(sqe->addr2); | |
3494 | ||
3495 | if (!io) | |
3496 | return 0; | |
3497 | ||
3498 | return move_addr_to_kernel(conn->addr, conn->addr_len, | |
3fbb51c1 | 3499 | &io->connect.address); |
f499a021 | 3500 | #else |
3fbb51c1 | 3501 | return -EOPNOTSUPP; |
f499a021 JA |
3502 | #endif |
3503 | } | |
3504 | ||
fc4df999 JA |
3505 | static int io_connect(struct io_kiocb *req, struct io_kiocb **nxt, |
3506 | bool force_nonblock) | |
f8e85cf2 JA |
3507 | { |
3508 | #if defined(CONFIG_NET) | |
f499a021 | 3509 | struct io_async_ctx __io, *io; |
f8e85cf2 | 3510 | unsigned file_flags; |
3fbb51c1 | 3511 | int ret; |
f8e85cf2 | 3512 | |
f499a021 JA |
3513 | if (req->io) { |
3514 | io = req->io; | |
3515 | } else { | |
3529d8c2 JA |
3516 | ret = move_addr_to_kernel(req->connect.addr, |
3517 | req->connect.addr_len, | |
3518 | &__io.connect.address); | |
f499a021 JA |
3519 | if (ret) |
3520 | goto out; | |
3521 | io = &__io; | |
3522 | } | |
3523 | ||
3fbb51c1 JA |
3524 | file_flags = force_nonblock ? O_NONBLOCK : 0; |
3525 | ||
3526 | ret = __sys_connect_file(req->file, &io->connect.address, | |
3527 | req->connect.addr_len, file_flags); | |
87f80d62 | 3528 | if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) { |
b7bb4f7d JA |
3529 | if (req->io) |
3530 | return -EAGAIN; | |
3531 | if (io_alloc_async_ctx(req)) { | |
f499a021 JA |
3532 | ret = -ENOMEM; |
3533 | goto out; | |
3534 | } | |
b7bb4f7d | 3535 | memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect)); |
f8e85cf2 | 3536 | return -EAGAIN; |
f499a021 | 3537 | } |
f8e85cf2 JA |
3538 | if (ret == -ERESTARTSYS) |
3539 | ret = -EINTR; | |
f499a021 | 3540 | out: |
4e88d6e7 JA |
3541 | if (ret < 0) |
3542 | req_set_fail_links(req); | |
f8e85cf2 JA |
3543 | io_cqring_add_event(req, ret); |
3544 | io_put_req_find_next(req, nxt); | |
3545 | return 0; | |
3546 | #else | |
3547 | return -EOPNOTSUPP; | |
3548 | #endif | |
3549 | } | |
3550 | ||
b41e9852 | 3551 | static bool io_poll_remove_one(struct io_kiocb *req) |
221c5eb2 JA |
3552 | { |
3553 | struct io_poll_iocb *poll = &req->poll; | |
b41e9852 | 3554 | bool do_complete = false; |
221c5eb2 JA |
3555 | |
3556 | spin_lock(&poll->head->lock); | |
3557 | WRITE_ONCE(poll->canceled, true); | |
392edb45 JA |
3558 | if (!list_empty(&poll->wait.entry)) { |
3559 | list_del_init(&poll->wait.entry); | |
b41e9852 | 3560 | do_complete = true; |
221c5eb2 JA |
3561 | } |
3562 | spin_unlock(&poll->head->lock); | |
78076bb6 | 3563 | hash_del(&req->hash_node); |
b41e9852 JA |
3564 | if (do_complete) { |
3565 | io_cqring_fill_event(req, -ECANCELED); | |
3566 | io_commit_cqring(req->ctx); | |
3567 | req->flags |= REQ_F_COMP_LOCKED; | |
3568 | io_put_req(req); | |
3569 | } | |
3570 | ||
3571 | return do_complete; | |
221c5eb2 JA |
3572 | } |
3573 | ||
3574 | static void io_poll_remove_all(struct io_ring_ctx *ctx) | |
3575 | { | |
78076bb6 | 3576 | struct hlist_node *tmp; |
221c5eb2 | 3577 | struct io_kiocb *req; |
78076bb6 | 3578 | int i; |
221c5eb2 JA |
3579 | |
3580 | spin_lock_irq(&ctx->completion_lock); | |
78076bb6 JA |
3581 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { |
3582 | struct hlist_head *list; | |
3583 | ||
3584 | list = &ctx->cancel_hash[i]; | |
3585 | hlist_for_each_entry_safe(req, tmp, list, hash_node) | |
3586 | io_poll_remove_one(req); | |
221c5eb2 JA |
3587 | } |
3588 | spin_unlock_irq(&ctx->completion_lock); | |
b41e9852 JA |
3589 | |
3590 | io_cqring_ev_posted(ctx); | |
221c5eb2 JA |
3591 | } |
3592 | ||
47f46768 JA |
3593 | static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr) |
3594 | { | |
78076bb6 | 3595 | struct hlist_head *list; |
47f46768 JA |
3596 | struct io_kiocb *req; |
3597 | ||
78076bb6 JA |
3598 | list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)]; |
3599 | hlist_for_each_entry(req, list, hash_node) { | |
b41e9852 JA |
3600 | if (sqe_addr != req->user_data) |
3601 | continue; | |
3602 | if (io_poll_remove_one(req)) | |
eac406c6 | 3603 | return 0; |
b41e9852 | 3604 | return -EALREADY; |
47f46768 JA |
3605 | } |
3606 | ||
3607 | return -ENOENT; | |
3608 | } | |
3609 | ||
3529d8c2 JA |
3610 | static int io_poll_remove_prep(struct io_kiocb *req, |
3611 | const struct io_uring_sqe *sqe) | |
0969e783 | 3612 | { |
0969e783 JA |
3613 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3614 | return -EINVAL; | |
3615 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index || | |
3616 | sqe->poll_events) | |
3617 | return -EINVAL; | |
3618 | ||
3619 | req->poll.addr = READ_ONCE(sqe->addr); | |
0969e783 JA |
3620 | return 0; |
3621 | } | |
3622 | ||
221c5eb2 JA |
3623 | /* |
3624 | * Find a running poll command that matches one specified in sqe->addr, | |
3625 | * and remove it if found. | |
3626 | */ | |
fc4df999 | 3627 | static int io_poll_remove(struct io_kiocb *req) |
221c5eb2 JA |
3628 | { |
3629 | struct io_ring_ctx *ctx = req->ctx; | |
0969e783 | 3630 | u64 addr; |
47f46768 | 3631 | int ret; |
221c5eb2 | 3632 | |
0969e783 | 3633 | addr = req->poll.addr; |
221c5eb2 | 3634 | spin_lock_irq(&ctx->completion_lock); |
0969e783 | 3635 | ret = io_poll_cancel(ctx, addr); |
221c5eb2 JA |
3636 | spin_unlock_irq(&ctx->completion_lock); |
3637 | ||
78e19bbe | 3638 | io_cqring_add_event(req, ret); |
4e88d6e7 JA |
3639 | if (ret < 0) |
3640 | req_set_fail_links(req); | |
e65ef56d | 3641 | io_put_req(req); |
221c5eb2 JA |
3642 | return 0; |
3643 | } | |
3644 | ||
b0dd8a41 | 3645 | static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error) |
221c5eb2 | 3646 | { |
a197f664 JL |
3647 | struct io_ring_ctx *ctx = req->ctx; |
3648 | ||
8c838788 | 3649 | req->poll.done = true; |
b0a20349 | 3650 | io_cqring_fill_event(req, error ? error : mangle_poll(mask)); |
8c838788 | 3651 | io_commit_cqring(ctx); |
221c5eb2 JA |
3652 | } |
3653 | ||
b41e9852 | 3654 | static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt) |
221c5eb2 | 3655 | { |
221c5eb2 | 3656 | struct io_ring_ctx *ctx = req->ctx; |
221c5eb2 | 3657 | |
221c5eb2 | 3658 | spin_lock_irq(&ctx->completion_lock); |
78076bb6 | 3659 | hash_del(&req->hash_node); |
b41e9852 JA |
3660 | io_poll_complete(req, req->result, 0); |
3661 | req->flags |= REQ_F_COMP_LOCKED; | |
3662 | io_put_req_find_next(req, nxt); | |
e94f141b JA |
3663 | spin_unlock_irq(&ctx->completion_lock); |
3664 | ||
3665 | io_cqring_ev_posted(ctx); | |
e94f141b JA |
3666 | } |
3667 | ||
b41e9852 | 3668 | static void io_poll_task_func(struct callback_head *cb) |
f0b493e6 | 3669 | { |
b41e9852 JA |
3670 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); |
3671 | struct io_kiocb *nxt = NULL; | |
f0b493e6 | 3672 | |
b41e9852 JA |
3673 | io_poll_task_handler(req, &nxt); |
3674 | if (nxt) | |
3675 | __io_queue_sqe(nxt, NULL); | |
f0b493e6 JA |
3676 | } |
3677 | ||
221c5eb2 JA |
3678 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, |
3679 | void *key) | |
3680 | { | |
c2f2eb7d JA |
3681 | struct io_kiocb *req = wait->private; |
3682 | struct io_poll_iocb *poll = &req->poll; | |
221c5eb2 | 3683 | __poll_t mask = key_to_poll(key); |
b41e9852 | 3684 | struct task_struct *tsk; |
221c5eb2 JA |
3685 | |
3686 | /* for instances that support it check for an event match first: */ | |
8c838788 JA |
3687 | if (mask && !(mask & poll->events)) |
3688 | return 0; | |
221c5eb2 | 3689 | |
392edb45 | 3690 | list_del_init(&poll->wait.entry); |
221c5eb2 | 3691 | |
b41e9852 JA |
3692 | tsk = req->task; |
3693 | req->result = mask; | |
3694 | init_task_work(&req->task_work, io_poll_task_func); | |
3695 | task_work_add(tsk, &req->task_work, true); | |
3696 | wake_up_process(tsk); | |
221c5eb2 JA |
3697 | return 1; |
3698 | } | |
3699 | ||
3700 | struct io_poll_table { | |
3701 | struct poll_table_struct pt; | |
3702 | struct io_kiocb *req; | |
3703 | int error; | |
3704 | }; | |
3705 | ||
3706 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, | |
3707 | struct poll_table_struct *p) | |
3708 | { | |
3709 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
3710 | ||
3711 | if (unlikely(pt->req->poll.head)) { | |
3712 | pt->error = -EINVAL; | |
3713 | return; | |
3714 | } | |
3715 | ||
3716 | pt->error = 0; | |
3717 | pt->req->poll.head = head; | |
392edb45 | 3718 | add_wait_queue(head, &pt->req->poll.wait); |
221c5eb2 JA |
3719 | } |
3720 | ||
eac406c6 JA |
3721 | static void io_poll_req_insert(struct io_kiocb *req) |
3722 | { | |
3723 | struct io_ring_ctx *ctx = req->ctx; | |
78076bb6 JA |
3724 | struct hlist_head *list; |
3725 | ||
3726 | list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)]; | |
3727 | hlist_add_head(&req->hash_node, list); | |
eac406c6 JA |
3728 | } |
3729 | ||
3529d8c2 | 3730 | static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
221c5eb2 JA |
3731 | { |
3732 | struct io_poll_iocb *poll = &req->poll; | |
221c5eb2 | 3733 | u16 events; |
221c5eb2 JA |
3734 | |
3735 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
3736 | return -EINVAL; | |
3737 | if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
3738 | return -EINVAL; | |
09bb8394 JA |
3739 | if (!poll->file) |
3740 | return -EBADF; | |
221c5eb2 | 3741 | |
221c5eb2 JA |
3742 | events = READ_ONCE(sqe->poll_events); |
3743 | poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP; | |
b41e9852 JA |
3744 | |
3745 | /* task will wait for requests on exit, don't need a ref */ | |
3746 | req->task = current; | |
0969e783 JA |
3747 | return 0; |
3748 | } | |
3749 | ||
3750 | static int io_poll_add(struct io_kiocb *req, struct io_kiocb **nxt) | |
3751 | { | |
3752 | struct io_poll_iocb *poll = &req->poll; | |
3753 | struct io_ring_ctx *ctx = req->ctx; | |
3754 | struct io_poll_table ipt; | |
3755 | bool cancel = false; | |
3756 | __poll_t mask; | |
0969e783 | 3757 | |
78076bb6 | 3758 | INIT_HLIST_NODE(&req->hash_node); |
221c5eb2 | 3759 | |
221c5eb2 | 3760 | poll->head = NULL; |
8c838788 | 3761 | poll->done = false; |
221c5eb2 JA |
3762 | poll->canceled = false; |
3763 | ||
3764 | ipt.pt._qproc = io_poll_queue_proc; | |
3765 | ipt.pt._key = poll->events; | |
3766 | ipt.req = req; | |
3767 | ipt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */ | |
3768 | ||
3769 | /* initialized the list so that we can do list_empty checks */ | |
392edb45 JA |
3770 | INIT_LIST_HEAD(&poll->wait.entry); |
3771 | init_waitqueue_func_entry(&poll->wait, io_poll_wake); | |
c2f2eb7d | 3772 | poll->wait.private = req; |
221c5eb2 | 3773 | |
36703247 JA |
3774 | INIT_LIST_HEAD(&req->list); |
3775 | ||
221c5eb2 | 3776 | mask = vfs_poll(poll->file, &ipt.pt) & poll->events; |
221c5eb2 JA |
3777 | |
3778 | spin_lock_irq(&ctx->completion_lock); | |
8c838788 JA |
3779 | if (likely(poll->head)) { |
3780 | spin_lock(&poll->head->lock); | |
392edb45 | 3781 | if (unlikely(list_empty(&poll->wait.entry))) { |
8c838788 JA |
3782 | if (ipt.error) |
3783 | cancel = true; | |
3784 | ipt.error = 0; | |
3785 | mask = 0; | |
3786 | } | |
3787 | if (mask || ipt.error) | |
392edb45 | 3788 | list_del_init(&poll->wait.entry); |
8c838788 JA |
3789 | else if (cancel) |
3790 | WRITE_ONCE(poll->canceled, true); | |
3791 | else if (!poll->done) /* actually waiting for an event */ | |
eac406c6 | 3792 | io_poll_req_insert(req); |
8c838788 JA |
3793 | spin_unlock(&poll->head->lock); |
3794 | } | |
3795 | if (mask) { /* no async, we'd stolen it */ | |
221c5eb2 | 3796 | ipt.error = 0; |
b0dd8a41 | 3797 | io_poll_complete(req, mask, 0); |
221c5eb2 | 3798 | } |
221c5eb2 JA |
3799 | spin_unlock_irq(&ctx->completion_lock); |
3800 | ||
8c838788 JA |
3801 | if (mask) { |
3802 | io_cqring_ev_posted(ctx); | |
ec9c02ad | 3803 | io_put_req_find_next(req, nxt); |
221c5eb2 | 3804 | } |
8c838788 | 3805 | return ipt.error; |
221c5eb2 JA |
3806 | } |
3807 | ||
5262f567 JA |
3808 | static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer) |
3809 | { | |
ad8a48ac JA |
3810 | struct io_timeout_data *data = container_of(timer, |
3811 | struct io_timeout_data, timer); | |
3812 | struct io_kiocb *req = data->req; | |
3813 | struct io_ring_ctx *ctx = req->ctx; | |
5262f567 JA |
3814 | unsigned long flags; |
3815 | ||
5262f567 JA |
3816 | atomic_inc(&ctx->cq_timeouts); |
3817 | ||
3818 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
ef03681a | 3819 | /* |
11365043 JA |
3820 | * We could be racing with timeout deletion. If the list is empty, |
3821 | * then timeout lookup already found it and will be handling it. | |
ef03681a | 3822 | */ |
842f9612 | 3823 | if (!list_empty(&req->list)) { |
11365043 | 3824 | struct io_kiocb *prev; |
5262f567 | 3825 | |
11365043 JA |
3826 | /* |
3827 | * Adjust the reqs sequence before the current one because it | |
d195a66e | 3828 | * will consume a slot in the cq_ring and the cq_tail |
11365043 JA |
3829 | * pointer will be increased, otherwise other timeout reqs may |
3830 | * return in advance without waiting for enough wait_nr. | |
3831 | */ | |
3832 | prev = req; | |
3833 | list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list) | |
3834 | prev->sequence++; | |
11365043 | 3835 | list_del_init(&req->list); |
11365043 | 3836 | } |
5262f567 | 3837 | |
78e19bbe | 3838 | io_cqring_fill_event(req, -ETIME); |
5262f567 JA |
3839 | io_commit_cqring(ctx); |
3840 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
3841 | ||
3842 | io_cqring_ev_posted(ctx); | |
4e88d6e7 | 3843 | req_set_fail_links(req); |
5262f567 JA |
3844 | io_put_req(req); |
3845 | return HRTIMER_NORESTART; | |
3846 | } | |
3847 | ||
47f46768 JA |
3848 | static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data) |
3849 | { | |
3850 | struct io_kiocb *req; | |
3851 | int ret = -ENOENT; | |
3852 | ||
3853 | list_for_each_entry(req, &ctx->timeout_list, list) { | |
3854 | if (user_data == req->user_data) { | |
3855 | list_del_init(&req->list); | |
3856 | ret = 0; | |
3857 | break; | |
3858 | } | |
3859 | } | |
3860 | ||
3861 | if (ret == -ENOENT) | |
3862 | return ret; | |
3863 | ||
2d28390a | 3864 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
47f46768 JA |
3865 | if (ret == -1) |
3866 | return -EALREADY; | |
3867 | ||
4e88d6e7 | 3868 | req_set_fail_links(req); |
47f46768 JA |
3869 | io_cqring_fill_event(req, -ECANCELED); |
3870 | io_put_req(req); | |
3871 | return 0; | |
3872 | } | |
3873 | ||
3529d8c2 JA |
3874 | static int io_timeout_remove_prep(struct io_kiocb *req, |
3875 | const struct io_uring_sqe *sqe) | |
b29472ee | 3876 | { |
b29472ee JA |
3877 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3878 | return -EINVAL; | |
3879 | if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len) | |
3880 | return -EINVAL; | |
3881 | ||
3882 | req->timeout.addr = READ_ONCE(sqe->addr); | |
3883 | req->timeout.flags = READ_ONCE(sqe->timeout_flags); | |
3884 | if (req->timeout.flags) | |
3885 | return -EINVAL; | |
3886 | ||
b29472ee JA |
3887 | return 0; |
3888 | } | |
3889 | ||
11365043 JA |
3890 | /* |
3891 | * Remove or update an existing timeout command | |
3892 | */ | |
fc4df999 | 3893 | static int io_timeout_remove(struct io_kiocb *req) |
11365043 JA |
3894 | { |
3895 | struct io_ring_ctx *ctx = req->ctx; | |
47f46768 | 3896 | int ret; |
11365043 | 3897 | |
11365043 | 3898 | spin_lock_irq(&ctx->completion_lock); |
b29472ee | 3899 | ret = io_timeout_cancel(ctx, req->timeout.addr); |
11365043 | 3900 | |
47f46768 | 3901 | io_cqring_fill_event(req, ret); |
11365043 JA |
3902 | io_commit_cqring(ctx); |
3903 | spin_unlock_irq(&ctx->completion_lock); | |
5262f567 | 3904 | io_cqring_ev_posted(ctx); |
4e88d6e7 JA |
3905 | if (ret < 0) |
3906 | req_set_fail_links(req); | |
ec9c02ad | 3907 | io_put_req(req); |
11365043 | 3908 | return 0; |
5262f567 JA |
3909 | } |
3910 | ||
3529d8c2 | 3911 | static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2d28390a | 3912 | bool is_timeout_link) |
5262f567 | 3913 | { |
ad8a48ac | 3914 | struct io_timeout_data *data; |
a41525ab | 3915 | unsigned flags; |
5262f567 | 3916 | |
ad8a48ac | 3917 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5262f567 | 3918 | return -EINVAL; |
ad8a48ac | 3919 | if (sqe->ioprio || sqe->buf_index || sqe->len != 1) |
a41525ab | 3920 | return -EINVAL; |
2d28390a JA |
3921 | if (sqe->off && is_timeout_link) |
3922 | return -EINVAL; | |
a41525ab JA |
3923 | flags = READ_ONCE(sqe->timeout_flags); |
3924 | if (flags & ~IORING_TIMEOUT_ABS) | |
5262f567 | 3925 | return -EINVAL; |
bdf20073 | 3926 | |
26a61679 JA |
3927 | req->timeout.count = READ_ONCE(sqe->off); |
3928 | ||
3529d8c2 | 3929 | if (!req->io && io_alloc_async_ctx(req)) |
26a61679 JA |
3930 | return -ENOMEM; |
3931 | ||
3932 | data = &req->io->timeout; | |
ad8a48ac | 3933 | data->req = req; |
ad8a48ac JA |
3934 | req->flags |= REQ_F_TIMEOUT; |
3935 | ||
3936 | if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr))) | |
5262f567 JA |
3937 | return -EFAULT; |
3938 | ||
11365043 | 3939 | if (flags & IORING_TIMEOUT_ABS) |
ad8a48ac | 3940 | data->mode = HRTIMER_MODE_ABS; |
11365043 | 3941 | else |
ad8a48ac | 3942 | data->mode = HRTIMER_MODE_REL; |
11365043 | 3943 | |
ad8a48ac JA |
3944 | hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode); |
3945 | return 0; | |
3946 | } | |
3947 | ||
fc4df999 | 3948 | static int io_timeout(struct io_kiocb *req) |
ad8a48ac JA |
3949 | { |
3950 | unsigned count; | |
3951 | struct io_ring_ctx *ctx = req->ctx; | |
3952 | struct io_timeout_data *data; | |
3953 | struct list_head *entry; | |
3954 | unsigned span = 0; | |
ad8a48ac | 3955 | |
2d28390a | 3956 | data = &req->io->timeout; |
93bd25bb | 3957 | |
5262f567 JA |
3958 | /* |
3959 | * sqe->off holds how many events that need to occur for this | |
93bd25bb JA |
3960 | * timeout event to be satisfied. If it isn't set, then this is |
3961 | * a pure timeout request, sequence isn't used. | |
5262f567 | 3962 | */ |
26a61679 | 3963 | count = req->timeout.count; |
93bd25bb JA |
3964 | if (!count) { |
3965 | req->flags |= REQ_F_TIMEOUT_NOSEQ; | |
3966 | spin_lock_irq(&ctx->completion_lock); | |
3967 | entry = ctx->timeout_list.prev; | |
3968 | goto add; | |
3969 | } | |
5262f567 JA |
3970 | |
3971 | req->sequence = ctx->cached_sq_head + count - 1; | |
2d28390a | 3972 | data->seq_offset = count; |
5262f567 JA |
3973 | |
3974 | /* | |
3975 | * Insertion sort, ensuring the first entry in the list is always | |
3976 | * the one we need first. | |
3977 | */ | |
5262f567 JA |
3978 | spin_lock_irq(&ctx->completion_lock); |
3979 | list_for_each_prev(entry, &ctx->timeout_list) { | |
3980 | struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list); | |
5da0fb1a | 3981 | unsigned nxt_sq_head; |
3982 | long long tmp, tmp_nxt; | |
2d28390a | 3983 | u32 nxt_offset = nxt->io->timeout.seq_offset; |
5262f567 | 3984 | |
93bd25bb JA |
3985 | if (nxt->flags & REQ_F_TIMEOUT_NOSEQ) |
3986 | continue; | |
3987 | ||
5da0fb1a | 3988 | /* |
3989 | * Since cached_sq_head + count - 1 can overflow, use type long | |
3990 | * long to store it. | |
3991 | */ | |
3992 | tmp = (long long)ctx->cached_sq_head + count - 1; | |
cc42e0ac PB |
3993 | nxt_sq_head = nxt->sequence - nxt_offset + 1; |
3994 | tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1; | |
5da0fb1a | 3995 | |
3996 | /* | |
3997 | * cached_sq_head may overflow, and it will never overflow twice | |
3998 | * once there is some timeout req still be valid. | |
3999 | */ | |
4000 | if (ctx->cached_sq_head < nxt_sq_head) | |
8b07a65a | 4001 | tmp += UINT_MAX; |
5da0fb1a | 4002 | |
a1f58ba4 | 4003 | if (tmp > tmp_nxt) |
5262f567 | 4004 | break; |
a1f58ba4 | 4005 | |
4006 | /* | |
4007 | * Sequence of reqs after the insert one and itself should | |
4008 | * be adjusted because each timeout req consumes a slot. | |
4009 | */ | |
4010 | span++; | |
4011 | nxt->sequence++; | |
5262f567 | 4012 | } |
a1f58ba4 | 4013 | req->sequence -= span; |
93bd25bb | 4014 | add: |
5262f567 | 4015 | list_add(&req->list, entry); |
ad8a48ac JA |
4016 | data->timer.function = io_timeout_fn; |
4017 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode); | |
5262f567 | 4018 | spin_unlock_irq(&ctx->completion_lock); |
5262f567 JA |
4019 | return 0; |
4020 | } | |
5262f567 | 4021 | |
62755e35 JA |
4022 | static bool io_cancel_cb(struct io_wq_work *work, void *data) |
4023 | { | |
4024 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
4025 | ||
4026 | return req->user_data == (unsigned long) data; | |
4027 | } | |
4028 | ||
e977d6d3 | 4029 | static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr) |
62755e35 | 4030 | { |
62755e35 | 4031 | enum io_wq_cancel cancel_ret; |
62755e35 JA |
4032 | int ret = 0; |
4033 | ||
62755e35 JA |
4034 | cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr); |
4035 | switch (cancel_ret) { | |
4036 | case IO_WQ_CANCEL_OK: | |
4037 | ret = 0; | |
4038 | break; | |
4039 | case IO_WQ_CANCEL_RUNNING: | |
4040 | ret = -EALREADY; | |
4041 | break; | |
4042 | case IO_WQ_CANCEL_NOTFOUND: | |
4043 | ret = -ENOENT; | |
4044 | break; | |
4045 | } | |
4046 | ||
e977d6d3 JA |
4047 | return ret; |
4048 | } | |
4049 | ||
47f46768 JA |
4050 | static void io_async_find_and_cancel(struct io_ring_ctx *ctx, |
4051 | struct io_kiocb *req, __u64 sqe_addr, | |
b0dd8a41 | 4052 | struct io_kiocb **nxt, int success_ret) |
47f46768 JA |
4053 | { |
4054 | unsigned long flags; | |
4055 | int ret; | |
4056 | ||
4057 | ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr); | |
4058 | if (ret != -ENOENT) { | |
4059 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
4060 | goto done; | |
4061 | } | |
4062 | ||
4063 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
4064 | ret = io_timeout_cancel(ctx, sqe_addr); | |
4065 | if (ret != -ENOENT) | |
4066 | goto done; | |
4067 | ret = io_poll_cancel(ctx, sqe_addr); | |
4068 | done: | |
b0dd8a41 JA |
4069 | if (!ret) |
4070 | ret = success_ret; | |
47f46768 JA |
4071 | io_cqring_fill_event(req, ret); |
4072 | io_commit_cqring(ctx); | |
4073 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
4074 | io_cqring_ev_posted(ctx); | |
4075 | ||
4e88d6e7 JA |
4076 | if (ret < 0) |
4077 | req_set_fail_links(req); | |
47f46768 JA |
4078 | io_put_req_find_next(req, nxt); |
4079 | } | |
4080 | ||
3529d8c2 JA |
4081 | static int io_async_cancel_prep(struct io_kiocb *req, |
4082 | const struct io_uring_sqe *sqe) | |
e977d6d3 | 4083 | { |
fbf23849 | 4084 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
e977d6d3 JA |
4085 | return -EINVAL; |
4086 | if (sqe->flags || sqe->ioprio || sqe->off || sqe->len || | |
4087 | sqe->cancel_flags) | |
4088 | return -EINVAL; | |
4089 | ||
fbf23849 JA |
4090 | req->cancel.addr = READ_ONCE(sqe->addr); |
4091 | return 0; | |
4092 | } | |
4093 | ||
4094 | static int io_async_cancel(struct io_kiocb *req, struct io_kiocb **nxt) | |
4095 | { | |
4096 | struct io_ring_ctx *ctx = req->ctx; | |
fbf23849 JA |
4097 | |
4098 | io_async_find_and_cancel(ctx, req, req->cancel.addr, nxt, 0); | |
5262f567 JA |
4099 | return 0; |
4100 | } | |
4101 | ||
05f3fb3c JA |
4102 | static int io_files_update_prep(struct io_kiocb *req, |
4103 | const struct io_uring_sqe *sqe) | |
4104 | { | |
4105 | if (sqe->flags || sqe->ioprio || sqe->rw_flags) | |
4106 | return -EINVAL; | |
4107 | ||
4108 | req->files_update.offset = READ_ONCE(sqe->off); | |
4109 | req->files_update.nr_args = READ_ONCE(sqe->len); | |
4110 | if (!req->files_update.nr_args) | |
4111 | return -EINVAL; | |
4112 | req->files_update.arg = READ_ONCE(sqe->addr); | |
4113 | return 0; | |
4114 | } | |
4115 | ||
4116 | static int io_files_update(struct io_kiocb *req, bool force_nonblock) | |
fbf23849 JA |
4117 | { |
4118 | struct io_ring_ctx *ctx = req->ctx; | |
05f3fb3c JA |
4119 | struct io_uring_files_update up; |
4120 | int ret; | |
fbf23849 | 4121 | |
f86cd20c | 4122 | if (force_nonblock) |
05f3fb3c | 4123 | return -EAGAIN; |
05f3fb3c JA |
4124 | |
4125 | up.offset = req->files_update.offset; | |
4126 | up.fds = req->files_update.arg; | |
4127 | ||
4128 | mutex_lock(&ctx->uring_lock); | |
4129 | ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args); | |
4130 | mutex_unlock(&ctx->uring_lock); | |
4131 | ||
4132 | if (ret < 0) | |
4133 | req_set_fail_links(req); | |
4134 | io_cqring_add_event(req, ret); | |
4135 | io_put_req(req); | |
5262f567 JA |
4136 | return 0; |
4137 | } | |
4138 | ||
3529d8c2 JA |
4139 | static int io_req_defer_prep(struct io_kiocb *req, |
4140 | const struct io_uring_sqe *sqe) | |
f67676d1 | 4141 | { |
e781573e | 4142 | ssize_t ret = 0; |
f67676d1 | 4143 | |
f86cd20c JA |
4144 | if (io_op_defs[req->opcode].file_table) { |
4145 | ret = io_grab_files(req); | |
4146 | if (unlikely(ret)) | |
4147 | return ret; | |
4148 | } | |
4149 | ||
cccf0ee8 JA |
4150 | io_req_work_grab_env(req, &io_op_defs[req->opcode]); |
4151 | ||
d625c6ee | 4152 | switch (req->opcode) { |
e781573e JA |
4153 | case IORING_OP_NOP: |
4154 | break; | |
f67676d1 JA |
4155 | case IORING_OP_READV: |
4156 | case IORING_OP_READ_FIXED: | |
3a6820f2 | 4157 | case IORING_OP_READ: |
3529d8c2 | 4158 | ret = io_read_prep(req, sqe, true); |
f67676d1 JA |
4159 | break; |
4160 | case IORING_OP_WRITEV: | |
4161 | case IORING_OP_WRITE_FIXED: | |
3a6820f2 | 4162 | case IORING_OP_WRITE: |
3529d8c2 | 4163 | ret = io_write_prep(req, sqe, true); |
f67676d1 | 4164 | break; |
0969e783 | 4165 | case IORING_OP_POLL_ADD: |
3529d8c2 | 4166 | ret = io_poll_add_prep(req, sqe); |
0969e783 JA |
4167 | break; |
4168 | case IORING_OP_POLL_REMOVE: | |
3529d8c2 | 4169 | ret = io_poll_remove_prep(req, sqe); |
0969e783 | 4170 | break; |
8ed8d3c3 | 4171 | case IORING_OP_FSYNC: |
3529d8c2 | 4172 | ret = io_prep_fsync(req, sqe); |
8ed8d3c3 JA |
4173 | break; |
4174 | case IORING_OP_SYNC_FILE_RANGE: | |
3529d8c2 | 4175 | ret = io_prep_sfr(req, sqe); |
8ed8d3c3 | 4176 | break; |
03b1230c | 4177 | case IORING_OP_SENDMSG: |
fddaface | 4178 | case IORING_OP_SEND: |
3529d8c2 | 4179 | ret = io_sendmsg_prep(req, sqe); |
03b1230c JA |
4180 | break; |
4181 | case IORING_OP_RECVMSG: | |
fddaface | 4182 | case IORING_OP_RECV: |
3529d8c2 | 4183 | ret = io_recvmsg_prep(req, sqe); |
03b1230c | 4184 | break; |
f499a021 | 4185 | case IORING_OP_CONNECT: |
3529d8c2 | 4186 | ret = io_connect_prep(req, sqe); |
f499a021 | 4187 | break; |
2d28390a | 4188 | case IORING_OP_TIMEOUT: |
3529d8c2 | 4189 | ret = io_timeout_prep(req, sqe, false); |
b7bb4f7d | 4190 | break; |
b29472ee | 4191 | case IORING_OP_TIMEOUT_REMOVE: |
3529d8c2 | 4192 | ret = io_timeout_remove_prep(req, sqe); |
b29472ee | 4193 | break; |
fbf23849 | 4194 | case IORING_OP_ASYNC_CANCEL: |
3529d8c2 | 4195 | ret = io_async_cancel_prep(req, sqe); |
fbf23849 | 4196 | break; |
2d28390a | 4197 | case IORING_OP_LINK_TIMEOUT: |
3529d8c2 | 4198 | ret = io_timeout_prep(req, sqe, true); |
b7bb4f7d | 4199 | break; |
8ed8d3c3 | 4200 | case IORING_OP_ACCEPT: |
3529d8c2 | 4201 | ret = io_accept_prep(req, sqe); |
8ed8d3c3 | 4202 | break; |
d63d1b5e JA |
4203 | case IORING_OP_FALLOCATE: |
4204 | ret = io_fallocate_prep(req, sqe); | |
4205 | break; | |
15b71abe JA |
4206 | case IORING_OP_OPENAT: |
4207 | ret = io_openat_prep(req, sqe); | |
4208 | break; | |
b5dba59e JA |
4209 | case IORING_OP_CLOSE: |
4210 | ret = io_close_prep(req, sqe); | |
4211 | break; | |
05f3fb3c JA |
4212 | case IORING_OP_FILES_UPDATE: |
4213 | ret = io_files_update_prep(req, sqe); | |
4214 | break; | |
eddc7ef5 JA |
4215 | case IORING_OP_STATX: |
4216 | ret = io_statx_prep(req, sqe); | |
4217 | break; | |
4840e418 JA |
4218 | case IORING_OP_FADVISE: |
4219 | ret = io_fadvise_prep(req, sqe); | |
4220 | break; | |
c1ca757b JA |
4221 | case IORING_OP_MADVISE: |
4222 | ret = io_madvise_prep(req, sqe); | |
4223 | break; | |
cebdb986 JA |
4224 | case IORING_OP_OPENAT2: |
4225 | ret = io_openat2_prep(req, sqe); | |
4226 | break; | |
3e4827b0 JA |
4227 | case IORING_OP_EPOLL_CTL: |
4228 | ret = io_epoll_ctl_prep(req, sqe); | |
4229 | break; | |
7d67af2c PB |
4230 | case IORING_OP_SPLICE: |
4231 | ret = io_splice_prep(req, sqe); | |
4232 | break; | |
f67676d1 | 4233 | default: |
e781573e JA |
4234 | printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n", |
4235 | req->opcode); | |
4236 | ret = -EINVAL; | |
b7bb4f7d | 4237 | break; |
f67676d1 JA |
4238 | } |
4239 | ||
b7bb4f7d | 4240 | return ret; |
f67676d1 JA |
4241 | } |
4242 | ||
3529d8c2 | 4243 | static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
de0617e4 | 4244 | { |
a197f664 | 4245 | struct io_ring_ctx *ctx = req->ctx; |
f67676d1 | 4246 | int ret; |
de0617e4 | 4247 | |
9d858b21 BL |
4248 | /* Still need defer if there is pending req in defer list. */ |
4249 | if (!req_need_defer(req) && list_empty(&ctx->defer_list)) | |
de0617e4 JA |
4250 | return 0; |
4251 | ||
3529d8c2 | 4252 | if (!req->io && io_alloc_async_ctx(req)) |
de0617e4 JA |
4253 | return -EAGAIN; |
4254 | ||
3529d8c2 | 4255 | ret = io_req_defer_prep(req, sqe); |
b7bb4f7d | 4256 | if (ret < 0) |
2d28390a | 4257 | return ret; |
2d28390a | 4258 | |
de0617e4 | 4259 | spin_lock_irq(&ctx->completion_lock); |
9d858b21 | 4260 | if (!req_need_defer(req) && list_empty(&ctx->defer_list)) { |
de0617e4 | 4261 | spin_unlock_irq(&ctx->completion_lock); |
de0617e4 JA |
4262 | return 0; |
4263 | } | |
4264 | ||
915967f6 | 4265 | trace_io_uring_defer(ctx, req, req->user_data); |
de0617e4 JA |
4266 | list_add_tail(&req->list, &ctx->defer_list); |
4267 | spin_unlock_irq(&ctx->completion_lock); | |
4268 | return -EIOCBQUEUED; | |
4269 | } | |
4270 | ||
99bc4c38 PB |
4271 | static void io_cleanup_req(struct io_kiocb *req) |
4272 | { | |
4273 | struct io_async_ctx *io = req->io; | |
4274 | ||
4275 | switch (req->opcode) { | |
4276 | case IORING_OP_READV: | |
4277 | case IORING_OP_READ_FIXED: | |
4278 | case IORING_OP_READ: | |
4279 | case IORING_OP_WRITEV: | |
4280 | case IORING_OP_WRITE_FIXED: | |
4281 | case IORING_OP_WRITE: | |
4282 | if (io->rw.iov != io->rw.fast_iov) | |
4283 | kfree(io->rw.iov); | |
4284 | break; | |
4285 | case IORING_OP_SENDMSG: | |
4286 | case IORING_OP_RECVMSG: | |
4287 | if (io->msg.iov != io->msg.fast_iov) | |
4288 | kfree(io->msg.iov); | |
4289 | break; | |
8fef80bf PB |
4290 | case IORING_OP_OPENAT: |
4291 | case IORING_OP_OPENAT2: | |
4292 | case IORING_OP_STATX: | |
4293 | putname(req->open.filename); | |
4294 | break; | |
7d67af2c PB |
4295 | case IORING_OP_SPLICE: |
4296 | io_put_file(req, req->splice.file_in, | |
4297 | (req->splice.flags & SPLICE_F_FD_IN_FIXED)); | |
4298 | break; | |
99bc4c38 PB |
4299 | } |
4300 | ||
4301 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
4302 | } | |
4303 | ||
3529d8c2 JA |
4304 | static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
4305 | struct io_kiocb **nxt, bool force_nonblock) | |
2b188cc1 | 4306 | { |
a197f664 | 4307 | struct io_ring_ctx *ctx = req->ctx; |
d625c6ee | 4308 | int ret; |
2b188cc1 | 4309 | |
d625c6ee | 4310 | switch (req->opcode) { |
2b188cc1 | 4311 | case IORING_OP_NOP: |
78e19bbe | 4312 | ret = io_nop(req); |
2b188cc1 JA |
4313 | break; |
4314 | case IORING_OP_READV: | |
edafccee | 4315 | case IORING_OP_READ_FIXED: |
3a6820f2 | 4316 | case IORING_OP_READ: |
3529d8c2 JA |
4317 | if (sqe) { |
4318 | ret = io_read_prep(req, sqe, force_nonblock); | |
4319 | if (ret < 0) | |
4320 | break; | |
4321 | } | |
267bc904 | 4322 | ret = io_read(req, nxt, force_nonblock); |
edafccee | 4323 | break; |
3529d8c2 | 4324 | case IORING_OP_WRITEV: |
edafccee | 4325 | case IORING_OP_WRITE_FIXED: |
3a6820f2 | 4326 | case IORING_OP_WRITE: |
3529d8c2 JA |
4327 | if (sqe) { |
4328 | ret = io_write_prep(req, sqe, force_nonblock); | |
4329 | if (ret < 0) | |
4330 | break; | |
4331 | } | |
267bc904 | 4332 | ret = io_write(req, nxt, force_nonblock); |
2b188cc1 | 4333 | break; |
c992fe29 | 4334 | case IORING_OP_FSYNC: |
3529d8c2 JA |
4335 | if (sqe) { |
4336 | ret = io_prep_fsync(req, sqe); | |
4337 | if (ret < 0) | |
4338 | break; | |
4339 | } | |
fc4df999 | 4340 | ret = io_fsync(req, nxt, force_nonblock); |
c992fe29 | 4341 | break; |
221c5eb2 | 4342 | case IORING_OP_POLL_ADD: |
3529d8c2 JA |
4343 | if (sqe) { |
4344 | ret = io_poll_add_prep(req, sqe); | |
4345 | if (ret) | |
4346 | break; | |
4347 | } | |
fc4df999 | 4348 | ret = io_poll_add(req, nxt); |
221c5eb2 JA |
4349 | break; |
4350 | case IORING_OP_POLL_REMOVE: | |
3529d8c2 JA |
4351 | if (sqe) { |
4352 | ret = io_poll_remove_prep(req, sqe); | |
4353 | if (ret < 0) | |
4354 | break; | |
4355 | } | |
fc4df999 | 4356 | ret = io_poll_remove(req); |
221c5eb2 | 4357 | break; |
5d17b4a4 | 4358 | case IORING_OP_SYNC_FILE_RANGE: |
3529d8c2 JA |
4359 | if (sqe) { |
4360 | ret = io_prep_sfr(req, sqe); | |
4361 | if (ret < 0) | |
4362 | break; | |
4363 | } | |
fc4df999 | 4364 | ret = io_sync_file_range(req, nxt, force_nonblock); |
5d17b4a4 | 4365 | break; |
0fa03c62 | 4366 | case IORING_OP_SENDMSG: |
fddaface | 4367 | case IORING_OP_SEND: |
3529d8c2 JA |
4368 | if (sqe) { |
4369 | ret = io_sendmsg_prep(req, sqe); | |
4370 | if (ret < 0) | |
4371 | break; | |
4372 | } | |
fddaface JA |
4373 | if (req->opcode == IORING_OP_SENDMSG) |
4374 | ret = io_sendmsg(req, nxt, force_nonblock); | |
4375 | else | |
4376 | ret = io_send(req, nxt, force_nonblock); | |
0fa03c62 | 4377 | break; |
aa1fa28f | 4378 | case IORING_OP_RECVMSG: |
fddaface | 4379 | case IORING_OP_RECV: |
3529d8c2 JA |
4380 | if (sqe) { |
4381 | ret = io_recvmsg_prep(req, sqe); | |
4382 | if (ret) | |
4383 | break; | |
4384 | } | |
fddaface JA |
4385 | if (req->opcode == IORING_OP_RECVMSG) |
4386 | ret = io_recvmsg(req, nxt, force_nonblock); | |
4387 | else | |
4388 | ret = io_recv(req, nxt, force_nonblock); | |
aa1fa28f | 4389 | break; |
5262f567 | 4390 | case IORING_OP_TIMEOUT: |
3529d8c2 JA |
4391 | if (sqe) { |
4392 | ret = io_timeout_prep(req, sqe, false); | |
4393 | if (ret) | |
4394 | break; | |
4395 | } | |
fc4df999 | 4396 | ret = io_timeout(req); |
5262f567 | 4397 | break; |
11365043 | 4398 | case IORING_OP_TIMEOUT_REMOVE: |
3529d8c2 JA |
4399 | if (sqe) { |
4400 | ret = io_timeout_remove_prep(req, sqe); | |
4401 | if (ret) | |
4402 | break; | |
4403 | } | |
fc4df999 | 4404 | ret = io_timeout_remove(req); |
11365043 | 4405 | break; |
17f2fe35 | 4406 | case IORING_OP_ACCEPT: |
3529d8c2 JA |
4407 | if (sqe) { |
4408 | ret = io_accept_prep(req, sqe); | |
4409 | if (ret) | |
4410 | break; | |
4411 | } | |
fc4df999 | 4412 | ret = io_accept(req, nxt, force_nonblock); |
17f2fe35 | 4413 | break; |
f8e85cf2 | 4414 | case IORING_OP_CONNECT: |
3529d8c2 JA |
4415 | if (sqe) { |
4416 | ret = io_connect_prep(req, sqe); | |
4417 | if (ret) | |
4418 | break; | |
4419 | } | |
fc4df999 | 4420 | ret = io_connect(req, nxt, force_nonblock); |
f8e85cf2 | 4421 | break; |
62755e35 | 4422 | case IORING_OP_ASYNC_CANCEL: |
3529d8c2 JA |
4423 | if (sqe) { |
4424 | ret = io_async_cancel_prep(req, sqe); | |
4425 | if (ret) | |
4426 | break; | |
4427 | } | |
fc4df999 | 4428 | ret = io_async_cancel(req, nxt); |
62755e35 | 4429 | break; |
d63d1b5e JA |
4430 | case IORING_OP_FALLOCATE: |
4431 | if (sqe) { | |
4432 | ret = io_fallocate_prep(req, sqe); | |
4433 | if (ret) | |
4434 | break; | |
4435 | } | |
4436 | ret = io_fallocate(req, nxt, force_nonblock); | |
4437 | break; | |
15b71abe JA |
4438 | case IORING_OP_OPENAT: |
4439 | if (sqe) { | |
4440 | ret = io_openat_prep(req, sqe); | |
4441 | if (ret) | |
4442 | break; | |
4443 | } | |
4444 | ret = io_openat(req, nxt, force_nonblock); | |
4445 | break; | |
b5dba59e JA |
4446 | case IORING_OP_CLOSE: |
4447 | if (sqe) { | |
4448 | ret = io_close_prep(req, sqe); | |
4449 | if (ret) | |
4450 | break; | |
4451 | } | |
4452 | ret = io_close(req, nxt, force_nonblock); | |
4453 | break; | |
05f3fb3c JA |
4454 | case IORING_OP_FILES_UPDATE: |
4455 | if (sqe) { | |
4456 | ret = io_files_update_prep(req, sqe); | |
4457 | if (ret) | |
4458 | break; | |
4459 | } | |
4460 | ret = io_files_update(req, force_nonblock); | |
4461 | break; | |
eddc7ef5 JA |
4462 | case IORING_OP_STATX: |
4463 | if (sqe) { | |
4464 | ret = io_statx_prep(req, sqe); | |
4465 | if (ret) | |
4466 | break; | |
4467 | } | |
4468 | ret = io_statx(req, nxt, force_nonblock); | |
4469 | break; | |
4840e418 JA |
4470 | case IORING_OP_FADVISE: |
4471 | if (sqe) { | |
4472 | ret = io_fadvise_prep(req, sqe); | |
4473 | if (ret) | |
4474 | break; | |
4475 | } | |
4476 | ret = io_fadvise(req, nxt, force_nonblock); | |
4477 | break; | |
c1ca757b JA |
4478 | case IORING_OP_MADVISE: |
4479 | if (sqe) { | |
4480 | ret = io_madvise_prep(req, sqe); | |
4481 | if (ret) | |
4482 | break; | |
4483 | } | |
4484 | ret = io_madvise(req, nxt, force_nonblock); | |
4485 | break; | |
cebdb986 JA |
4486 | case IORING_OP_OPENAT2: |
4487 | if (sqe) { | |
4488 | ret = io_openat2_prep(req, sqe); | |
4489 | if (ret) | |
4490 | break; | |
4491 | } | |
4492 | ret = io_openat2(req, nxt, force_nonblock); | |
4493 | break; | |
3e4827b0 JA |
4494 | case IORING_OP_EPOLL_CTL: |
4495 | if (sqe) { | |
4496 | ret = io_epoll_ctl_prep(req, sqe); | |
4497 | if (ret) | |
4498 | break; | |
4499 | } | |
4500 | ret = io_epoll_ctl(req, nxt, force_nonblock); | |
4501 | break; | |
7d67af2c PB |
4502 | case IORING_OP_SPLICE: |
4503 | if (sqe) { | |
4504 | ret = io_splice_prep(req, sqe); | |
4505 | if (ret < 0) | |
4506 | break; | |
4507 | } | |
4508 | ret = io_splice(req, nxt, force_nonblock); | |
4509 | break; | |
2b188cc1 JA |
4510 | default: |
4511 | ret = -EINVAL; | |
4512 | break; | |
4513 | } | |
4514 | ||
def596e9 JA |
4515 | if (ret) |
4516 | return ret; | |
4517 | ||
4518 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
11ba820b JA |
4519 | const bool in_async = io_wq_current_is_worker(); |
4520 | ||
9e645e11 | 4521 | if (req->result == -EAGAIN) |
def596e9 JA |
4522 | return -EAGAIN; |
4523 | ||
11ba820b JA |
4524 | /* workqueue context doesn't hold uring_lock, grab it now */ |
4525 | if (in_async) | |
4526 | mutex_lock(&ctx->uring_lock); | |
4527 | ||
def596e9 | 4528 | io_iopoll_req_issued(req); |
11ba820b JA |
4529 | |
4530 | if (in_async) | |
4531 | mutex_unlock(&ctx->uring_lock); | |
def596e9 JA |
4532 | } |
4533 | ||
4534 | return 0; | |
2b188cc1 JA |
4535 | } |
4536 | ||
561fb04a | 4537 | static void io_wq_submit_work(struct io_wq_work **workptr) |
2b188cc1 | 4538 | { |
561fb04a | 4539 | struct io_wq_work *work = *workptr; |
2b188cc1 | 4540 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
561fb04a JA |
4541 | struct io_kiocb *nxt = NULL; |
4542 | int ret = 0; | |
2b188cc1 | 4543 | |
0c9d5ccd JA |
4544 | /* if NO_CANCEL is set, we must still run the work */ |
4545 | if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) == | |
4546 | IO_WQ_WORK_CANCEL) { | |
561fb04a | 4547 | ret = -ECANCELED; |
0c9d5ccd | 4548 | } |
31b51510 | 4549 | |
561fb04a | 4550 | if (!ret) { |
561fb04a | 4551 | do { |
3529d8c2 | 4552 | ret = io_issue_sqe(req, NULL, &nxt, false); |
561fb04a JA |
4553 | /* |
4554 | * We can get EAGAIN for polled IO even though we're | |
4555 | * forcing a sync submission from here, since we can't | |
4556 | * wait for request slots on the block side. | |
4557 | */ | |
4558 | if (ret != -EAGAIN) | |
4559 | break; | |
4560 | cond_resched(); | |
4561 | } while (1); | |
4562 | } | |
31b51510 | 4563 | |
561fb04a | 4564 | /* drop submission reference */ |
ec9c02ad | 4565 | io_put_req(req); |
817869d2 | 4566 | |
561fb04a | 4567 | if (ret) { |
4e88d6e7 | 4568 | req_set_fail_links(req); |
78e19bbe | 4569 | io_cqring_add_event(req, ret); |
817869d2 | 4570 | io_put_req(req); |
edafccee | 4571 | } |
2b188cc1 | 4572 | |
561fb04a | 4573 | /* if a dependent link is ready, pass it back */ |
78912934 JA |
4574 | if (!ret && nxt) |
4575 | io_wq_assign_next(workptr, nxt); | |
2b188cc1 JA |
4576 | } |
4577 | ||
15b71abe | 4578 | static int io_req_needs_file(struct io_kiocb *req, int fd) |
9e3aa61a | 4579 | { |
d3656344 | 4580 | if (!io_op_defs[req->opcode].needs_file) |
9e3aa61a | 4581 | return 0; |
0b5faf6b | 4582 | if ((fd == -1 || fd == AT_FDCWD) && io_op_defs[req->opcode].fd_non_neg) |
d3656344 JA |
4583 | return 0; |
4584 | return 1; | |
09bb8394 JA |
4585 | } |
4586 | ||
65e19f54 JA |
4587 | static inline struct file *io_file_from_index(struct io_ring_ctx *ctx, |
4588 | int index) | |
4589 | { | |
4590 | struct fixed_file_table *table; | |
4591 | ||
05f3fb3c JA |
4592 | table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT]; |
4593 | return table->files[index & IORING_FILE_TABLE_MASK];; | |
65e19f54 JA |
4594 | } |
4595 | ||
8da11c19 PB |
4596 | static int io_file_get(struct io_submit_state *state, struct io_kiocb *req, |
4597 | int fd, struct file **out_file, bool fixed) | |
09bb8394 | 4598 | { |
a197f664 | 4599 | struct io_ring_ctx *ctx = req->ctx; |
8da11c19 | 4600 | struct file *file; |
09bb8394 | 4601 | |
8da11c19 | 4602 | if (fixed) { |
05f3fb3c | 4603 | if (unlikely(!ctx->file_data || |
09bb8394 JA |
4604 | (unsigned) fd >= ctx->nr_user_files)) |
4605 | return -EBADF; | |
b7620121 | 4606 | fd = array_index_nospec(fd, ctx->nr_user_files); |
8da11c19 PB |
4607 | file = io_file_from_index(ctx, fd); |
4608 | if (!file) | |
08a45173 | 4609 | return -EBADF; |
05f3fb3c | 4610 | percpu_ref_get(&ctx->file_data->refs); |
09bb8394 | 4611 | } else { |
c826bd7a | 4612 | trace_io_uring_file_get(ctx, fd); |
8da11c19 PB |
4613 | file = __io_file_get(state, fd); |
4614 | if (unlikely(!file)) | |
09bb8394 JA |
4615 | return -EBADF; |
4616 | } | |
4617 | ||
8da11c19 | 4618 | *out_file = file; |
09bb8394 JA |
4619 | return 0; |
4620 | } | |
4621 | ||
8da11c19 PB |
4622 | static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req, |
4623 | const struct io_uring_sqe *sqe) | |
4624 | { | |
4625 | unsigned flags; | |
4626 | int fd; | |
4627 | bool fixed; | |
4628 | ||
4629 | flags = READ_ONCE(sqe->flags); | |
4630 | fd = READ_ONCE(sqe->fd); | |
4631 | ||
4632 | if (!io_req_needs_file(req, fd)) | |
4633 | return 0; | |
4634 | ||
4635 | fixed = (flags & IOSQE_FIXED_FILE); | |
4636 | if (unlikely(!fixed && req->needs_fixed_file)) | |
4637 | return -EBADF; | |
4638 | ||
4639 | return io_file_get(state, req, fd, &req->file, fixed); | |
4640 | } | |
4641 | ||
a197f664 | 4642 | static int io_grab_files(struct io_kiocb *req) |
fcb323cc JA |
4643 | { |
4644 | int ret = -EBADF; | |
a197f664 | 4645 | struct io_ring_ctx *ctx = req->ctx; |
fcb323cc | 4646 | |
f86cd20c JA |
4647 | if (req->work.files) |
4648 | return 0; | |
b14cca0c | 4649 | if (!ctx->ring_file) |
b5dba59e JA |
4650 | return -EBADF; |
4651 | ||
fcb323cc JA |
4652 | rcu_read_lock(); |
4653 | spin_lock_irq(&ctx->inflight_lock); | |
4654 | /* | |
4655 | * We use the f_ops->flush() handler to ensure that we can flush | |
4656 | * out work accessing these files if the fd is closed. Check if | |
4657 | * the fd has changed since we started down this path, and disallow | |
4658 | * this operation if it has. | |
4659 | */ | |
b14cca0c | 4660 | if (fcheck(ctx->ring_fd) == ctx->ring_file) { |
fcb323cc JA |
4661 | list_add(&req->inflight_entry, &ctx->inflight_list); |
4662 | req->flags |= REQ_F_INFLIGHT; | |
4663 | req->work.files = current->files; | |
4664 | ret = 0; | |
4665 | } | |
4666 | spin_unlock_irq(&ctx->inflight_lock); | |
4667 | rcu_read_unlock(); | |
4668 | ||
4669 | return ret; | |
4670 | } | |
4671 | ||
2665abfd | 4672 | static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer) |
2b188cc1 | 4673 | { |
ad8a48ac JA |
4674 | struct io_timeout_data *data = container_of(timer, |
4675 | struct io_timeout_data, timer); | |
4676 | struct io_kiocb *req = data->req; | |
2665abfd JA |
4677 | struct io_ring_ctx *ctx = req->ctx; |
4678 | struct io_kiocb *prev = NULL; | |
4679 | unsigned long flags; | |
2665abfd JA |
4680 | |
4681 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
4682 | ||
4683 | /* | |
4684 | * We don't expect the list to be empty, that will only happen if we | |
4685 | * race with the completion of the linked work. | |
4686 | */ | |
4493233e PB |
4687 | if (!list_empty(&req->link_list)) { |
4688 | prev = list_entry(req->link_list.prev, struct io_kiocb, | |
4689 | link_list); | |
5d960724 | 4690 | if (refcount_inc_not_zero(&prev->refs)) { |
4493233e | 4691 | list_del_init(&req->link_list); |
5d960724 JA |
4692 | prev->flags &= ~REQ_F_LINK_TIMEOUT; |
4693 | } else | |
76a46e06 | 4694 | prev = NULL; |
2665abfd JA |
4695 | } |
4696 | ||
4697 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
4698 | ||
4699 | if (prev) { | |
4e88d6e7 | 4700 | req_set_fail_links(prev); |
b0dd8a41 JA |
4701 | io_async_find_and_cancel(ctx, req, prev->user_data, NULL, |
4702 | -ETIME); | |
76a46e06 | 4703 | io_put_req(prev); |
47f46768 JA |
4704 | } else { |
4705 | io_cqring_add_event(req, -ETIME); | |
4706 | io_put_req(req); | |
2665abfd | 4707 | } |
2665abfd JA |
4708 | return HRTIMER_NORESTART; |
4709 | } | |
4710 | ||
ad8a48ac | 4711 | static void io_queue_linked_timeout(struct io_kiocb *req) |
2665abfd | 4712 | { |
76a46e06 | 4713 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd | 4714 | |
76a46e06 JA |
4715 | /* |
4716 | * If the list is now empty, then our linked request finished before | |
4717 | * we got a chance to setup the timer | |
4718 | */ | |
4719 | spin_lock_irq(&ctx->completion_lock); | |
4493233e | 4720 | if (!list_empty(&req->link_list)) { |
2d28390a | 4721 | struct io_timeout_data *data = &req->io->timeout; |
94ae5e77 | 4722 | |
ad8a48ac JA |
4723 | data->timer.function = io_link_timeout_fn; |
4724 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), | |
4725 | data->mode); | |
2665abfd | 4726 | } |
76a46e06 | 4727 | spin_unlock_irq(&ctx->completion_lock); |
2665abfd | 4728 | |
2665abfd | 4729 | /* drop submission reference */ |
76a46e06 JA |
4730 | io_put_req(req); |
4731 | } | |
2665abfd | 4732 | |
ad8a48ac | 4733 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) |
2665abfd JA |
4734 | { |
4735 | struct io_kiocb *nxt; | |
4736 | ||
4737 | if (!(req->flags & REQ_F_LINK)) | |
4738 | return NULL; | |
4739 | ||
4493233e PB |
4740 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, |
4741 | link_list); | |
d625c6ee | 4742 | if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT) |
76a46e06 | 4743 | return NULL; |
2665abfd | 4744 | |
76a46e06 | 4745 | req->flags |= REQ_F_LINK_TIMEOUT; |
76a46e06 | 4746 | return nxt; |
2665abfd JA |
4747 | } |
4748 | ||
3529d8c2 | 4749 | static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2b188cc1 | 4750 | { |
4a0a7a18 | 4751 | struct io_kiocb *linked_timeout; |
f9bd67f6 | 4752 | struct io_kiocb *nxt = NULL; |
193155c8 | 4753 | const struct cred *old_creds = NULL; |
e0c5c576 | 4754 | int ret; |
2b188cc1 | 4755 | |
4a0a7a18 JA |
4756 | again: |
4757 | linked_timeout = io_prep_linked_timeout(req); | |
4758 | ||
193155c8 JA |
4759 | if (req->work.creds && req->work.creds != current_cred()) { |
4760 | if (old_creds) | |
4761 | revert_creds(old_creds); | |
4762 | if (old_creds == req->work.creds) | |
4763 | old_creds = NULL; /* restored original creds */ | |
4764 | else | |
4765 | old_creds = override_creds(req->work.creds); | |
4766 | } | |
4767 | ||
3529d8c2 | 4768 | ret = io_issue_sqe(req, sqe, &nxt, true); |
491381ce JA |
4769 | |
4770 | /* | |
4771 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
4772 | * doesn't support non-blocking read/write attempts | |
4773 | */ | |
4774 | if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) || | |
4775 | (req->flags & REQ_F_MUST_PUNT))) { | |
86a761f8 | 4776 | punt: |
f86cd20c | 4777 | if (io_op_defs[req->opcode].file_table) { |
bbad27b2 PB |
4778 | ret = io_grab_files(req); |
4779 | if (ret) | |
4780 | goto err; | |
2b188cc1 | 4781 | } |
bbad27b2 PB |
4782 | |
4783 | /* | |
4784 | * Queued up for async execution, worker will release | |
4785 | * submit reference when the iocb is actually submitted. | |
4786 | */ | |
4787 | io_queue_async_work(req); | |
4a0a7a18 | 4788 | goto done_req; |
2b188cc1 | 4789 | } |
e65ef56d | 4790 | |
fcb323cc | 4791 | err: |
76a46e06 | 4792 | /* drop submission reference */ |
2a44f467 | 4793 | io_put_req_find_next(req, &nxt); |
e65ef56d | 4794 | |
f9bd67f6 | 4795 | if (linked_timeout) { |
76a46e06 | 4796 | if (!ret) |
f9bd67f6 | 4797 | io_queue_linked_timeout(linked_timeout); |
76a46e06 | 4798 | else |
f9bd67f6 | 4799 | io_put_req(linked_timeout); |
76a46e06 JA |
4800 | } |
4801 | ||
e65ef56d | 4802 | /* and drop final reference, if we failed */ |
9e645e11 | 4803 | if (ret) { |
78e19bbe | 4804 | io_cqring_add_event(req, ret); |
4e88d6e7 | 4805 | req_set_fail_links(req); |
e65ef56d | 4806 | io_put_req(req); |
9e645e11 | 4807 | } |
4a0a7a18 JA |
4808 | done_req: |
4809 | if (nxt) { | |
4810 | req = nxt; | |
4811 | nxt = NULL; | |
86a761f8 PB |
4812 | |
4813 | if (req->flags & REQ_F_FORCE_ASYNC) | |
4814 | goto punt; | |
4a0a7a18 JA |
4815 | goto again; |
4816 | } | |
193155c8 JA |
4817 | if (old_creds) |
4818 | revert_creds(old_creds); | |
2b188cc1 JA |
4819 | } |
4820 | ||
3529d8c2 | 4821 | static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4fe2c963 JL |
4822 | { |
4823 | int ret; | |
4824 | ||
3529d8c2 | 4825 | ret = io_req_defer(req, sqe); |
4fe2c963 JL |
4826 | if (ret) { |
4827 | if (ret != -EIOCBQUEUED) { | |
1118591a | 4828 | fail_req: |
78e19bbe | 4829 | io_cqring_add_event(req, ret); |
4e88d6e7 | 4830 | req_set_fail_links(req); |
78e19bbe | 4831 | io_double_put_req(req); |
4fe2c963 | 4832 | } |
2550878f | 4833 | } else if (req->flags & REQ_F_FORCE_ASYNC) { |
1118591a PB |
4834 | ret = io_req_defer_prep(req, sqe); |
4835 | if (unlikely(ret < 0)) | |
4836 | goto fail_req; | |
ce35a47a JA |
4837 | /* |
4838 | * Never try inline submit of IOSQE_ASYNC is set, go straight | |
4839 | * to async execution. | |
4840 | */ | |
4841 | req->work.flags |= IO_WQ_WORK_CONCURRENT; | |
4842 | io_queue_async_work(req); | |
4843 | } else { | |
3529d8c2 | 4844 | __io_queue_sqe(req, sqe); |
ce35a47a | 4845 | } |
4fe2c963 JL |
4846 | } |
4847 | ||
1b4a51b6 | 4848 | static inline void io_queue_link_head(struct io_kiocb *req) |
4fe2c963 | 4849 | { |
94ae5e77 | 4850 | if (unlikely(req->flags & REQ_F_FAIL_LINK)) { |
1b4a51b6 PB |
4851 | io_cqring_add_event(req, -ECANCELED); |
4852 | io_double_put_req(req); | |
4853 | } else | |
3529d8c2 | 4854 | io_queue_sqe(req, NULL); |
4fe2c963 JL |
4855 | } |
4856 | ||
4e88d6e7 | 4857 | #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \ |
ce35a47a | 4858 | IOSQE_IO_HARDLINK | IOSQE_ASYNC) |
9e645e11 | 4859 | |
3529d8c2 JA |
4860 | static bool io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
4861 | struct io_submit_state *state, struct io_kiocb **link) | |
9e645e11 | 4862 | { |
a197f664 | 4863 | struct io_ring_ctx *ctx = req->ctx; |
32fe525b | 4864 | unsigned int sqe_flags; |
75c6a039 | 4865 | int ret, id; |
9e645e11 | 4866 | |
32fe525b | 4867 | sqe_flags = READ_ONCE(sqe->flags); |
9e645e11 JA |
4868 | |
4869 | /* enforce forwards compatibility on users */ | |
32fe525b | 4870 | if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) { |
9e645e11 | 4871 | ret = -EINVAL; |
196be95c | 4872 | goto err_req; |
9e645e11 JA |
4873 | } |
4874 | ||
75c6a039 JA |
4875 | id = READ_ONCE(sqe->personality); |
4876 | if (id) { | |
193155c8 JA |
4877 | req->work.creds = idr_find(&ctx->personality_idr, id); |
4878 | if (unlikely(!req->work.creds)) { | |
75c6a039 JA |
4879 | ret = -EINVAL; |
4880 | goto err_req; | |
4881 | } | |
193155c8 | 4882 | get_cred(req->work.creds); |
75c6a039 JA |
4883 | } |
4884 | ||
6b47ee6e | 4885 | /* same numerical values with corresponding REQ_F_*, safe to copy */ |
8da11c19 PB |
4886 | req->flags |= sqe_flags & (IOSQE_IO_DRAIN | IOSQE_IO_HARDLINK | |
4887 | IOSQE_ASYNC | IOSQE_FIXED_FILE); | |
9e645e11 | 4888 | |
3529d8c2 | 4889 | ret = io_req_set_file(state, req, sqe); |
9e645e11 JA |
4890 | if (unlikely(ret)) { |
4891 | err_req: | |
78e19bbe JA |
4892 | io_cqring_add_event(req, ret); |
4893 | io_double_put_req(req); | |
2e6e1fde | 4894 | return false; |
9e645e11 JA |
4895 | } |
4896 | ||
9e645e11 JA |
4897 | /* |
4898 | * If we already have a head request, queue this one for async | |
4899 | * submittal once the head completes. If we don't have a head but | |
4900 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
4901 | * submitted sync once the chain is complete. If none of those | |
4902 | * conditions are true (normal request), then just queue it. | |
4903 | */ | |
4904 | if (*link) { | |
9d76377f | 4905 | struct io_kiocb *head = *link; |
4e88d6e7 | 4906 | |
8cdf2193 PB |
4907 | /* |
4908 | * Taking sequential execution of a link, draining both sides | |
4909 | * of the link also fullfils IOSQE_IO_DRAIN semantics for all | |
4910 | * requests in the link. So, it drains the head and the | |
4911 | * next after the link request. The last one is done via | |
4912 | * drain_next flag to persist the effect across calls. | |
4913 | */ | |
711be031 PB |
4914 | if (sqe_flags & IOSQE_IO_DRAIN) { |
4915 | head->flags |= REQ_F_IO_DRAIN; | |
4916 | ctx->drain_next = 1; | |
4917 | } | |
b7bb4f7d | 4918 | if (io_alloc_async_ctx(req)) { |
9e645e11 JA |
4919 | ret = -EAGAIN; |
4920 | goto err_req; | |
4921 | } | |
4922 | ||
3529d8c2 | 4923 | ret = io_req_defer_prep(req, sqe); |
2d28390a | 4924 | if (ret) { |
4e88d6e7 | 4925 | /* fail even hard links since we don't submit */ |
9d76377f | 4926 | head->flags |= REQ_F_FAIL_LINK; |
f67676d1 | 4927 | goto err_req; |
2d28390a | 4928 | } |
9d76377f PB |
4929 | trace_io_uring_link(ctx, req, head); |
4930 | list_add_tail(&req->link_list, &head->link_list); | |
32fe525b PB |
4931 | |
4932 | /* last request of a link, enqueue the link */ | |
4933 | if (!(sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK))) { | |
4934 | io_queue_link_head(head); | |
4935 | *link = NULL; | |
4936 | } | |
9e645e11 | 4937 | } else { |
711be031 PB |
4938 | if (unlikely(ctx->drain_next)) { |
4939 | req->flags |= REQ_F_IO_DRAIN; | |
4940 | req->ctx->drain_next = 0; | |
4941 | } | |
4942 | if (sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) { | |
4943 | req->flags |= REQ_F_LINK; | |
711be031 PB |
4944 | INIT_LIST_HEAD(&req->link_list); |
4945 | ret = io_req_defer_prep(req, sqe); | |
4946 | if (ret) | |
4947 | req->flags |= REQ_F_FAIL_LINK; | |
4948 | *link = req; | |
4949 | } else { | |
4950 | io_queue_sqe(req, sqe); | |
4951 | } | |
9e645e11 | 4952 | } |
2e6e1fde PB |
4953 | |
4954 | return true; | |
9e645e11 JA |
4955 | } |
4956 | ||
9a56a232 JA |
4957 | /* |
4958 | * Batched submission is done, ensure local IO is flushed out. | |
4959 | */ | |
4960 | static void io_submit_state_end(struct io_submit_state *state) | |
4961 | { | |
4962 | blk_finish_plug(&state->plug); | |
3d6770fb | 4963 | io_file_put(state); |
2579f913 | 4964 | if (state->free_reqs) |
6c8a3134 | 4965 | kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs); |
9a56a232 JA |
4966 | } |
4967 | ||
4968 | /* | |
4969 | * Start submission side cache. | |
4970 | */ | |
4971 | static void io_submit_state_start(struct io_submit_state *state, | |
22efde59 | 4972 | unsigned int max_ios) |
9a56a232 JA |
4973 | { |
4974 | blk_start_plug(&state->plug); | |
2579f913 | 4975 | state->free_reqs = 0; |
9a56a232 JA |
4976 | state->file = NULL; |
4977 | state->ios_left = max_ios; | |
4978 | } | |
4979 | ||
2b188cc1 JA |
4980 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
4981 | { | |
75b28aff | 4982 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 4983 | |
caf582c6 PB |
4984 | /* |
4985 | * Ensure any loads from the SQEs are done at this point, | |
4986 | * since once we write the new head, the application could | |
4987 | * write new data to them. | |
4988 | */ | |
4989 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
4990 | } |
4991 | ||
2b188cc1 | 4992 | /* |
3529d8c2 | 4993 | * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory |
2b188cc1 JA |
4994 | * that is mapped by userspace. This means that care needs to be taken to |
4995 | * ensure that reads are stable, as we cannot rely on userspace always | |
4996 | * being a good citizen. If members of the sqe are validated and then later | |
4997 | * used, it's important that those reads are done through READ_ONCE() to | |
4998 | * prevent a re-load down the line. | |
4999 | */ | |
3529d8c2 JA |
5000 | static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req, |
5001 | const struct io_uring_sqe **sqe_ptr) | |
2b188cc1 | 5002 | { |
75b28aff | 5003 | u32 *sq_array = ctx->sq_array; |
2b188cc1 JA |
5004 | unsigned head; |
5005 | ||
5006 | /* | |
5007 | * The cached sq head (or cq tail) serves two purposes: | |
5008 | * | |
5009 | * 1) allows us to batch the cost of updating the user visible | |
5010 | * head updates. | |
5011 | * 2) allows the kernel side to track the head on its own, even | |
5012 | * though the application is the one updating it. | |
5013 | */ | |
ee7d46d9 | 5014 | head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]); |
9835d6fa | 5015 | if (likely(head < ctx->sq_entries)) { |
cf6fd4bd PB |
5016 | /* |
5017 | * All io need record the previous position, if LINK vs DARIN, | |
5018 | * it can be used to mark the position of the first IO in the | |
5019 | * link list. | |
5020 | */ | |
5021 | req->sequence = ctx->cached_sq_head; | |
3529d8c2 JA |
5022 | *sqe_ptr = &ctx->sq_sqes[head]; |
5023 | req->opcode = READ_ONCE((*sqe_ptr)->opcode); | |
5024 | req->user_data = READ_ONCE((*sqe_ptr)->user_data); | |
2b188cc1 JA |
5025 | ctx->cached_sq_head++; |
5026 | return true; | |
5027 | } | |
5028 | ||
5029 | /* drop invalid entries */ | |
5030 | ctx->cached_sq_head++; | |
498ccd9e | 5031 | ctx->cached_sq_dropped++; |
ee7d46d9 | 5032 | WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped); |
2b188cc1 JA |
5033 | return false; |
5034 | } | |
5035 | ||
fb5ccc98 | 5036 | static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr, |
ae9428ca PB |
5037 | struct file *ring_file, int ring_fd, |
5038 | struct mm_struct **mm, bool async) | |
6c271ce2 JA |
5039 | { |
5040 | struct io_submit_state state, *statep = NULL; | |
9e645e11 | 5041 | struct io_kiocb *link = NULL; |
9e645e11 | 5042 | int i, submitted = 0; |
95a1b3ff | 5043 | bool mm_fault = false; |
6c271ce2 | 5044 | |
c4a2ed72 | 5045 | /* if we have a backlog and couldn't flush it all, return BUSY */ |
ad3eb2c8 JA |
5046 | if (test_bit(0, &ctx->sq_check_overflow)) { |
5047 | if (!list_empty(&ctx->cq_overflow_list) && | |
5048 | !io_cqring_overflow_flush(ctx, false)) | |
5049 | return -EBUSY; | |
5050 | } | |
6c271ce2 | 5051 | |
ee7d46d9 PB |
5052 | /* make sure SQ entry isn't read before tail */ |
5053 | nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx)); | |
9ef4f124 | 5054 | |
2b85edfc PB |
5055 | if (!percpu_ref_tryget_many(&ctx->refs, nr)) |
5056 | return -EAGAIN; | |
6c271ce2 JA |
5057 | |
5058 | if (nr > IO_PLUG_THRESHOLD) { | |
22efde59 | 5059 | io_submit_state_start(&state, nr); |
6c271ce2 JA |
5060 | statep = &state; |
5061 | } | |
5062 | ||
b14cca0c PB |
5063 | ctx->ring_fd = ring_fd; |
5064 | ctx->ring_file = ring_file; | |
5065 | ||
6c271ce2 | 5066 | for (i = 0; i < nr; i++) { |
3529d8c2 | 5067 | const struct io_uring_sqe *sqe; |
196be95c | 5068 | struct io_kiocb *req; |
1cb1edb2 | 5069 | int err; |
fb5ccc98 | 5070 | |
196be95c PB |
5071 | req = io_get_req(ctx, statep); |
5072 | if (unlikely(!req)) { | |
5073 | if (!submitted) | |
5074 | submitted = -EAGAIN; | |
fb5ccc98 | 5075 | break; |
196be95c | 5076 | } |
3529d8c2 | 5077 | if (!io_get_sqring(ctx, req, &sqe)) { |
2b85edfc | 5078 | __io_req_do_free(req); |
196be95c PB |
5079 | break; |
5080 | } | |
fb5ccc98 | 5081 | |
d3656344 JA |
5082 | /* will complete beyond this point, count as submitted */ |
5083 | submitted++; | |
5084 | ||
5085 | if (unlikely(req->opcode >= IORING_OP_LAST)) { | |
1cb1edb2 PB |
5086 | err = -EINVAL; |
5087 | fail_req: | |
5088 | io_cqring_add_event(req, err); | |
d3656344 | 5089 | io_double_put_req(req); |
196be95c PB |
5090 | break; |
5091 | } | |
fb5ccc98 | 5092 | |
d3656344 | 5093 | if (io_op_defs[req->opcode].needs_mm && !*mm) { |
95a1b3ff | 5094 | mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm); |
1cb1edb2 PB |
5095 | if (unlikely(mm_fault)) { |
5096 | err = -EFAULT; | |
5097 | goto fail_req; | |
95a1b3ff | 5098 | } |
1cb1edb2 PB |
5099 | use_mm(ctx->sqo_mm); |
5100 | *mm = ctx->sqo_mm; | |
9e645e11 | 5101 | } |
9e645e11 | 5102 | |
cf6fd4bd | 5103 | req->needs_fixed_file = async; |
354420f7 JA |
5104 | trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data, |
5105 | true, async); | |
3529d8c2 | 5106 | if (!io_submit_sqe(req, sqe, statep, &link)) |
2e6e1fde | 5107 | break; |
6c271ce2 JA |
5108 | } |
5109 | ||
9466f437 PB |
5110 | if (unlikely(submitted != nr)) { |
5111 | int ref_used = (submitted == -EAGAIN) ? 0 : submitted; | |
5112 | ||
5113 | percpu_ref_put_many(&ctx->refs, nr - ref_used); | |
5114 | } | |
9e645e11 | 5115 | if (link) |
1b4a51b6 | 5116 | io_queue_link_head(link); |
6c271ce2 JA |
5117 | if (statep) |
5118 | io_submit_state_end(&state); | |
5119 | ||
ae9428ca PB |
5120 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
5121 | io_commit_sqring(ctx); | |
5122 | ||
6c271ce2 JA |
5123 | return submitted; |
5124 | } | |
5125 | ||
5126 | static int io_sq_thread(void *data) | |
5127 | { | |
6c271ce2 JA |
5128 | struct io_ring_ctx *ctx = data; |
5129 | struct mm_struct *cur_mm = NULL; | |
181e448d | 5130 | const struct cred *old_cred; |
6c271ce2 JA |
5131 | mm_segment_t old_fs; |
5132 | DEFINE_WAIT(wait); | |
6c271ce2 | 5133 | unsigned long timeout; |
bdcd3eab | 5134 | int ret = 0; |
6c271ce2 | 5135 | |
206aefde | 5136 | complete(&ctx->completions[1]); |
a4c0b3de | 5137 | |
6c271ce2 JA |
5138 | old_fs = get_fs(); |
5139 | set_fs(USER_DS); | |
181e448d | 5140 | old_cred = override_creds(ctx->creds); |
6c271ce2 | 5141 | |
bdcd3eab | 5142 | timeout = jiffies + ctx->sq_thread_idle; |
2bbcd6d3 | 5143 | while (!kthread_should_park()) { |
fb5ccc98 | 5144 | unsigned int to_submit; |
6c271ce2 | 5145 | |
bdcd3eab | 5146 | if (!list_empty(&ctx->poll_list)) { |
6c271ce2 JA |
5147 | unsigned nr_events = 0; |
5148 | ||
bdcd3eab XW |
5149 | mutex_lock(&ctx->uring_lock); |
5150 | if (!list_empty(&ctx->poll_list)) | |
5151 | io_iopoll_getevents(ctx, &nr_events, 0); | |
5152 | else | |
6c271ce2 | 5153 | timeout = jiffies + ctx->sq_thread_idle; |
bdcd3eab | 5154 | mutex_unlock(&ctx->uring_lock); |
6c271ce2 JA |
5155 | } |
5156 | ||
fb5ccc98 | 5157 | to_submit = io_sqring_entries(ctx); |
c1edbf5f JA |
5158 | |
5159 | /* | |
5160 | * If submit got -EBUSY, flag us as needing the application | |
5161 | * to enter the kernel to reap and flush events. | |
5162 | */ | |
5163 | if (!to_submit || ret == -EBUSY) { | |
7143b5ac SG |
5164 | /* |
5165 | * Drop cur_mm before scheduling, we can't hold it for | |
5166 | * long periods (or over schedule()). Do this before | |
5167 | * adding ourselves to the waitqueue, as the unuse/drop | |
5168 | * may sleep. | |
5169 | */ | |
5170 | if (cur_mm) { | |
5171 | unuse_mm(cur_mm); | |
5172 | mmput(cur_mm); | |
5173 | cur_mm = NULL; | |
5174 | } | |
5175 | ||
6c271ce2 JA |
5176 | /* |
5177 | * We're polling. If we're within the defined idle | |
5178 | * period, then let us spin without work before going | |
c1edbf5f JA |
5179 | * to sleep. The exception is if we got EBUSY doing |
5180 | * more IO, we should wait for the application to | |
5181 | * reap events and wake us up. | |
6c271ce2 | 5182 | */ |
bdcd3eab | 5183 | if (!list_empty(&ctx->poll_list) || |
df069d80 JA |
5184 | (!time_after(jiffies, timeout) && ret != -EBUSY && |
5185 | !percpu_ref_is_dying(&ctx->refs))) { | |
b41e9852 JA |
5186 | if (current->task_works) |
5187 | task_work_run(); | |
9831a90c | 5188 | cond_resched(); |
6c271ce2 JA |
5189 | continue; |
5190 | } | |
5191 | ||
6c271ce2 JA |
5192 | prepare_to_wait(&ctx->sqo_wait, &wait, |
5193 | TASK_INTERRUPTIBLE); | |
5194 | ||
bdcd3eab XW |
5195 | /* |
5196 | * While doing polled IO, before going to sleep, we need | |
5197 | * to check if there are new reqs added to poll_list, it | |
5198 | * is because reqs may have been punted to io worker and | |
5199 | * will be added to poll_list later, hence check the | |
5200 | * poll_list again. | |
5201 | */ | |
5202 | if ((ctx->flags & IORING_SETUP_IOPOLL) && | |
5203 | !list_empty_careful(&ctx->poll_list)) { | |
5204 | finish_wait(&ctx->sqo_wait, &wait); | |
5205 | continue; | |
5206 | } | |
5207 | ||
6c271ce2 | 5208 | /* Tell userspace we may need a wakeup call */ |
75b28aff | 5209 | ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP; |
0d7bae69 SB |
5210 | /* make sure to read SQ tail after writing flags */ |
5211 | smp_mb(); | |
6c271ce2 | 5212 | |
fb5ccc98 | 5213 | to_submit = io_sqring_entries(ctx); |
c1edbf5f | 5214 | if (!to_submit || ret == -EBUSY) { |
2bbcd6d3 | 5215 | if (kthread_should_park()) { |
6c271ce2 JA |
5216 | finish_wait(&ctx->sqo_wait, &wait); |
5217 | break; | |
5218 | } | |
b41e9852 JA |
5219 | if (current->task_works) { |
5220 | task_work_run(); | |
5221 | continue; | |
5222 | } | |
6c271ce2 JA |
5223 | if (signal_pending(current)) |
5224 | flush_signals(current); | |
5225 | schedule(); | |
5226 | finish_wait(&ctx->sqo_wait, &wait); | |
5227 | ||
75b28aff | 5228 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
5229 | continue; |
5230 | } | |
5231 | finish_wait(&ctx->sqo_wait, &wait); | |
5232 | ||
75b28aff | 5233 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
5234 | } |
5235 | ||
8a4955ff | 5236 | mutex_lock(&ctx->uring_lock); |
1d7bb1d5 | 5237 | ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true); |
8a4955ff | 5238 | mutex_unlock(&ctx->uring_lock); |
bdcd3eab | 5239 | timeout = jiffies + ctx->sq_thread_idle; |
6c271ce2 JA |
5240 | } |
5241 | ||
b41e9852 JA |
5242 | if (current->task_works) |
5243 | task_work_run(); | |
5244 | ||
6c271ce2 JA |
5245 | set_fs(old_fs); |
5246 | if (cur_mm) { | |
5247 | unuse_mm(cur_mm); | |
5248 | mmput(cur_mm); | |
5249 | } | |
181e448d | 5250 | revert_creds(old_cred); |
06058632 | 5251 | |
2bbcd6d3 | 5252 | kthread_parkme(); |
06058632 | 5253 | |
6c271ce2 JA |
5254 | return 0; |
5255 | } | |
5256 | ||
bda52162 JA |
5257 | struct io_wait_queue { |
5258 | struct wait_queue_entry wq; | |
5259 | struct io_ring_ctx *ctx; | |
5260 | unsigned to_wait; | |
5261 | unsigned nr_timeouts; | |
5262 | }; | |
5263 | ||
1d7bb1d5 | 5264 | static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush) |
bda52162 JA |
5265 | { |
5266 | struct io_ring_ctx *ctx = iowq->ctx; | |
5267 | ||
5268 | /* | |
d195a66e | 5269 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
5270 | * started waiting. For timeouts, we always want to return to userspace, |
5271 | * regardless of event count. | |
5272 | */ | |
1d7bb1d5 | 5273 | return io_cqring_events(ctx, noflush) >= iowq->to_wait || |
bda52162 JA |
5274 | atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
5275 | } | |
5276 | ||
5277 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
5278 | int wake_flags, void *key) | |
5279 | { | |
5280 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
5281 | wq); | |
5282 | ||
1d7bb1d5 JA |
5283 | /* use noflush == true, as we can't safely rely on locking context */ |
5284 | if (!io_should_wake(iowq, true)) | |
bda52162 JA |
5285 | return -1; |
5286 | ||
5287 | return autoremove_wake_function(curr, mode, wake_flags, key); | |
5288 | } | |
5289 | ||
2b188cc1 JA |
5290 | /* |
5291 | * Wait until events become available, if we don't already have some. The | |
5292 | * application must reap them itself, as they reside on the shared cq ring. | |
5293 | */ | |
5294 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
5295 | const sigset_t __user *sig, size_t sigsz) | |
5296 | { | |
bda52162 JA |
5297 | struct io_wait_queue iowq = { |
5298 | .wq = { | |
5299 | .private = current, | |
5300 | .func = io_wake_function, | |
5301 | .entry = LIST_HEAD_INIT(iowq.wq.entry), | |
5302 | }, | |
5303 | .ctx = ctx, | |
5304 | .to_wait = min_events, | |
5305 | }; | |
75b28aff | 5306 | struct io_rings *rings = ctx->rings; |
e9ffa5c2 | 5307 | int ret = 0; |
2b188cc1 | 5308 | |
b41e9852 JA |
5309 | do { |
5310 | if (io_cqring_events(ctx, false) >= min_events) | |
5311 | return 0; | |
5312 | if (!current->task_works) | |
5313 | break; | |
5314 | task_work_run(); | |
5315 | } while (1); | |
2b188cc1 JA |
5316 | |
5317 | if (sig) { | |
9e75ad5d AB |
5318 | #ifdef CONFIG_COMPAT |
5319 | if (in_compat_syscall()) | |
5320 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 5321 | sigsz); |
9e75ad5d AB |
5322 | else |
5323 | #endif | |
b772434b | 5324 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 5325 | |
2b188cc1 JA |
5326 | if (ret) |
5327 | return ret; | |
5328 | } | |
5329 | ||
bda52162 | 5330 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
c826bd7a | 5331 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 JA |
5332 | do { |
5333 | prepare_to_wait_exclusive(&ctx->wait, &iowq.wq, | |
5334 | TASK_INTERRUPTIBLE); | |
b41e9852 JA |
5335 | if (current->task_works) |
5336 | task_work_run(); | |
1d7bb1d5 | 5337 | if (io_should_wake(&iowq, false)) |
bda52162 JA |
5338 | break; |
5339 | schedule(); | |
5340 | if (signal_pending(current)) { | |
e9ffa5c2 | 5341 | ret = -EINTR; |
bda52162 JA |
5342 | break; |
5343 | } | |
5344 | } while (1); | |
5345 | finish_wait(&ctx->wait, &iowq.wq); | |
5346 | ||
e9ffa5c2 | 5347 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 5348 | |
75b28aff | 5349 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
5350 | } |
5351 | ||
6b06314c JA |
5352 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
5353 | { | |
5354 | #if defined(CONFIG_UNIX) | |
5355 | if (ctx->ring_sock) { | |
5356 | struct sock *sock = ctx->ring_sock->sk; | |
5357 | struct sk_buff *skb; | |
5358 | ||
5359 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
5360 | kfree_skb(skb); | |
5361 | } | |
5362 | #else | |
5363 | int i; | |
5364 | ||
65e19f54 JA |
5365 | for (i = 0; i < ctx->nr_user_files; i++) { |
5366 | struct file *file; | |
5367 | ||
5368 | file = io_file_from_index(ctx, i); | |
5369 | if (file) | |
5370 | fput(file); | |
5371 | } | |
6b06314c JA |
5372 | #endif |
5373 | } | |
5374 | ||
05f3fb3c JA |
5375 | static void io_file_ref_kill(struct percpu_ref *ref) |
5376 | { | |
5377 | struct fixed_file_data *data; | |
5378 | ||
5379 | data = container_of(ref, struct fixed_file_data, refs); | |
5380 | complete(&data->done); | |
5381 | } | |
5382 | ||
6b06314c JA |
5383 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) |
5384 | { | |
05f3fb3c | 5385 | struct fixed_file_data *data = ctx->file_data; |
65e19f54 JA |
5386 | unsigned nr_tables, i; |
5387 | ||
05f3fb3c | 5388 | if (!data) |
6b06314c JA |
5389 | return -ENXIO; |
5390 | ||
05f3fb3c | 5391 | percpu_ref_kill_and_confirm(&data->refs, io_file_ref_kill); |
e46a7950 | 5392 | flush_work(&data->ref_work); |
2faf852d JA |
5393 | wait_for_completion(&data->done); |
5394 | io_ring_file_ref_flush(data); | |
05f3fb3c JA |
5395 | percpu_ref_exit(&data->refs); |
5396 | ||
6b06314c | 5397 | __io_sqe_files_unregister(ctx); |
65e19f54 JA |
5398 | nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE); |
5399 | for (i = 0; i < nr_tables; i++) | |
05f3fb3c JA |
5400 | kfree(data->table[i].files); |
5401 | kfree(data->table); | |
5402 | kfree(data); | |
5403 | ctx->file_data = NULL; | |
6b06314c JA |
5404 | ctx->nr_user_files = 0; |
5405 | return 0; | |
5406 | } | |
5407 | ||
6c271ce2 JA |
5408 | static void io_sq_thread_stop(struct io_ring_ctx *ctx) |
5409 | { | |
5410 | if (ctx->sqo_thread) { | |
206aefde | 5411 | wait_for_completion(&ctx->completions[1]); |
2bbcd6d3 RP |
5412 | /* |
5413 | * The park is a bit of a work-around, without it we get | |
5414 | * warning spews on shutdown with SQPOLL set and affinity | |
5415 | * set to a single CPU. | |
5416 | */ | |
06058632 | 5417 | kthread_park(ctx->sqo_thread); |
6c271ce2 JA |
5418 | kthread_stop(ctx->sqo_thread); |
5419 | ctx->sqo_thread = NULL; | |
5420 | } | |
5421 | } | |
5422 | ||
6b06314c JA |
5423 | static void io_finish_async(struct io_ring_ctx *ctx) |
5424 | { | |
6c271ce2 JA |
5425 | io_sq_thread_stop(ctx); |
5426 | ||
561fb04a JA |
5427 | if (ctx->io_wq) { |
5428 | io_wq_destroy(ctx->io_wq); | |
5429 | ctx->io_wq = NULL; | |
6b06314c JA |
5430 | } |
5431 | } | |
5432 | ||
5433 | #if defined(CONFIG_UNIX) | |
6b06314c JA |
5434 | /* |
5435 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
5436 | * the io_uring can be safely unregistered on process exit, even if we have | |
5437 | * loops in the file referencing. | |
5438 | */ | |
5439 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
5440 | { | |
5441 | struct sock *sk = ctx->ring_sock->sk; | |
5442 | struct scm_fp_list *fpl; | |
5443 | struct sk_buff *skb; | |
08a45173 | 5444 | int i, nr_files; |
6b06314c JA |
5445 | |
5446 | if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) { | |
5447 | unsigned long inflight = ctx->user->unix_inflight + nr; | |
5448 | ||
5449 | if (inflight > task_rlimit(current, RLIMIT_NOFILE)) | |
5450 | return -EMFILE; | |
5451 | } | |
5452 | ||
5453 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); | |
5454 | if (!fpl) | |
5455 | return -ENOMEM; | |
5456 | ||
5457 | skb = alloc_skb(0, GFP_KERNEL); | |
5458 | if (!skb) { | |
5459 | kfree(fpl); | |
5460 | return -ENOMEM; | |
5461 | } | |
5462 | ||
5463 | skb->sk = sk; | |
6b06314c | 5464 | |
08a45173 | 5465 | nr_files = 0; |
6b06314c JA |
5466 | fpl->user = get_uid(ctx->user); |
5467 | for (i = 0; i < nr; i++) { | |
65e19f54 JA |
5468 | struct file *file = io_file_from_index(ctx, i + offset); |
5469 | ||
5470 | if (!file) | |
08a45173 | 5471 | continue; |
65e19f54 | 5472 | fpl->fp[nr_files] = get_file(file); |
08a45173 JA |
5473 | unix_inflight(fpl->user, fpl->fp[nr_files]); |
5474 | nr_files++; | |
6b06314c JA |
5475 | } |
5476 | ||
08a45173 JA |
5477 | if (nr_files) { |
5478 | fpl->max = SCM_MAX_FD; | |
5479 | fpl->count = nr_files; | |
5480 | UNIXCB(skb).fp = fpl; | |
05f3fb3c | 5481 | skb->destructor = unix_destruct_scm; |
08a45173 JA |
5482 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); |
5483 | skb_queue_head(&sk->sk_receive_queue, skb); | |
6b06314c | 5484 | |
08a45173 JA |
5485 | for (i = 0; i < nr_files; i++) |
5486 | fput(fpl->fp[i]); | |
5487 | } else { | |
5488 | kfree_skb(skb); | |
5489 | kfree(fpl); | |
5490 | } | |
6b06314c JA |
5491 | |
5492 | return 0; | |
5493 | } | |
5494 | ||
5495 | /* | |
5496 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
5497 | * causes regular reference counting to break down. We rely on the UNIX | |
5498 | * garbage collection to take care of this problem for us. | |
5499 | */ | |
5500 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
5501 | { | |
5502 | unsigned left, total; | |
5503 | int ret = 0; | |
5504 | ||
5505 | total = 0; | |
5506 | left = ctx->nr_user_files; | |
5507 | while (left) { | |
5508 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
5509 | |
5510 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
5511 | if (ret) | |
5512 | break; | |
5513 | left -= this_files; | |
5514 | total += this_files; | |
5515 | } | |
5516 | ||
5517 | if (!ret) | |
5518 | return 0; | |
5519 | ||
5520 | while (total < ctx->nr_user_files) { | |
65e19f54 JA |
5521 | struct file *file = io_file_from_index(ctx, total); |
5522 | ||
5523 | if (file) | |
5524 | fput(file); | |
6b06314c JA |
5525 | total++; |
5526 | } | |
5527 | ||
5528 | return ret; | |
5529 | } | |
5530 | #else | |
5531 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
5532 | { | |
5533 | return 0; | |
5534 | } | |
5535 | #endif | |
5536 | ||
65e19f54 JA |
5537 | static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables, |
5538 | unsigned nr_files) | |
5539 | { | |
5540 | int i; | |
5541 | ||
5542 | for (i = 0; i < nr_tables; i++) { | |
05f3fb3c | 5543 | struct fixed_file_table *table = &ctx->file_data->table[i]; |
65e19f54 JA |
5544 | unsigned this_files; |
5545 | ||
5546 | this_files = min(nr_files, IORING_MAX_FILES_TABLE); | |
5547 | table->files = kcalloc(this_files, sizeof(struct file *), | |
5548 | GFP_KERNEL); | |
5549 | if (!table->files) | |
5550 | break; | |
5551 | nr_files -= this_files; | |
5552 | } | |
5553 | ||
5554 | if (i == nr_tables) | |
5555 | return 0; | |
5556 | ||
5557 | for (i = 0; i < nr_tables; i++) { | |
05f3fb3c | 5558 | struct fixed_file_table *table = &ctx->file_data->table[i]; |
65e19f54 JA |
5559 | kfree(table->files); |
5560 | } | |
5561 | return 1; | |
5562 | } | |
5563 | ||
05f3fb3c JA |
5564 | static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file) |
5565 | { | |
5566 | #if defined(CONFIG_UNIX) | |
5567 | struct sock *sock = ctx->ring_sock->sk; | |
5568 | struct sk_buff_head list, *head = &sock->sk_receive_queue; | |
5569 | struct sk_buff *skb; | |
5570 | int i; | |
5571 | ||
5572 | __skb_queue_head_init(&list); | |
5573 | ||
5574 | /* | |
5575 | * Find the skb that holds this file in its SCM_RIGHTS. When found, | |
5576 | * remove this entry and rearrange the file array. | |
5577 | */ | |
5578 | skb = skb_dequeue(head); | |
5579 | while (skb) { | |
5580 | struct scm_fp_list *fp; | |
5581 | ||
5582 | fp = UNIXCB(skb).fp; | |
5583 | for (i = 0; i < fp->count; i++) { | |
5584 | int left; | |
5585 | ||
5586 | if (fp->fp[i] != file) | |
5587 | continue; | |
5588 | ||
5589 | unix_notinflight(fp->user, fp->fp[i]); | |
5590 | left = fp->count - 1 - i; | |
5591 | if (left) { | |
5592 | memmove(&fp->fp[i], &fp->fp[i + 1], | |
5593 | left * sizeof(struct file *)); | |
5594 | } | |
5595 | fp->count--; | |
5596 | if (!fp->count) { | |
5597 | kfree_skb(skb); | |
5598 | skb = NULL; | |
5599 | } else { | |
5600 | __skb_queue_tail(&list, skb); | |
5601 | } | |
5602 | fput(file); | |
5603 | file = NULL; | |
5604 | break; | |
5605 | } | |
5606 | ||
5607 | if (!file) | |
5608 | break; | |
5609 | ||
5610 | __skb_queue_tail(&list, skb); | |
5611 | ||
5612 | skb = skb_dequeue(head); | |
5613 | } | |
5614 | ||
5615 | if (skb_peek(&list)) { | |
5616 | spin_lock_irq(&head->lock); | |
5617 | while ((skb = __skb_dequeue(&list)) != NULL) | |
5618 | __skb_queue_tail(head, skb); | |
5619 | spin_unlock_irq(&head->lock); | |
5620 | } | |
5621 | #else | |
5622 | fput(file); | |
5623 | #endif | |
5624 | } | |
5625 | ||
5626 | struct io_file_put { | |
5627 | struct llist_node llist; | |
5628 | struct file *file; | |
5629 | struct completion *done; | |
5630 | }; | |
5631 | ||
2faf852d | 5632 | static void io_ring_file_ref_flush(struct fixed_file_data *data) |
65e19f54 | 5633 | { |
05f3fb3c | 5634 | struct io_file_put *pfile, *tmp; |
05f3fb3c | 5635 | struct llist_node *node; |
65e19f54 | 5636 | |
05f3fb3c JA |
5637 | while ((node = llist_del_all(&data->put_llist)) != NULL) { |
5638 | llist_for_each_entry_safe(pfile, tmp, node, llist) { | |
5639 | io_ring_file_put(data->ctx, pfile->file); | |
5640 | if (pfile->done) | |
5641 | complete(pfile->done); | |
5642 | else | |
5643 | kfree(pfile); | |
5644 | } | |
65e19f54 | 5645 | } |
2faf852d | 5646 | } |
65e19f54 | 5647 | |
2faf852d JA |
5648 | static void io_ring_file_ref_switch(struct work_struct *work) |
5649 | { | |
5650 | struct fixed_file_data *data; | |
65e19f54 | 5651 | |
2faf852d JA |
5652 | data = container_of(work, struct fixed_file_data, ref_work); |
5653 | io_ring_file_ref_flush(data); | |
05f3fb3c JA |
5654 | percpu_ref_switch_to_percpu(&data->refs); |
5655 | } | |
65e19f54 | 5656 | |
05f3fb3c JA |
5657 | static void io_file_data_ref_zero(struct percpu_ref *ref) |
5658 | { | |
5659 | struct fixed_file_data *data; | |
5660 | ||
5661 | data = container_of(ref, struct fixed_file_data, refs); | |
5662 | ||
2faf852d JA |
5663 | /* |
5664 | * We can't safely switch from inside this context, punt to wq. If | |
5665 | * the table ref is going away, the table is being unregistered. | |
5666 | * Don't queue up the async work for that case, the caller will | |
5667 | * handle it. | |
5668 | */ | |
5669 | if (!percpu_ref_is_dying(&data->refs)) | |
5670 | queue_work(system_wq, &data->ref_work); | |
65e19f54 JA |
5671 | } |
5672 | ||
6b06314c JA |
5673 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, |
5674 | unsigned nr_args) | |
5675 | { | |
5676 | __s32 __user *fds = (__s32 __user *) arg; | |
65e19f54 | 5677 | unsigned nr_tables; |
05f3fb3c | 5678 | struct file *file; |
6b06314c JA |
5679 | int fd, ret = 0; |
5680 | unsigned i; | |
5681 | ||
05f3fb3c | 5682 | if (ctx->file_data) |
6b06314c JA |
5683 | return -EBUSY; |
5684 | if (!nr_args) | |
5685 | return -EINVAL; | |
5686 | if (nr_args > IORING_MAX_FIXED_FILES) | |
5687 | return -EMFILE; | |
5688 | ||
05f3fb3c JA |
5689 | ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL); |
5690 | if (!ctx->file_data) | |
5691 | return -ENOMEM; | |
5692 | ctx->file_data->ctx = ctx; | |
5693 | init_completion(&ctx->file_data->done); | |
5694 | ||
65e19f54 | 5695 | nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE); |
05f3fb3c JA |
5696 | ctx->file_data->table = kcalloc(nr_tables, |
5697 | sizeof(struct fixed_file_table), | |
65e19f54 | 5698 | GFP_KERNEL); |
05f3fb3c JA |
5699 | if (!ctx->file_data->table) { |
5700 | kfree(ctx->file_data); | |
5701 | ctx->file_data = NULL; | |
6b06314c | 5702 | return -ENOMEM; |
05f3fb3c JA |
5703 | } |
5704 | ||
5705 | if (percpu_ref_init(&ctx->file_data->refs, io_file_data_ref_zero, | |
5706 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) { | |
5707 | kfree(ctx->file_data->table); | |
5708 | kfree(ctx->file_data); | |
5709 | ctx->file_data = NULL; | |
6b06314c | 5710 | return -ENOMEM; |
05f3fb3c JA |
5711 | } |
5712 | ctx->file_data->put_llist.first = NULL; | |
5713 | INIT_WORK(&ctx->file_data->ref_work, io_ring_file_ref_switch); | |
6b06314c | 5714 | |
65e19f54 | 5715 | if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) { |
05f3fb3c JA |
5716 | percpu_ref_exit(&ctx->file_data->refs); |
5717 | kfree(ctx->file_data->table); | |
5718 | kfree(ctx->file_data); | |
5719 | ctx->file_data = NULL; | |
65e19f54 JA |
5720 | return -ENOMEM; |
5721 | } | |
5722 | ||
08a45173 | 5723 | for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { |
65e19f54 JA |
5724 | struct fixed_file_table *table; |
5725 | unsigned index; | |
5726 | ||
6b06314c JA |
5727 | ret = -EFAULT; |
5728 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) | |
5729 | break; | |
08a45173 JA |
5730 | /* allow sparse sets */ |
5731 | if (fd == -1) { | |
5732 | ret = 0; | |
5733 | continue; | |
5734 | } | |
6b06314c | 5735 | |
05f3fb3c | 5736 | table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT]; |
65e19f54 | 5737 | index = i & IORING_FILE_TABLE_MASK; |
05f3fb3c | 5738 | file = fget(fd); |
6b06314c JA |
5739 | |
5740 | ret = -EBADF; | |
05f3fb3c | 5741 | if (!file) |
6b06314c | 5742 | break; |
05f3fb3c | 5743 | |
6b06314c JA |
5744 | /* |
5745 | * Don't allow io_uring instances to be registered. If UNIX | |
5746 | * isn't enabled, then this causes a reference cycle and this | |
5747 | * instance can never get freed. If UNIX is enabled we'll | |
5748 | * handle it just fine, but there's still no point in allowing | |
5749 | * a ring fd as it doesn't support regular read/write anyway. | |
5750 | */ | |
05f3fb3c JA |
5751 | if (file->f_op == &io_uring_fops) { |
5752 | fput(file); | |
6b06314c JA |
5753 | break; |
5754 | } | |
6b06314c | 5755 | ret = 0; |
05f3fb3c | 5756 | table->files[index] = file; |
6b06314c JA |
5757 | } |
5758 | ||
5759 | if (ret) { | |
65e19f54 | 5760 | for (i = 0; i < ctx->nr_user_files; i++) { |
65e19f54 JA |
5761 | file = io_file_from_index(ctx, i); |
5762 | if (file) | |
5763 | fput(file); | |
5764 | } | |
5765 | for (i = 0; i < nr_tables; i++) | |
05f3fb3c | 5766 | kfree(ctx->file_data->table[i].files); |
6b06314c | 5767 | |
05f3fb3c JA |
5768 | kfree(ctx->file_data->table); |
5769 | kfree(ctx->file_data); | |
5770 | ctx->file_data = NULL; | |
6b06314c JA |
5771 | ctx->nr_user_files = 0; |
5772 | return ret; | |
5773 | } | |
5774 | ||
5775 | ret = io_sqe_files_scm(ctx); | |
5776 | if (ret) | |
5777 | io_sqe_files_unregister(ctx); | |
5778 | ||
5779 | return ret; | |
5780 | } | |
5781 | ||
c3a31e60 JA |
5782 | static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file, |
5783 | int index) | |
5784 | { | |
5785 | #if defined(CONFIG_UNIX) | |
5786 | struct sock *sock = ctx->ring_sock->sk; | |
5787 | struct sk_buff_head *head = &sock->sk_receive_queue; | |
5788 | struct sk_buff *skb; | |
5789 | ||
5790 | /* | |
5791 | * See if we can merge this file into an existing skb SCM_RIGHTS | |
5792 | * file set. If there's no room, fall back to allocating a new skb | |
5793 | * and filling it in. | |
5794 | */ | |
5795 | spin_lock_irq(&head->lock); | |
5796 | skb = skb_peek(head); | |
5797 | if (skb) { | |
5798 | struct scm_fp_list *fpl = UNIXCB(skb).fp; | |
5799 | ||
5800 | if (fpl->count < SCM_MAX_FD) { | |
5801 | __skb_unlink(skb, head); | |
5802 | spin_unlock_irq(&head->lock); | |
5803 | fpl->fp[fpl->count] = get_file(file); | |
5804 | unix_inflight(fpl->user, fpl->fp[fpl->count]); | |
5805 | fpl->count++; | |
5806 | spin_lock_irq(&head->lock); | |
5807 | __skb_queue_head(head, skb); | |
5808 | } else { | |
5809 | skb = NULL; | |
5810 | } | |
5811 | } | |
5812 | spin_unlock_irq(&head->lock); | |
5813 | ||
5814 | if (skb) { | |
5815 | fput(file); | |
5816 | return 0; | |
5817 | } | |
5818 | ||
5819 | return __io_sqe_files_scm(ctx, 1, index); | |
5820 | #else | |
5821 | return 0; | |
5822 | #endif | |
5823 | } | |
5824 | ||
05f3fb3c | 5825 | static void io_atomic_switch(struct percpu_ref *ref) |
c3a31e60 | 5826 | { |
05f3fb3c JA |
5827 | struct fixed_file_data *data; |
5828 | ||
dd3db2a3 JA |
5829 | /* |
5830 | * Juggle reference to ensure we hit zero, if needed, so we can | |
5831 | * switch back to percpu mode | |
5832 | */ | |
05f3fb3c | 5833 | data = container_of(ref, struct fixed_file_data, refs); |
dd3db2a3 JA |
5834 | percpu_ref_put(&data->refs); |
5835 | percpu_ref_get(&data->refs); | |
05f3fb3c JA |
5836 | } |
5837 | ||
5838 | static bool io_queue_file_removal(struct fixed_file_data *data, | |
5839 | struct file *file) | |
5840 | { | |
5841 | struct io_file_put *pfile, pfile_stack; | |
5842 | DECLARE_COMPLETION_ONSTACK(done); | |
5843 | ||
5844 | /* | |
5845 | * If we fail allocating the struct we need for doing async reomval | |
5846 | * of this file, just punt to sync and wait for it. | |
5847 | */ | |
5848 | pfile = kzalloc(sizeof(*pfile), GFP_KERNEL); | |
5849 | if (!pfile) { | |
5850 | pfile = &pfile_stack; | |
5851 | pfile->done = &done; | |
5852 | } | |
5853 | ||
5854 | pfile->file = file; | |
5855 | llist_add(&pfile->llist, &data->put_llist); | |
5856 | ||
5857 | if (pfile == &pfile_stack) { | |
dd3db2a3 | 5858 | percpu_ref_switch_to_atomic(&data->refs, io_atomic_switch); |
05f3fb3c JA |
5859 | wait_for_completion(&done); |
5860 | flush_work(&data->ref_work); | |
5861 | return false; | |
5862 | } | |
5863 | ||
5864 | return true; | |
5865 | } | |
5866 | ||
5867 | static int __io_sqe_files_update(struct io_ring_ctx *ctx, | |
5868 | struct io_uring_files_update *up, | |
5869 | unsigned nr_args) | |
5870 | { | |
5871 | struct fixed_file_data *data = ctx->file_data; | |
5872 | bool ref_switch = false; | |
5873 | struct file *file; | |
c3a31e60 JA |
5874 | __s32 __user *fds; |
5875 | int fd, i, err; | |
5876 | __u32 done; | |
5877 | ||
05f3fb3c | 5878 | if (check_add_overflow(up->offset, nr_args, &done)) |
c3a31e60 JA |
5879 | return -EOVERFLOW; |
5880 | if (done > ctx->nr_user_files) | |
5881 | return -EINVAL; | |
5882 | ||
5883 | done = 0; | |
05f3fb3c | 5884 | fds = u64_to_user_ptr(up->fds); |
c3a31e60 | 5885 | while (nr_args) { |
65e19f54 JA |
5886 | struct fixed_file_table *table; |
5887 | unsigned index; | |
5888 | ||
c3a31e60 JA |
5889 | err = 0; |
5890 | if (copy_from_user(&fd, &fds[done], sizeof(fd))) { | |
5891 | err = -EFAULT; | |
5892 | break; | |
5893 | } | |
05f3fb3c JA |
5894 | i = array_index_nospec(up->offset, ctx->nr_user_files); |
5895 | table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT]; | |
65e19f54 JA |
5896 | index = i & IORING_FILE_TABLE_MASK; |
5897 | if (table->files[index]) { | |
05f3fb3c | 5898 | file = io_file_from_index(ctx, index); |
65e19f54 | 5899 | table->files[index] = NULL; |
05f3fb3c JA |
5900 | if (io_queue_file_removal(data, file)) |
5901 | ref_switch = true; | |
c3a31e60 JA |
5902 | } |
5903 | if (fd != -1) { | |
c3a31e60 JA |
5904 | file = fget(fd); |
5905 | if (!file) { | |
5906 | err = -EBADF; | |
5907 | break; | |
5908 | } | |
5909 | /* | |
5910 | * Don't allow io_uring instances to be registered. If | |
5911 | * UNIX isn't enabled, then this causes a reference | |
5912 | * cycle and this instance can never get freed. If UNIX | |
5913 | * is enabled we'll handle it just fine, but there's | |
5914 | * still no point in allowing a ring fd as it doesn't | |
5915 | * support regular read/write anyway. | |
5916 | */ | |
5917 | if (file->f_op == &io_uring_fops) { | |
5918 | fput(file); | |
5919 | err = -EBADF; | |
5920 | break; | |
5921 | } | |
65e19f54 | 5922 | table->files[index] = file; |
c3a31e60 JA |
5923 | err = io_sqe_file_register(ctx, file, i); |
5924 | if (err) | |
5925 | break; | |
5926 | } | |
5927 | nr_args--; | |
5928 | done++; | |
05f3fb3c JA |
5929 | up->offset++; |
5930 | } | |
5931 | ||
dd3db2a3 | 5932 | if (ref_switch) |
05f3fb3c | 5933 | percpu_ref_switch_to_atomic(&data->refs, io_atomic_switch); |
c3a31e60 JA |
5934 | |
5935 | return done ? done : err; | |
5936 | } | |
05f3fb3c JA |
5937 | static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg, |
5938 | unsigned nr_args) | |
5939 | { | |
5940 | struct io_uring_files_update up; | |
5941 | ||
5942 | if (!ctx->file_data) | |
5943 | return -ENXIO; | |
5944 | if (!nr_args) | |
5945 | return -EINVAL; | |
5946 | if (copy_from_user(&up, arg, sizeof(up))) | |
5947 | return -EFAULT; | |
5948 | if (up.resv) | |
5949 | return -EINVAL; | |
5950 | ||
5951 | return __io_sqe_files_update(ctx, &up, nr_args); | |
5952 | } | |
c3a31e60 | 5953 | |
7d723065 JA |
5954 | static void io_put_work(struct io_wq_work *work) |
5955 | { | |
5956 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
5957 | ||
5958 | io_put_req(req); | |
5959 | } | |
5960 | ||
5961 | static void io_get_work(struct io_wq_work *work) | |
5962 | { | |
5963 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
5964 | ||
5965 | refcount_inc(&req->refs); | |
5966 | } | |
5967 | ||
24369c2e PB |
5968 | static int io_init_wq_offload(struct io_ring_ctx *ctx, |
5969 | struct io_uring_params *p) | |
5970 | { | |
5971 | struct io_wq_data data; | |
5972 | struct fd f; | |
5973 | struct io_ring_ctx *ctx_attach; | |
5974 | unsigned int concurrency; | |
5975 | int ret = 0; | |
5976 | ||
5977 | data.user = ctx->user; | |
5978 | data.get_work = io_get_work; | |
5979 | data.put_work = io_put_work; | |
5980 | ||
5981 | if (!(p->flags & IORING_SETUP_ATTACH_WQ)) { | |
5982 | /* Do QD, or 4 * CPUS, whatever is smallest */ | |
5983 | concurrency = min(ctx->sq_entries, 4 * num_online_cpus()); | |
5984 | ||
5985 | ctx->io_wq = io_wq_create(concurrency, &data); | |
5986 | if (IS_ERR(ctx->io_wq)) { | |
5987 | ret = PTR_ERR(ctx->io_wq); | |
5988 | ctx->io_wq = NULL; | |
5989 | } | |
5990 | return ret; | |
5991 | } | |
5992 | ||
5993 | f = fdget(p->wq_fd); | |
5994 | if (!f.file) | |
5995 | return -EBADF; | |
5996 | ||
5997 | if (f.file->f_op != &io_uring_fops) { | |
5998 | ret = -EINVAL; | |
5999 | goto out_fput; | |
6000 | } | |
6001 | ||
6002 | ctx_attach = f.file->private_data; | |
6003 | /* @io_wq is protected by holding the fd */ | |
6004 | if (!io_wq_get(ctx_attach->io_wq, &data)) { | |
6005 | ret = -EINVAL; | |
6006 | goto out_fput; | |
6007 | } | |
6008 | ||
6009 | ctx->io_wq = ctx_attach->io_wq; | |
6010 | out_fput: | |
6011 | fdput(f); | |
6012 | return ret; | |
6013 | } | |
6014 | ||
6c271ce2 JA |
6015 | static int io_sq_offload_start(struct io_ring_ctx *ctx, |
6016 | struct io_uring_params *p) | |
2b188cc1 JA |
6017 | { |
6018 | int ret; | |
6019 | ||
6c271ce2 | 6020 | init_waitqueue_head(&ctx->sqo_wait); |
2b188cc1 JA |
6021 | mmgrab(current->mm); |
6022 | ctx->sqo_mm = current->mm; | |
6023 | ||
6c271ce2 | 6024 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3ec482d1 JA |
6025 | ret = -EPERM; |
6026 | if (!capable(CAP_SYS_ADMIN)) | |
6027 | goto err; | |
6028 | ||
917257da JA |
6029 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
6030 | if (!ctx->sq_thread_idle) | |
6031 | ctx->sq_thread_idle = HZ; | |
6032 | ||
6c271ce2 | 6033 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 6034 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 6035 | |
917257da | 6036 | ret = -EINVAL; |
44a9bd18 JA |
6037 | if (cpu >= nr_cpu_ids) |
6038 | goto err; | |
7889f44d | 6039 | if (!cpu_online(cpu)) |
917257da JA |
6040 | goto err; |
6041 | ||
6c271ce2 JA |
6042 | ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread, |
6043 | ctx, cpu, | |
6044 | "io_uring-sq"); | |
6045 | } else { | |
6046 | ctx->sqo_thread = kthread_create(io_sq_thread, ctx, | |
6047 | "io_uring-sq"); | |
6048 | } | |
6049 | if (IS_ERR(ctx->sqo_thread)) { | |
6050 | ret = PTR_ERR(ctx->sqo_thread); | |
6051 | ctx->sqo_thread = NULL; | |
6052 | goto err; | |
6053 | } | |
6054 | wake_up_process(ctx->sqo_thread); | |
6055 | } else if (p->flags & IORING_SETUP_SQ_AFF) { | |
6056 | /* Can't have SQ_AFF without SQPOLL */ | |
6057 | ret = -EINVAL; | |
6058 | goto err; | |
6059 | } | |
6060 | ||
24369c2e PB |
6061 | ret = io_init_wq_offload(ctx, p); |
6062 | if (ret) | |
2b188cc1 | 6063 | goto err; |
2b188cc1 JA |
6064 | |
6065 | return 0; | |
6066 | err: | |
54a91f3b | 6067 | io_finish_async(ctx); |
2b188cc1 JA |
6068 | mmdrop(ctx->sqo_mm); |
6069 | ctx->sqo_mm = NULL; | |
6070 | return ret; | |
6071 | } | |
6072 | ||
6073 | static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages) | |
6074 | { | |
6075 | atomic_long_sub(nr_pages, &user->locked_vm); | |
6076 | } | |
6077 | ||
6078 | static int io_account_mem(struct user_struct *user, unsigned long nr_pages) | |
6079 | { | |
6080 | unsigned long page_limit, cur_pages, new_pages; | |
6081 | ||
6082 | /* Don't allow more pages than we can safely lock */ | |
6083 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
6084 | ||
6085 | do { | |
6086 | cur_pages = atomic_long_read(&user->locked_vm); | |
6087 | new_pages = cur_pages + nr_pages; | |
6088 | if (new_pages > page_limit) | |
6089 | return -ENOMEM; | |
6090 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
6091 | new_pages) != cur_pages); | |
6092 | ||
6093 | return 0; | |
6094 | } | |
6095 | ||
6096 | static void io_mem_free(void *ptr) | |
6097 | { | |
52e04ef4 MR |
6098 | struct page *page; |
6099 | ||
6100 | if (!ptr) | |
6101 | return; | |
2b188cc1 | 6102 | |
52e04ef4 | 6103 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
6104 | if (put_page_testzero(page)) |
6105 | free_compound_page(page); | |
6106 | } | |
6107 | ||
6108 | static void *io_mem_alloc(size_t size) | |
6109 | { | |
6110 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
6111 | __GFP_NORETRY; | |
6112 | ||
6113 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
6114 | } | |
6115 | ||
75b28aff HV |
6116 | static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries, |
6117 | size_t *sq_offset) | |
6118 | { | |
6119 | struct io_rings *rings; | |
6120 | size_t off, sq_array_size; | |
6121 | ||
6122 | off = struct_size(rings, cqes, cq_entries); | |
6123 | if (off == SIZE_MAX) | |
6124 | return SIZE_MAX; | |
6125 | ||
6126 | #ifdef CONFIG_SMP | |
6127 | off = ALIGN(off, SMP_CACHE_BYTES); | |
6128 | if (off == 0) | |
6129 | return SIZE_MAX; | |
6130 | #endif | |
6131 | ||
6132 | sq_array_size = array_size(sizeof(u32), sq_entries); | |
6133 | if (sq_array_size == SIZE_MAX) | |
6134 | return SIZE_MAX; | |
6135 | ||
6136 | if (check_add_overflow(off, sq_array_size, &off)) | |
6137 | return SIZE_MAX; | |
6138 | ||
6139 | if (sq_offset) | |
6140 | *sq_offset = off; | |
6141 | ||
6142 | return off; | |
6143 | } | |
6144 | ||
2b188cc1 JA |
6145 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) |
6146 | { | |
75b28aff | 6147 | size_t pages; |
2b188cc1 | 6148 | |
75b28aff HV |
6149 | pages = (size_t)1 << get_order( |
6150 | rings_size(sq_entries, cq_entries, NULL)); | |
6151 | pages += (size_t)1 << get_order( | |
6152 | array_size(sizeof(struct io_uring_sqe), sq_entries)); | |
2b188cc1 | 6153 | |
75b28aff | 6154 | return pages; |
2b188cc1 JA |
6155 | } |
6156 | ||
edafccee JA |
6157 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
6158 | { | |
6159 | int i, j; | |
6160 | ||
6161 | if (!ctx->user_bufs) | |
6162 | return -ENXIO; | |
6163 | ||
6164 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
6165 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
6166 | ||
6167 | for (j = 0; j < imu->nr_bvecs; j++) | |
f1f6a7dd | 6168 | unpin_user_page(imu->bvec[j].bv_page); |
edafccee JA |
6169 | |
6170 | if (ctx->account_mem) | |
6171 | io_unaccount_mem(ctx->user, imu->nr_bvecs); | |
d4ef6475 | 6172 | kvfree(imu->bvec); |
edafccee JA |
6173 | imu->nr_bvecs = 0; |
6174 | } | |
6175 | ||
6176 | kfree(ctx->user_bufs); | |
6177 | ctx->user_bufs = NULL; | |
6178 | ctx->nr_user_bufs = 0; | |
6179 | return 0; | |
6180 | } | |
6181 | ||
6182 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
6183 | void __user *arg, unsigned index) | |
6184 | { | |
6185 | struct iovec __user *src; | |
6186 | ||
6187 | #ifdef CONFIG_COMPAT | |
6188 | if (ctx->compat) { | |
6189 | struct compat_iovec __user *ciovs; | |
6190 | struct compat_iovec ciov; | |
6191 | ||
6192 | ciovs = (struct compat_iovec __user *) arg; | |
6193 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
6194 | return -EFAULT; | |
6195 | ||
d55e5f5b | 6196 | dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base); |
edafccee JA |
6197 | dst->iov_len = ciov.iov_len; |
6198 | return 0; | |
6199 | } | |
6200 | #endif | |
6201 | src = (struct iovec __user *) arg; | |
6202 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
6203 | return -EFAULT; | |
6204 | return 0; | |
6205 | } | |
6206 | ||
6207 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, | |
6208 | unsigned nr_args) | |
6209 | { | |
6210 | struct vm_area_struct **vmas = NULL; | |
6211 | struct page **pages = NULL; | |
6212 | int i, j, got_pages = 0; | |
6213 | int ret = -EINVAL; | |
6214 | ||
6215 | if (ctx->user_bufs) | |
6216 | return -EBUSY; | |
6217 | if (!nr_args || nr_args > UIO_MAXIOV) | |
6218 | return -EINVAL; | |
6219 | ||
6220 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
6221 | GFP_KERNEL); | |
6222 | if (!ctx->user_bufs) | |
6223 | return -ENOMEM; | |
6224 | ||
6225 | for (i = 0; i < nr_args; i++) { | |
6226 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
6227 | unsigned long off, start, end, ubuf; | |
6228 | int pret, nr_pages; | |
6229 | struct iovec iov; | |
6230 | size_t size; | |
6231 | ||
6232 | ret = io_copy_iov(ctx, &iov, arg, i); | |
6233 | if (ret) | |
a278682d | 6234 | goto err; |
edafccee JA |
6235 | |
6236 | /* | |
6237 | * Don't impose further limits on the size and buffer | |
6238 | * constraints here, we'll -EINVAL later when IO is | |
6239 | * submitted if they are wrong. | |
6240 | */ | |
6241 | ret = -EFAULT; | |
6242 | if (!iov.iov_base || !iov.iov_len) | |
6243 | goto err; | |
6244 | ||
6245 | /* arbitrary limit, but we need something */ | |
6246 | if (iov.iov_len > SZ_1G) | |
6247 | goto err; | |
6248 | ||
6249 | ubuf = (unsigned long) iov.iov_base; | |
6250 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
6251 | start = ubuf >> PAGE_SHIFT; | |
6252 | nr_pages = end - start; | |
6253 | ||
6254 | if (ctx->account_mem) { | |
6255 | ret = io_account_mem(ctx->user, nr_pages); | |
6256 | if (ret) | |
6257 | goto err; | |
6258 | } | |
6259 | ||
6260 | ret = 0; | |
6261 | if (!pages || nr_pages > got_pages) { | |
6262 | kfree(vmas); | |
6263 | kfree(pages); | |
d4ef6475 | 6264 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), |
edafccee | 6265 | GFP_KERNEL); |
d4ef6475 | 6266 | vmas = kvmalloc_array(nr_pages, |
edafccee JA |
6267 | sizeof(struct vm_area_struct *), |
6268 | GFP_KERNEL); | |
6269 | if (!pages || !vmas) { | |
6270 | ret = -ENOMEM; | |
6271 | if (ctx->account_mem) | |
6272 | io_unaccount_mem(ctx->user, nr_pages); | |
6273 | goto err; | |
6274 | } | |
6275 | got_pages = nr_pages; | |
6276 | } | |
6277 | ||
d4ef6475 | 6278 | imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec), |
edafccee JA |
6279 | GFP_KERNEL); |
6280 | ret = -ENOMEM; | |
6281 | if (!imu->bvec) { | |
6282 | if (ctx->account_mem) | |
6283 | io_unaccount_mem(ctx->user, nr_pages); | |
6284 | goto err; | |
6285 | } | |
6286 | ||
6287 | ret = 0; | |
6288 | down_read(¤t->mm->mmap_sem); | |
2113b05d | 6289 | pret = pin_user_pages(ubuf, nr_pages, |
932f4a63 IW |
6290 | FOLL_WRITE | FOLL_LONGTERM, |
6291 | pages, vmas); | |
edafccee JA |
6292 | if (pret == nr_pages) { |
6293 | /* don't support file backed memory */ | |
6294 | for (j = 0; j < nr_pages; j++) { | |
6295 | struct vm_area_struct *vma = vmas[j]; | |
6296 | ||
6297 | if (vma->vm_file && | |
6298 | !is_file_hugepages(vma->vm_file)) { | |
6299 | ret = -EOPNOTSUPP; | |
6300 | break; | |
6301 | } | |
6302 | } | |
6303 | } else { | |
6304 | ret = pret < 0 ? pret : -EFAULT; | |
6305 | } | |
6306 | up_read(¤t->mm->mmap_sem); | |
6307 | if (ret) { | |
6308 | /* | |
6309 | * if we did partial map, or found file backed vmas, | |
6310 | * release any pages we did get | |
6311 | */ | |
27c4d3a3 | 6312 | if (pret > 0) |
f1f6a7dd | 6313 | unpin_user_pages(pages, pret); |
edafccee JA |
6314 | if (ctx->account_mem) |
6315 | io_unaccount_mem(ctx->user, nr_pages); | |
d4ef6475 | 6316 | kvfree(imu->bvec); |
edafccee JA |
6317 | goto err; |
6318 | } | |
6319 | ||
6320 | off = ubuf & ~PAGE_MASK; | |
6321 | size = iov.iov_len; | |
6322 | for (j = 0; j < nr_pages; j++) { | |
6323 | size_t vec_len; | |
6324 | ||
6325 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
6326 | imu->bvec[j].bv_page = pages[j]; | |
6327 | imu->bvec[j].bv_len = vec_len; | |
6328 | imu->bvec[j].bv_offset = off; | |
6329 | off = 0; | |
6330 | size -= vec_len; | |
6331 | } | |
6332 | /* store original address for later verification */ | |
6333 | imu->ubuf = ubuf; | |
6334 | imu->len = iov.iov_len; | |
6335 | imu->nr_bvecs = nr_pages; | |
6336 | ||
6337 | ctx->nr_user_bufs++; | |
6338 | } | |
d4ef6475 MR |
6339 | kvfree(pages); |
6340 | kvfree(vmas); | |
edafccee JA |
6341 | return 0; |
6342 | err: | |
d4ef6475 MR |
6343 | kvfree(pages); |
6344 | kvfree(vmas); | |
edafccee JA |
6345 | io_sqe_buffer_unregister(ctx); |
6346 | return ret; | |
6347 | } | |
6348 | ||
9b402849 JA |
6349 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
6350 | { | |
6351 | __s32 __user *fds = arg; | |
6352 | int fd; | |
6353 | ||
6354 | if (ctx->cq_ev_fd) | |
6355 | return -EBUSY; | |
6356 | ||
6357 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
6358 | return -EFAULT; | |
6359 | ||
6360 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
6361 | if (IS_ERR(ctx->cq_ev_fd)) { | |
6362 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
6363 | ctx->cq_ev_fd = NULL; | |
6364 | return ret; | |
6365 | } | |
6366 | ||
6367 | return 0; | |
6368 | } | |
6369 | ||
6370 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
6371 | { | |
6372 | if (ctx->cq_ev_fd) { | |
6373 | eventfd_ctx_put(ctx->cq_ev_fd); | |
6374 | ctx->cq_ev_fd = NULL; | |
6375 | return 0; | |
6376 | } | |
6377 | ||
6378 | return -ENXIO; | |
6379 | } | |
6380 | ||
2b188cc1 JA |
6381 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
6382 | { | |
6b06314c | 6383 | io_finish_async(ctx); |
2b188cc1 JA |
6384 | if (ctx->sqo_mm) |
6385 | mmdrop(ctx->sqo_mm); | |
def596e9 JA |
6386 | |
6387 | io_iopoll_reap_events(ctx); | |
edafccee | 6388 | io_sqe_buffer_unregister(ctx); |
6b06314c | 6389 | io_sqe_files_unregister(ctx); |
9b402849 | 6390 | io_eventfd_unregister(ctx); |
41726c9a | 6391 | idr_destroy(&ctx->personality_idr); |
def596e9 | 6392 | |
2b188cc1 | 6393 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
6394 | if (ctx->ring_sock) { |
6395 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 6396 | sock_release(ctx->ring_sock); |
355e8d26 | 6397 | } |
2b188cc1 JA |
6398 | #endif |
6399 | ||
75b28aff | 6400 | io_mem_free(ctx->rings); |
2b188cc1 | 6401 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
6402 | |
6403 | percpu_ref_exit(&ctx->refs); | |
6404 | if (ctx->account_mem) | |
6405 | io_unaccount_mem(ctx->user, | |
6406 | ring_pages(ctx->sq_entries, ctx->cq_entries)); | |
6407 | free_uid(ctx->user); | |
181e448d | 6408 | put_cred(ctx->creds); |
206aefde | 6409 | kfree(ctx->completions); |
78076bb6 | 6410 | kfree(ctx->cancel_hash); |
0ddf92e8 | 6411 | kmem_cache_free(req_cachep, ctx->fallback_req); |
2b188cc1 JA |
6412 | kfree(ctx); |
6413 | } | |
6414 | ||
6415 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
6416 | { | |
6417 | struct io_ring_ctx *ctx = file->private_data; | |
6418 | __poll_t mask = 0; | |
6419 | ||
6420 | poll_wait(file, &ctx->cq_wait, wait); | |
4f7067c3 SB |
6421 | /* |
6422 | * synchronizes with barrier from wq_has_sleeper call in | |
6423 | * io_commit_cqring | |
6424 | */ | |
2b188cc1 | 6425 | smp_rmb(); |
75b28aff HV |
6426 | if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head != |
6427 | ctx->rings->sq_ring_entries) | |
2b188cc1 | 6428 | mask |= EPOLLOUT | EPOLLWRNORM; |
63e5d81f | 6429 | if (io_cqring_events(ctx, false)) |
2b188cc1 JA |
6430 | mask |= EPOLLIN | EPOLLRDNORM; |
6431 | ||
6432 | return mask; | |
6433 | } | |
6434 | ||
6435 | static int io_uring_fasync(int fd, struct file *file, int on) | |
6436 | { | |
6437 | struct io_ring_ctx *ctx = file->private_data; | |
6438 | ||
6439 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
6440 | } | |
6441 | ||
071698e1 JA |
6442 | static int io_remove_personalities(int id, void *p, void *data) |
6443 | { | |
6444 | struct io_ring_ctx *ctx = data; | |
6445 | const struct cred *cred; | |
6446 | ||
6447 | cred = idr_remove(&ctx->personality_idr, id); | |
6448 | if (cred) | |
6449 | put_cred(cred); | |
6450 | return 0; | |
6451 | } | |
6452 | ||
2b188cc1 JA |
6453 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
6454 | { | |
6455 | mutex_lock(&ctx->uring_lock); | |
6456 | percpu_ref_kill(&ctx->refs); | |
6457 | mutex_unlock(&ctx->uring_lock); | |
6458 | ||
df069d80 JA |
6459 | /* |
6460 | * Wait for sq thread to idle, if we have one. It won't spin on new | |
6461 | * work after we've killed the ctx ref above. This is important to do | |
6462 | * before we cancel existing commands, as the thread could otherwise | |
6463 | * be queueing new work post that. If that's work we need to cancel, | |
6464 | * it could cause shutdown to hang. | |
6465 | */ | |
6466 | while (ctx->sqo_thread && !wq_has_sleeper(&ctx->sqo_wait)) | |
6467 | cpu_relax(); | |
6468 | ||
5262f567 | 6469 | io_kill_timeouts(ctx); |
221c5eb2 | 6470 | io_poll_remove_all(ctx); |
561fb04a JA |
6471 | |
6472 | if (ctx->io_wq) | |
6473 | io_wq_cancel_all(ctx->io_wq); | |
6474 | ||
def596e9 | 6475 | io_iopoll_reap_events(ctx); |
15dff286 JA |
6476 | /* if we failed setting up the ctx, we might not have any rings */ |
6477 | if (ctx->rings) | |
6478 | io_cqring_overflow_flush(ctx, true); | |
071698e1 | 6479 | idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx); |
206aefde | 6480 | wait_for_completion(&ctx->completions[0]); |
2b188cc1 JA |
6481 | io_ring_ctx_free(ctx); |
6482 | } | |
6483 | ||
6484 | static int io_uring_release(struct inode *inode, struct file *file) | |
6485 | { | |
6486 | struct io_ring_ctx *ctx = file->private_data; | |
6487 | ||
6488 | file->private_data = NULL; | |
6489 | io_ring_ctx_wait_and_kill(ctx); | |
6490 | return 0; | |
6491 | } | |
6492 | ||
fcb323cc JA |
6493 | static void io_uring_cancel_files(struct io_ring_ctx *ctx, |
6494 | struct files_struct *files) | |
6495 | { | |
6496 | struct io_kiocb *req; | |
6497 | DEFINE_WAIT(wait); | |
6498 | ||
6499 | while (!list_empty_careful(&ctx->inflight_list)) { | |
768134d4 | 6500 | struct io_kiocb *cancel_req = NULL; |
fcb323cc JA |
6501 | |
6502 | spin_lock_irq(&ctx->inflight_lock); | |
6503 | list_for_each_entry(req, &ctx->inflight_list, inflight_entry) { | |
768134d4 JA |
6504 | if (req->work.files != files) |
6505 | continue; | |
6506 | /* req is being completed, ignore */ | |
6507 | if (!refcount_inc_not_zero(&req->refs)) | |
6508 | continue; | |
6509 | cancel_req = req; | |
6510 | break; | |
fcb323cc | 6511 | } |
768134d4 | 6512 | if (cancel_req) |
fcb323cc | 6513 | prepare_to_wait(&ctx->inflight_wait, &wait, |
768134d4 | 6514 | TASK_UNINTERRUPTIBLE); |
fcb323cc JA |
6515 | spin_unlock_irq(&ctx->inflight_lock); |
6516 | ||
768134d4 JA |
6517 | /* We need to keep going until we don't find a matching req */ |
6518 | if (!cancel_req) | |
fcb323cc | 6519 | break; |
2f6d9b9d | 6520 | |
2ca10259 JA |
6521 | if (cancel_req->flags & REQ_F_OVERFLOW) { |
6522 | spin_lock_irq(&ctx->completion_lock); | |
6523 | list_del(&cancel_req->list); | |
6524 | cancel_req->flags &= ~REQ_F_OVERFLOW; | |
6525 | if (list_empty(&ctx->cq_overflow_list)) { | |
6526 | clear_bit(0, &ctx->sq_check_overflow); | |
6527 | clear_bit(0, &ctx->cq_check_overflow); | |
6528 | } | |
6529 | spin_unlock_irq(&ctx->completion_lock); | |
6530 | ||
6531 | WRITE_ONCE(ctx->rings->cq_overflow, | |
6532 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
6533 | ||
6534 | /* | |
6535 | * Put inflight ref and overflow ref. If that's | |
6536 | * all we had, then we're done with this request. | |
6537 | */ | |
6538 | if (refcount_sub_and_test(2, &cancel_req->refs)) { | |
6539 | io_put_req(cancel_req); | |
6540 | continue; | |
6541 | } | |
6542 | } | |
6543 | ||
2f6d9b9d BL |
6544 | io_wq_cancel_work(ctx->io_wq, &cancel_req->work); |
6545 | io_put_req(cancel_req); | |
fcb323cc JA |
6546 | schedule(); |
6547 | } | |
768134d4 | 6548 | finish_wait(&ctx->inflight_wait, &wait); |
fcb323cc JA |
6549 | } |
6550 | ||
6551 | static int io_uring_flush(struct file *file, void *data) | |
6552 | { | |
6553 | struct io_ring_ctx *ctx = file->private_data; | |
6554 | ||
6555 | io_uring_cancel_files(ctx, data); | |
6ab23144 JA |
6556 | |
6557 | /* | |
6558 | * If the task is going away, cancel work it may have pending | |
6559 | */ | |
6560 | if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) | |
6561 | io_wq_cancel_pid(ctx->io_wq, task_pid_vnr(current)); | |
6562 | ||
fcb323cc JA |
6563 | return 0; |
6564 | } | |
6565 | ||
6c5c240e RP |
6566 | static void *io_uring_validate_mmap_request(struct file *file, |
6567 | loff_t pgoff, size_t sz) | |
2b188cc1 | 6568 | { |
2b188cc1 | 6569 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 6570 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
6571 | struct page *page; |
6572 | void *ptr; | |
6573 | ||
6574 | switch (offset) { | |
6575 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
6576 | case IORING_OFF_CQ_RING: |
6577 | ptr = ctx->rings; | |
2b188cc1 JA |
6578 | break; |
6579 | case IORING_OFF_SQES: | |
6580 | ptr = ctx->sq_sqes; | |
6581 | break; | |
2b188cc1 | 6582 | default: |
6c5c240e | 6583 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
6584 | } |
6585 | ||
6586 | page = virt_to_head_page(ptr); | |
a50b854e | 6587 | if (sz > page_size(page)) |
6c5c240e RP |
6588 | return ERR_PTR(-EINVAL); |
6589 | ||
6590 | return ptr; | |
6591 | } | |
6592 | ||
6593 | #ifdef CONFIG_MMU | |
6594 | ||
6595 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
6596 | { | |
6597 | size_t sz = vma->vm_end - vma->vm_start; | |
6598 | unsigned long pfn; | |
6599 | void *ptr; | |
6600 | ||
6601 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
6602 | if (IS_ERR(ptr)) | |
6603 | return PTR_ERR(ptr); | |
2b188cc1 JA |
6604 | |
6605 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
6606 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
6607 | } | |
6608 | ||
6c5c240e RP |
6609 | #else /* !CONFIG_MMU */ |
6610 | ||
6611 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
6612 | { | |
6613 | return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL; | |
6614 | } | |
6615 | ||
6616 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
6617 | { | |
6618 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
6619 | } | |
6620 | ||
6621 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
6622 | unsigned long addr, unsigned long len, | |
6623 | unsigned long pgoff, unsigned long flags) | |
6624 | { | |
6625 | void *ptr; | |
6626 | ||
6627 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
6628 | if (IS_ERR(ptr)) | |
6629 | return PTR_ERR(ptr); | |
6630 | ||
6631 | return (unsigned long) ptr; | |
6632 | } | |
6633 | ||
6634 | #endif /* !CONFIG_MMU */ | |
6635 | ||
2b188cc1 JA |
6636 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
6637 | u32, min_complete, u32, flags, const sigset_t __user *, sig, | |
6638 | size_t, sigsz) | |
6639 | { | |
6640 | struct io_ring_ctx *ctx; | |
6641 | long ret = -EBADF; | |
6642 | int submitted = 0; | |
6643 | struct fd f; | |
6644 | ||
b41e9852 JA |
6645 | if (current->task_works) |
6646 | task_work_run(); | |
6647 | ||
6c271ce2 | 6648 | if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP)) |
2b188cc1 JA |
6649 | return -EINVAL; |
6650 | ||
6651 | f = fdget(fd); | |
6652 | if (!f.file) | |
6653 | return -EBADF; | |
6654 | ||
6655 | ret = -EOPNOTSUPP; | |
6656 | if (f.file->f_op != &io_uring_fops) | |
6657 | goto out_fput; | |
6658 | ||
6659 | ret = -ENXIO; | |
6660 | ctx = f.file->private_data; | |
6661 | if (!percpu_ref_tryget(&ctx->refs)) | |
6662 | goto out_fput; | |
6663 | ||
6c271ce2 JA |
6664 | /* |
6665 | * For SQ polling, the thread will do all submissions and completions. | |
6666 | * Just return the requested submit count, and wake the thread if | |
6667 | * we were asked to. | |
6668 | */ | |
b2a9eada | 6669 | ret = 0; |
6c271ce2 | 6670 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
c1edbf5f JA |
6671 | if (!list_empty_careful(&ctx->cq_overflow_list)) |
6672 | io_cqring_overflow_flush(ctx, false); | |
6c271ce2 JA |
6673 | if (flags & IORING_ENTER_SQ_WAKEUP) |
6674 | wake_up(&ctx->sqo_wait); | |
6675 | submitted = to_submit; | |
b2a9eada | 6676 | } else if (to_submit) { |
ae9428ca | 6677 | struct mm_struct *cur_mm; |
2b188cc1 JA |
6678 | |
6679 | mutex_lock(&ctx->uring_lock); | |
ae9428ca PB |
6680 | /* already have mm, so io_submit_sqes() won't try to grab it */ |
6681 | cur_mm = ctx->sqo_mm; | |
6682 | submitted = io_submit_sqes(ctx, to_submit, f.file, fd, | |
6683 | &cur_mm, false); | |
2b188cc1 | 6684 | mutex_unlock(&ctx->uring_lock); |
7c504e65 PB |
6685 | |
6686 | if (submitted != to_submit) | |
6687 | goto out; | |
2b188cc1 JA |
6688 | } |
6689 | if (flags & IORING_ENTER_GETEVENTS) { | |
def596e9 JA |
6690 | unsigned nr_events = 0; |
6691 | ||
2b188cc1 JA |
6692 | min_complete = min(min_complete, ctx->cq_entries); |
6693 | ||
def596e9 | 6694 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 | 6695 | ret = io_iopoll_check(ctx, &nr_events, min_complete); |
def596e9 JA |
6696 | } else { |
6697 | ret = io_cqring_wait(ctx, min_complete, sig, sigsz); | |
6698 | } | |
2b188cc1 JA |
6699 | } |
6700 | ||
7c504e65 | 6701 | out: |
6805b32e | 6702 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
6703 | out_fput: |
6704 | fdput(f); | |
6705 | return submitted ? submitted : ret; | |
6706 | } | |
6707 | ||
bebdb65e | 6708 | #ifdef CONFIG_PROC_FS |
87ce955b JA |
6709 | static int io_uring_show_cred(int id, void *p, void *data) |
6710 | { | |
6711 | const struct cred *cred = p; | |
6712 | struct seq_file *m = data; | |
6713 | struct user_namespace *uns = seq_user_ns(m); | |
6714 | struct group_info *gi; | |
6715 | kernel_cap_t cap; | |
6716 | unsigned __capi; | |
6717 | int g; | |
6718 | ||
6719 | seq_printf(m, "%5d\n", id); | |
6720 | seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid)); | |
6721 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid)); | |
6722 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid)); | |
6723 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid)); | |
6724 | seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid)); | |
6725 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid)); | |
6726 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid)); | |
6727 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid)); | |
6728 | seq_puts(m, "\n\tGroups:\t"); | |
6729 | gi = cred->group_info; | |
6730 | for (g = 0; g < gi->ngroups; g++) { | |
6731 | seq_put_decimal_ull(m, g ? " " : "", | |
6732 | from_kgid_munged(uns, gi->gid[g])); | |
6733 | } | |
6734 | seq_puts(m, "\n\tCapEff:\t"); | |
6735 | cap = cred->cap_effective; | |
6736 | CAP_FOR_EACH_U32(__capi) | |
6737 | seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8); | |
6738 | seq_putc(m, '\n'); | |
6739 | return 0; | |
6740 | } | |
6741 | ||
6742 | static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m) | |
6743 | { | |
6744 | int i; | |
6745 | ||
6746 | mutex_lock(&ctx->uring_lock); | |
6747 | seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files); | |
6748 | for (i = 0; i < ctx->nr_user_files; i++) { | |
6749 | struct fixed_file_table *table; | |
6750 | struct file *f; | |
6751 | ||
6752 | table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT]; | |
6753 | f = table->files[i & IORING_FILE_TABLE_MASK]; | |
6754 | if (f) | |
6755 | seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname); | |
6756 | else | |
6757 | seq_printf(m, "%5u: <none>\n", i); | |
6758 | } | |
6759 | seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs); | |
6760 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
6761 | struct io_mapped_ubuf *buf = &ctx->user_bufs[i]; | |
6762 | ||
6763 | seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, | |
6764 | (unsigned int) buf->len); | |
6765 | } | |
6766 | if (!idr_is_empty(&ctx->personality_idr)) { | |
6767 | seq_printf(m, "Personalities:\n"); | |
6768 | idr_for_each(&ctx->personality_idr, io_uring_show_cred, m); | |
6769 | } | |
6770 | mutex_unlock(&ctx->uring_lock); | |
6771 | } | |
6772 | ||
6773 | static void io_uring_show_fdinfo(struct seq_file *m, struct file *f) | |
6774 | { | |
6775 | struct io_ring_ctx *ctx = f->private_data; | |
6776 | ||
6777 | if (percpu_ref_tryget(&ctx->refs)) { | |
6778 | __io_uring_show_fdinfo(ctx, m); | |
6779 | percpu_ref_put(&ctx->refs); | |
6780 | } | |
6781 | } | |
bebdb65e | 6782 | #endif |
87ce955b | 6783 | |
2b188cc1 JA |
6784 | static const struct file_operations io_uring_fops = { |
6785 | .release = io_uring_release, | |
fcb323cc | 6786 | .flush = io_uring_flush, |
2b188cc1 | 6787 | .mmap = io_uring_mmap, |
6c5c240e RP |
6788 | #ifndef CONFIG_MMU |
6789 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
6790 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
6791 | #endif | |
2b188cc1 JA |
6792 | .poll = io_uring_poll, |
6793 | .fasync = io_uring_fasync, | |
bebdb65e | 6794 | #ifdef CONFIG_PROC_FS |
87ce955b | 6795 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 6796 | #endif |
2b188cc1 JA |
6797 | }; |
6798 | ||
6799 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
6800 | struct io_uring_params *p) | |
6801 | { | |
75b28aff HV |
6802 | struct io_rings *rings; |
6803 | size_t size, sq_array_offset; | |
2b188cc1 | 6804 | |
75b28aff HV |
6805 | size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset); |
6806 | if (size == SIZE_MAX) | |
6807 | return -EOVERFLOW; | |
6808 | ||
6809 | rings = io_mem_alloc(size); | |
6810 | if (!rings) | |
2b188cc1 JA |
6811 | return -ENOMEM; |
6812 | ||
75b28aff HV |
6813 | ctx->rings = rings; |
6814 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
6815 | rings->sq_ring_mask = p->sq_entries - 1; | |
6816 | rings->cq_ring_mask = p->cq_entries - 1; | |
6817 | rings->sq_ring_entries = p->sq_entries; | |
6818 | rings->cq_ring_entries = p->cq_entries; | |
6819 | ctx->sq_mask = rings->sq_ring_mask; | |
6820 | ctx->cq_mask = rings->cq_ring_mask; | |
6821 | ctx->sq_entries = rings->sq_ring_entries; | |
6822 | ctx->cq_entries = rings->cq_ring_entries; | |
2b188cc1 JA |
6823 | |
6824 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 JA |
6825 | if (size == SIZE_MAX) { |
6826 | io_mem_free(ctx->rings); | |
6827 | ctx->rings = NULL; | |
2b188cc1 | 6828 | return -EOVERFLOW; |
eb065d30 | 6829 | } |
2b188cc1 JA |
6830 | |
6831 | ctx->sq_sqes = io_mem_alloc(size); | |
eb065d30 JA |
6832 | if (!ctx->sq_sqes) { |
6833 | io_mem_free(ctx->rings); | |
6834 | ctx->rings = NULL; | |
2b188cc1 | 6835 | return -ENOMEM; |
eb065d30 | 6836 | } |
2b188cc1 | 6837 | |
2b188cc1 JA |
6838 | return 0; |
6839 | } | |
6840 | ||
6841 | /* | |
6842 | * Allocate an anonymous fd, this is what constitutes the application | |
6843 | * visible backing of an io_uring instance. The application mmaps this | |
6844 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
6845 | * we have to tie this fd to a socket for file garbage collection purposes. | |
6846 | */ | |
6847 | static int io_uring_get_fd(struct io_ring_ctx *ctx) | |
6848 | { | |
6849 | struct file *file; | |
6850 | int ret; | |
6851 | ||
6852 | #if defined(CONFIG_UNIX) | |
6853 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, | |
6854 | &ctx->ring_sock); | |
6855 | if (ret) | |
6856 | return ret; | |
6857 | #endif | |
6858 | ||
6859 | ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
6860 | if (ret < 0) | |
6861 | goto err; | |
6862 | ||
6863 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, | |
6864 | O_RDWR | O_CLOEXEC); | |
6865 | if (IS_ERR(file)) { | |
6866 | put_unused_fd(ret); | |
6867 | ret = PTR_ERR(file); | |
6868 | goto err; | |
6869 | } | |
6870 | ||
6871 | #if defined(CONFIG_UNIX) | |
6872 | ctx->ring_sock->file = file; | |
6873 | #endif | |
6874 | fd_install(ret, file); | |
6875 | return ret; | |
6876 | err: | |
6877 | #if defined(CONFIG_UNIX) | |
6878 | sock_release(ctx->ring_sock); | |
6879 | ctx->ring_sock = NULL; | |
6880 | #endif | |
6881 | return ret; | |
6882 | } | |
6883 | ||
6884 | static int io_uring_create(unsigned entries, struct io_uring_params *p) | |
6885 | { | |
6886 | struct user_struct *user = NULL; | |
6887 | struct io_ring_ctx *ctx; | |
6888 | bool account_mem; | |
6889 | int ret; | |
6890 | ||
8110c1a6 | 6891 | if (!entries) |
2b188cc1 | 6892 | return -EINVAL; |
8110c1a6 JA |
6893 | if (entries > IORING_MAX_ENTRIES) { |
6894 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
6895 | return -EINVAL; | |
6896 | entries = IORING_MAX_ENTRIES; | |
6897 | } | |
2b188cc1 JA |
6898 | |
6899 | /* | |
6900 | * Use twice as many entries for the CQ ring. It's possible for the | |
6901 | * application to drive a higher depth than the size of the SQ ring, | |
6902 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
6903 | * some flexibility in overcommitting a bit. If the application has |
6904 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
6905 | * of CQ ring entries manually. | |
2b188cc1 JA |
6906 | */ |
6907 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
6908 | if (p->flags & IORING_SETUP_CQSIZE) { |
6909 | /* | |
6910 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
6911 | * to a power-of-two, if it isn't already. We do NOT impose | |
6912 | * any cq vs sq ring sizing. | |
6913 | */ | |
8110c1a6 | 6914 | if (p->cq_entries < p->sq_entries) |
33a107f0 | 6915 | return -EINVAL; |
8110c1a6 JA |
6916 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
6917 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
6918 | return -EINVAL; | |
6919 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
6920 | } | |
33a107f0 JA |
6921 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
6922 | } else { | |
6923 | p->cq_entries = 2 * p->sq_entries; | |
6924 | } | |
2b188cc1 JA |
6925 | |
6926 | user = get_uid(current_user()); | |
6927 | account_mem = !capable(CAP_IPC_LOCK); | |
6928 | ||
6929 | if (account_mem) { | |
6930 | ret = io_account_mem(user, | |
6931 | ring_pages(p->sq_entries, p->cq_entries)); | |
6932 | if (ret) { | |
6933 | free_uid(user); | |
6934 | return ret; | |
6935 | } | |
6936 | } | |
6937 | ||
6938 | ctx = io_ring_ctx_alloc(p); | |
6939 | if (!ctx) { | |
6940 | if (account_mem) | |
6941 | io_unaccount_mem(user, ring_pages(p->sq_entries, | |
6942 | p->cq_entries)); | |
6943 | free_uid(user); | |
6944 | return -ENOMEM; | |
6945 | } | |
6946 | ctx->compat = in_compat_syscall(); | |
6947 | ctx->account_mem = account_mem; | |
6948 | ctx->user = user; | |
0b8c0ec7 | 6949 | ctx->creds = get_current_cred(); |
2b188cc1 JA |
6950 | |
6951 | ret = io_allocate_scq_urings(ctx, p); | |
6952 | if (ret) | |
6953 | goto err; | |
6954 | ||
6c271ce2 | 6955 | ret = io_sq_offload_start(ctx, p); |
2b188cc1 JA |
6956 | if (ret) |
6957 | goto err; | |
6958 | ||
2b188cc1 | 6959 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
6960 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
6961 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
6962 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
6963 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
6964 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
6965 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
6966 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
6967 | |
6968 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
6969 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
6970 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
6971 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
6972 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
6973 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
6974 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
ac90f249 | 6975 | |
044c1ab3 JA |
6976 | /* |
6977 | * Install ring fd as the very last thing, so we don't risk someone | |
6978 | * having closed it before we finish setup | |
6979 | */ | |
6980 | ret = io_uring_get_fd(ctx); | |
6981 | if (ret < 0) | |
6982 | goto err; | |
6983 | ||
da8c9690 | 6984 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
cccf0ee8 JA |
6985 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | |
6986 | IORING_FEAT_CUR_PERSONALITY; | |
c826bd7a | 6987 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
6988 | return ret; |
6989 | err: | |
6990 | io_ring_ctx_wait_and_kill(ctx); | |
6991 | return ret; | |
6992 | } | |
6993 | ||
6994 | /* | |
6995 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
6996 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
6997 | * params structure passed in. | |
6998 | */ | |
6999 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
7000 | { | |
7001 | struct io_uring_params p; | |
7002 | long ret; | |
7003 | int i; | |
7004 | ||
7005 | if (copy_from_user(&p, params, sizeof(p))) | |
7006 | return -EFAULT; | |
7007 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
7008 | if (p.resv[i]) | |
7009 | return -EINVAL; | |
7010 | } | |
7011 | ||
6c271ce2 | 7012 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 7013 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
24369c2e | 7014 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ)) |
2b188cc1 JA |
7015 | return -EINVAL; |
7016 | ||
7017 | ret = io_uring_create(entries, &p); | |
7018 | if (ret < 0) | |
7019 | return ret; | |
7020 | ||
7021 | if (copy_to_user(params, &p, sizeof(p))) | |
7022 | return -EFAULT; | |
7023 | ||
7024 | return ret; | |
7025 | } | |
7026 | ||
7027 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
7028 | struct io_uring_params __user *, params) | |
7029 | { | |
7030 | return io_uring_setup(entries, params); | |
7031 | } | |
7032 | ||
66f4af93 JA |
7033 | static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args) |
7034 | { | |
7035 | struct io_uring_probe *p; | |
7036 | size_t size; | |
7037 | int i, ret; | |
7038 | ||
7039 | size = struct_size(p, ops, nr_args); | |
7040 | if (size == SIZE_MAX) | |
7041 | return -EOVERFLOW; | |
7042 | p = kzalloc(size, GFP_KERNEL); | |
7043 | if (!p) | |
7044 | return -ENOMEM; | |
7045 | ||
7046 | ret = -EFAULT; | |
7047 | if (copy_from_user(p, arg, size)) | |
7048 | goto out; | |
7049 | ret = -EINVAL; | |
7050 | if (memchr_inv(p, 0, size)) | |
7051 | goto out; | |
7052 | ||
7053 | p->last_op = IORING_OP_LAST - 1; | |
7054 | if (nr_args > IORING_OP_LAST) | |
7055 | nr_args = IORING_OP_LAST; | |
7056 | ||
7057 | for (i = 0; i < nr_args; i++) { | |
7058 | p->ops[i].op = i; | |
7059 | if (!io_op_defs[i].not_supported) | |
7060 | p->ops[i].flags = IO_URING_OP_SUPPORTED; | |
7061 | } | |
7062 | p->ops_len = i; | |
7063 | ||
7064 | ret = 0; | |
7065 | if (copy_to_user(arg, p, size)) | |
7066 | ret = -EFAULT; | |
7067 | out: | |
7068 | kfree(p); | |
7069 | return ret; | |
7070 | } | |
7071 | ||
071698e1 JA |
7072 | static int io_register_personality(struct io_ring_ctx *ctx) |
7073 | { | |
7074 | const struct cred *creds = get_current_cred(); | |
7075 | int id; | |
7076 | ||
7077 | id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1, | |
7078 | USHRT_MAX, GFP_KERNEL); | |
7079 | if (id < 0) | |
7080 | put_cred(creds); | |
7081 | return id; | |
7082 | } | |
7083 | ||
7084 | static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) | |
7085 | { | |
7086 | const struct cred *old_creds; | |
7087 | ||
7088 | old_creds = idr_remove(&ctx->personality_idr, id); | |
7089 | if (old_creds) { | |
7090 | put_cred(old_creds); | |
7091 | return 0; | |
7092 | } | |
7093 | ||
7094 | return -EINVAL; | |
7095 | } | |
7096 | ||
7097 | static bool io_register_op_must_quiesce(int op) | |
7098 | { | |
7099 | switch (op) { | |
7100 | case IORING_UNREGISTER_FILES: | |
7101 | case IORING_REGISTER_FILES_UPDATE: | |
7102 | case IORING_REGISTER_PROBE: | |
7103 | case IORING_REGISTER_PERSONALITY: | |
7104 | case IORING_UNREGISTER_PERSONALITY: | |
7105 | return false; | |
7106 | default: | |
7107 | return true; | |
7108 | } | |
7109 | } | |
7110 | ||
edafccee JA |
7111 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
7112 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
7113 | __releases(ctx->uring_lock) |
7114 | __acquires(ctx->uring_lock) | |
edafccee JA |
7115 | { |
7116 | int ret; | |
7117 | ||
35fa71a0 JA |
7118 | /* |
7119 | * We're inside the ring mutex, if the ref is already dying, then | |
7120 | * someone else killed the ctx or is already going through | |
7121 | * io_uring_register(). | |
7122 | */ | |
7123 | if (percpu_ref_is_dying(&ctx->refs)) | |
7124 | return -ENXIO; | |
7125 | ||
071698e1 | 7126 | if (io_register_op_must_quiesce(opcode)) { |
05f3fb3c | 7127 | percpu_ref_kill(&ctx->refs); |
b19062a5 | 7128 | |
05f3fb3c JA |
7129 | /* |
7130 | * Drop uring mutex before waiting for references to exit. If | |
7131 | * another thread is currently inside io_uring_enter() it might | |
7132 | * need to grab the uring_lock to make progress. If we hold it | |
7133 | * here across the drain wait, then we can deadlock. It's safe | |
7134 | * to drop the mutex here, since no new references will come in | |
7135 | * after we've killed the percpu ref. | |
7136 | */ | |
7137 | mutex_unlock(&ctx->uring_lock); | |
c150368b | 7138 | ret = wait_for_completion_interruptible(&ctx->completions[0]); |
05f3fb3c | 7139 | mutex_lock(&ctx->uring_lock); |
c150368b JA |
7140 | if (ret) { |
7141 | percpu_ref_resurrect(&ctx->refs); | |
7142 | ret = -EINTR; | |
7143 | goto out; | |
7144 | } | |
05f3fb3c | 7145 | } |
edafccee JA |
7146 | |
7147 | switch (opcode) { | |
7148 | case IORING_REGISTER_BUFFERS: | |
7149 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
7150 | break; | |
7151 | case IORING_UNREGISTER_BUFFERS: | |
7152 | ret = -EINVAL; | |
7153 | if (arg || nr_args) | |
7154 | break; | |
7155 | ret = io_sqe_buffer_unregister(ctx); | |
7156 | break; | |
6b06314c JA |
7157 | case IORING_REGISTER_FILES: |
7158 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
7159 | break; | |
7160 | case IORING_UNREGISTER_FILES: | |
7161 | ret = -EINVAL; | |
7162 | if (arg || nr_args) | |
7163 | break; | |
7164 | ret = io_sqe_files_unregister(ctx); | |
7165 | break; | |
c3a31e60 JA |
7166 | case IORING_REGISTER_FILES_UPDATE: |
7167 | ret = io_sqe_files_update(ctx, arg, nr_args); | |
7168 | break; | |
9b402849 | 7169 | case IORING_REGISTER_EVENTFD: |
f2842ab5 | 7170 | case IORING_REGISTER_EVENTFD_ASYNC: |
9b402849 JA |
7171 | ret = -EINVAL; |
7172 | if (nr_args != 1) | |
7173 | break; | |
7174 | ret = io_eventfd_register(ctx, arg); | |
f2842ab5 JA |
7175 | if (ret) |
7176 | break; | |
7177 | if (opcode == IORING_REGISTER_EVENTFD_ASYNC) | |
7178 | ctx->eventfd_async = 1; | |
7179 | else | |
7180 | ctx->eventfd_async = 0; | |
9b402849 JA |
7181 | break; |
7182 | case IORING_UNREGISTER_EVENTFD: | |
7183 | ret = -EINVAL; | |
7184 | if (arg || nr_args) | |
7185 | break; | |
7186 | ret = io_eventfd_unregister(ctx); | |
7187 | break; | |
66f4af93 JA |
7188 | case IORING_REGISTER_PROBE: |
7189 | ret = -EINVAL; | |
7190 | if (!arg || nr_args > 256) | |
7191 | break; | |
7192 | ret = io_probe(ctx, arg, nr_args); | |
7193 | break; | |
071698e1 JA |
7194 | case IORING_REGISTER_PERSONALITY: |
7195 | ret = -EINVAL; | |
7196 | if (arg || nr_args) | |
7197 | break; | |
7198 | ret = io_register_personality(ctx); | |
7199 | break; | |
7200 | case IORING_UNREGISTER_PERSONALITY: | |
7201 | ret = -EINVAL; | |
7202 | if (arg) | |
7203 | break; | |
7204 | ret = io_unregister_personality(ctx, nr_args); | |
7205 | break; | |
edafccee JA |
7206 | default: |
7207 | ret = -EINVAL; | |
7208 | break; | |
7209 | } | |
7210 | ||
071698e1 | 7211 | if (io_register_op_must_quiesce(opcode)) { |
05f3fb3c | 7212 | /* bring the ctx back to life */ |
05f3fb3c | 7213 | percpu_ref_reinit(&ctx->refs); |
c150368b JA |
7214 | out: |
7215 | reinit_completion(&ctx->completions[0]); | |
05f3fb3c | 7216 | } |
edafccee JA |
7217 | return ret; |
7218 | } | |
7219 | ||
7220 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
7221 | void __user *, arg, unsigned int, nr_args) | |
7222 | { | |
7223 | struct io_ring_ctx *ctx; | |
7224 | long ret = -EBADF; | |
7225 | struct fd f; | |
7226 | ||
7227 | f = fdget(fd); | |
7228 | if (!f.file) | |
7229 | return -EBADF; | |
7230 | ||
7231 | ret = -EOPNOTSUPP; | |
7232 | if (f.file->f_op != &io_uring_fops) | |
7233 | goto out_fput; | |
7234 | ||
7235 | ctx = f.file->private_data; | |
7236 | ||
7237 | mutex_lock(&ctx->uring_lock); | |
7238 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
7239 | mutex_unlock(&ctx->uring_lock); | |
c826bd7a DD |
7240 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, |
7241 | ctx->cq_ev_fd != NULL, ret); | |
edafccee JA |
7242 | out_fput: |
7243 | fdput(f); | |
7244 | return ret; | |
7245 | } | |
7246 | ||
2b188cc1 JA |
7247 | static int __init io_uring_init(void) |
7248 | { | |
d7f62e82 SM |
7249 | #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \ |
7250 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ | |
7251 | BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \ | |
7252 | } while (0) | |
7253 | ||
7254 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
7255 | __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename) | |
7256 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); | |
7257 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
7258 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
7259 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
7260 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
7261 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
7262 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
7263 | BUILD_BUG_SQE_ELEM(16, __u64, addr); | |
7d67af2c | 7264 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
7265 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
7266 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
7267 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
7268 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
7269 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
7270 | BUILD_BUG_SQE_ELEM(28, __u16, poll_events); | |
7271 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); | |
7272 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
7273 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
7274 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
7275 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
7276 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
7277 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
7278 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 7279 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
d7f62e82 SM |
7280 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
7281 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
7282 | BUILD_BUG_SQE_ELEM(42, __u16, personality); | |
7d67af2c | 7283 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
d7f62e82 | 7284 | |
d3656344 | 7285 | BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST); |
2b188cc1 JA |
7286 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); |
7287 | return 0; | |
7288 | }; | |
7289 | __initcall(io_uring_init); |