]>
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 | ||
5a2e745d JA |
198 | struct io_buffer { |
199 | struct list_head list; | |
200 | __u64 addr; | |
201 | __s32 len; | |
202 | __u16 bid; | |
203 | }; | |
204 | ||
2b188cc1 JA |
205 | struct io_ring_ctx { |
206 | struct { | |
207 | struct percpu_ref refs; | |
208 | } ____cacheline_aligned_in_smp; | |
209 | ||
210 | struct { | |
211 | unsigned int flags; | |
e1d85334 RD |
212 | unsigned int compat: 1; |
213 | unsigned int account_mem: 1; | |
214 | unsigned int cq_overflow_flushed: 1; | |
215 | unsigned int drain_next: 1; | |
216 | unsigned int eventfd_async: 1; | |
2b188cc1 | 217 | |
75b28aff HV |
218 | /* |
219 | * Ring buffer of indices into array of io_uring_sqe, which is | |
220 | * mmapped by the application using the IORING_OFF_SQES offset. | |
221 | * | |
222 | * This indirection could e.g. be used to assign fixed | |
223 | * io_uring_sqe entries to operations and only submit them to | |
224 | * the queue when needed. | |
225 | * | |
226 | * The kernel modifies neither the indices array nor the entries | |
227 | * array. | |
228 | */ | |
229 | u32 *sq_array; | |
2b188cc1 JA |
230 | unsigned cached_sq_head; |
231 | unsigned sq_entries; | |
232 | unsigned sq_mask; | |
6c271ce2 | 233 | unsigned sq_thread_idle; |
498ccd9e | 234 | unsigned cached_sq_dropped; |
206aefde | 235 | atomic_t cached_cq_overflow; |
ad3eb2c8 | 236 | unsigned long sq_check_overflow; |
de0617e4 JA |
237 | |
238 | struct list_head defer_list; | |
5262f567 | 239 | struct list_head timeout_list; |
1d7bb1d5 | 240 | struct list_head cq_overflow_list; |
fcb323cc JA |
241 | |
242 | wait_queue_head_t inflight_wait; | |
ad3eb2c8 | 243 | struct io_uring_sqe *sq_sqes; |
2b188cc1 JA |
244 | } ____cacheline_aligned_in_smp; |
245 | ||
206aefde JA |
246 | struct io_rings *rings; |
247 | ||
2b188cc1 | 248 | /* IO offload */ |
561fb04a | 249 | struct io_wq *io_wq; |
6c271ce2 | 250 | struct task_struct *sqo_thread; /* if using sq thread polling */ |
2b188cc1 | 251 | struct mm_struct *sqo_mm; |
6c271ce2 | 252 | wait_queue_head_t sqo_wait; |
75b28aff | 253 | |
6b06314c JA |
254 | /* |
255 | * If used, fixed file set. Writers must ensure that ->refs is dead, | |
256 | * readers must ensure that ->refs is alive as long as the file* is | |
257 | * used. Only updated through io_uring_register(2). | |
258 | */ | |
05f3fb3c | 259 | struct fixed_file_data *file_data; |
6b06314c | 260 | unsigned nr_user_files; |
b14cca0c PB |
261 | int ring_fd; |
262 | struct file *ring_file; | |
6b06314c | 263 | |
edafccee JA |
264 | /* if used, fixed mapped user buffers */ |
265 | unsigned nr_user_bufs; | |
266 | struct io_mapped_ubuf *user_bufs; | |
267 | ||
2b188cc1 JA |
268 | struct user_struct *user; |
269 | ||
0b8c0ec7 | 270 | const struct cred *creds; |
181e448d | 271 | |
206aefde JA |
272 | /* 0 is for ctx quiesce/reinit/free, 1 is for sqo_thread started */ |
273 | struct completion *completions; | |
274 | ||
0ddf92e8 JA |
275 | /* if all else fails... */ |
276 | struct io_kiocb *fallback_req; | |
277 | ||
206aefde JA |
278 | #if defined(CONFIG_UNIX) |
279 | struct socket *ring_sock; | |
280 | #endif | |
281 | ||
5a2e745d JA |
282 | struct idr io_buffer_idr; |
283 | ||
071698e1 JA |
284 | struct idr personality_idr; |
285 | ||
206aefde JA |
286 | struct { |
287 | unsigned cached_cq_tail; | |
288 | unsigned cq_entries; | |
289 | unsigned cq_mask; | |
290 | atomic_t cq_timeouts; | |
ad3eb2c8 | 291 | unsigned long cq_check_overflow; |
206aefde JA |
292 | struct wait_queue_head cq_wait; |
293 | struct fasync_struct *cq_fasync; | |
294 | struct eventfd_ctx *cq_ev_fd; | |
295 | } ____cacheline_aligned_in_smp; | |
2b188cc1 JA |
296 | |
297 | struct { | |
298 | struct mutex uring_lock; | |
299 | wait_queue_head_t wait; | |
300 | } ____cacheline_aligned_in_smp; | |
301 | ||
302 | struct { | |
303 | spinlock_t completion_lock; | |
e94f141b | 304 | |
def596e9 JA |
305 | /* |
306 | * ->poll_list is protected by the ctx->uring_lock for | |
307 | * io_uring instances that don't use IORING_SETUP_SQPOLL. | |
308 | * For SQPOLL, only the single threaded io_sq_thread() will | |
309 | * manipulate the list, hence no extra locking is needed there. | |
310 | */ | |
311 | struct list_head poll_list; | |
78076bb6 JA |
312 | struct hlist_head *cancel_hash; |
313 | unsigned cancel_hash_bits; | |
e94f141b | 314 | bool poll_multi_file; |
31b51510 | 315 | |
fcb323cc JA |
316 | spinlock_t inflight_lock; |
317 | struct list_head inflight_list; | |
2b188cc1 | 318 | } ____cacheline_aligned_in_smp; |
2b188cc1 JA |
319 | }; |
320 | ||
09bb8394 JA |
321 | /* |
322 | * First field must be the file pointer in all the | |
323 | * iocb unions! See also 'struct kiocb' in <linux/fs.h> | |
324 | */ | |
221c5eb2 JA |
325 | struct io_poll_iocb { |
326 | struct file *file; | |
0969e783 JA |
327 | union { |
328 | struct wait_queue_head *head; | |
329 | u64 addr; | |
330 | }; | |
221c5eb2 | 331 | __poll_t events; |
8c838788 | 332 | bool done; |
221c5eb2 | 333 | bool canceled; |
392edb45 | 334 | struct wait_queue_entry wait; |
221c5eb2 JA |
335 | }; |
336 | ||
b5dba59e JA |
337 | struct io_close { |
338 | struct file *file; | |
339 | struct file *put_file; | |
340 | int fd; | |
341 | }; | |
342 | ||
ad8a48ac JA |
343 | struct io_timeout_data { |
344 | struct io_kiocb *req; | |
345 | struct hrtimer timer; | |
346 | struct timespec64 ts; | |
347 | enum hrtimer_mode mode; | |
cc42e0ac | 348 | u32 seq_offset; |
ad8a48ac JA |
349 | }; |
350 | ||
8ed8d3c3 JA |
351 | struct io_accept { |
352 | struct file *file; | |
353 | struct sockaddr __user *addr; | |
354 | int __user *addr_len; | |
355 | int flags; | |
356 | }; | |
357 | ||
358 | struct io_sync { | |
359 | struct file *file; | |
360 | loff_t len; | |
361 | loff_t off; | |
362 | int flags; | |
d63d1b5e | 363 | int mode; |
8ed8d3c3 JA |
364 | }; |
365 | ||
fbf23849 JA |
366 | struct io_cancel { |
367 | struct file *file; | |
368 | u64 addr; | |
369 | }; | |
370 | ||
b29472ee JA |
371 | struct io_timeout { |
372 | struct file *file; | |
373 | u64 addr; | |
374 | int flags; | |
26a61679 | 375 | unsigned count; |
b29472ee JA |
376 | }; |
377 | ||
9adbd45d JA |
378 | struct io_rw { |
379 | /* NOTE: kiocb has the file as the first member, so don't do it here */ | |
380 | struct kiocb kiocb; | |
381 | u64 addr; | |
382 | u64 len; | |
383 | }; | |
384 | ||
3fbb51c1 JA |
385 | struct io_connect { |
386 | struct file *file; | |
387 | struct sockaddr __user *addr; | |
388 | int addr_len; | |
389 | }; | |
390 | ||
e47293fd JA |
391 | struct io_sr_msg { |
392 | struct file *file; | |
fddaface JA |
393 | union { |
394 | struct user_msghdr __user *msg; | |
395 | void __user *buf; | |
396 | }; | |
e47293fd | 397 | int msg_flags; |
fddaface | 398 | size_t len; |
e47293fd JA |
399 | }; |
400 | ||
15b71abe JA |
401 | struct io_open { |
402 | struct file *file; | |
403 | int dfd; | |
eddc7ef5 | 404 | union { |
eddc7ef5 JA |
405 | unsigned mask; |
406 | }; | |
15b71abe | 407 | struct filename *filename; |
eddc7ef5 | 408 | struct statx __user *buffer; |
c12cedf2 | 409 | struct open_how how; |
15b71abe JA |
410 | }; |
411 | ||
05f3fb3c JA |
412 | struct io_files_update { |
413 | struct file *file; | |
414 | u64 arg; | |
415 | u32 nr_args; | |
416 | u32 offset; | |
417 | }; | |
418 | ||
4840e418 JA |
419 | struct io_fadvise { |
420 | struct file *file; | |
421 | u64 offset; | |
422 | u32 len; | |
423 | u32 advice; | |
424 | }; | |
425 | ||
c1ca757b JA |
426 | struct io_madvise { |
427 | struct file *file; | |
428 | u64 addr; | |
429 | u32 len; | |
430 | u32 advice; | |
431 | }; | |
432 | ||
3e4827b0 JA |
433 | struct io_epoll { |
434 | struct file *file; | |
435 | int epfd; | |
436 | int op; | |
437 | int fd; | |
438 | struct epoll_event event; | |
e47293fd JA |
439 | }; |
440 | ||
7d67af2c PB |
441 | struct io_splice { |
442 | struct file *file_out; | |
443 | struct file *file_in; | |
444 | loff_t off_out; | |
445 | loff_t off_in; | |
446 | u64 len; | |
447 | unsigned int flags; | |
448 | }; | |
449 | ||
ddf0322d JA |
450 | struct io_provide_buf { |
451 | struct file *file; | |
452 | __u64 addr; | |
453 | __s32 len; | |
454 | __u32 bgid; | |
455 | __u16 nbufs; | |
456 | __u16 bid; | |
457 | }; | |
458 | ||
f499a021 JA |
459 | struct io_async_connect { |
460 | struct sockaddr_storage address; | |
461 | }; | |
462 | ||
03b1230c JA |
463 | struct io_async_msghdr { |
464 | struct iovec fast_iov[UIO_FASTIOV]; | |
465 | struct iovec *iov; | |
466 | struct sockaddr __user *uaddr; | |
467 | struct msghdr msg; | |
b537916c | 468 | struct sockaddr_storage addr; |
03b1230c JA |
469 | }; |
470 | ||
f67676d1 JA |
471 | struct io_async_rw { |
472 | struct iovec fast_iov[UIO_FASTIOV]; | |
473 | struct iovec *iov; | |
474 | ssize_t nr_segs; | |
475 | ssize_t size; | |
476 | }; | |
477 | ||
1a6b74fc | 478 | struct io_async_ctx { |
f67676d1 JA |
479 | union { |
480 | struct io_async_rw rw; | |
03b1230c | 481 | struct io_async_msghdr msg; |
f499a021 | 482 | struct io_async_connect connect; |
2d28390a | 483 | struct io_timeout_data timeout; |
f67676d1 | 484 | }; |
1a6b74fc JA |
485 | }; |
486 | ||
6b47ee6e PB |
487 | enum { |
488 | REQ_F_FIXED_FILE_BIT = IOSQE_FIXED_FILE_BIT, | |
489 | REQ_F_IO_DRAIN_BIT = IOSQE_IO_DRAIN_BIT, | |
490 | REQ_F_LINK_BIT = IOSQE_IO_LINK_BIT, | |
491 | REQ_F_HARDLINK_BIT = IOSQE_IO_HARDLINK_BIT, | |
492 | REQ_F_FORCE_ASYNC_BIT = IOSQE_ASYNC_BIT, | |
493 | ||
494 | REQ_F_LINK_NEXT_BIT, | |
495 | REQ_F_FAIL_LINK_BIT, | |
496 | REQ_F_INFLIGHT_BIT, | |
497 | REQ_F_CUR_POS_BIT, | |
498 | REQ_F_NOWAIT_BIT, | |
499 | REQ_F_IOPOLL_COMPLETED_BIT, | |
500 | REQ_F_LINK_TIMEOUT_BIT, | |
501 | REQ_F_TIMEOUT_BIT, | |
502 | REQ_F_ISREG_BIT, | |
503 | REQ_F_MUST_PUNT_BIT, | |
504 | REQ_F_TIMEOUT_NOSEQ_BIT, | |
505 | REQ_F_COMP_LOCKED_BIT, | |
99bc4c38 | 506 | REQ_F_NEED_CLEANUP_BIT, |
2ca10259 | 507 | REQ_F_OVERFLOW_BIT, |
d7718a9d | 508 | REQ_F_POLLED_BIT, |
6b47ee6e PB |
509 | }; |
510 | ||
511 | enum { | |
512 | /* ctx owns file */ | |
513 | REQ_F_FIXED_FILE = BIT(REQ_F_FIXED_FILE_BIT), | |
514 | /* drain existing IO first */ | |
515 | REQ_F_IO_DRAIN = BIT(REQ_F_IO_DRAIN_BIT), | |
516 | /* linked sqes */ | |
517 | REQ_F_LINK = BIT(REQ_F_LINK_BIT), | |
518 | /* doesn't sever on completion < 0 */ | |
519 | REQ_F_HARDLINK = BIT(REQ_F_HARDLINK_BIT), | |
520 | /* IOSQE_ASYNC */ | |
521 | REQ_F_FORCE_ASYNC = BIT(REQ_F_FORCE_ASYNC_BIT), | |
522 | ||
523 | /* already grabbed next link */ | |
524 | REQ_F_LINK_NEXT = BIT(REQ_F_LINK_NEXT_BIT), | |
525 | /* fail rest of links */ | |
526 | REQ_F_FAIL_LINK = BIT(REQ_F_FAIL_LINK_BIT), | |
527 | /* on inflight list */ | |
528 | REQ_F_INFLIGHT = BIT(REQ_F_INFLIGHT_BIT), | |
529 | /* read/write uses file position */ | |
530 | REQ_F_CUR_POS = BIT(REQ_F_CUR_POS_BIT), | |
531 | /* must not punt to workers */ | |
532 | REQ_F_NOWAIT = BIT(REQ_F_NOWAIT_BIT), | |
533 | /* polled IO has completed */ | |
534 | REQ_F_IOPOLL_COMPLETED = BIT(REQ_F_IOPOLL_COMPLETED_BIT), | |
535 | /* has linked timeout */ | |
536 | REQ_F_LINK_TIMEOUT = BIT(REQ_F_LINK_TIMEOUT_BIT), | |
537 | /* timeout request */ | |
538 | REQ_F_TIMEOUT = BIT(REQ_F_TIMEOUT_BIT), | |
539 | /* regular file */ | |
540 | REQ_F_ISREG = BIT(REQ_F_ISREG_BIT), | |
541 | /* must be punted even for NONBLOCK */ | |
542 | REQ_F_MUST_PUNT = BIT(REQ_F_MUST_PUNT_BIT), | |
543 | /* no timeout sequence */ | |
544 | REQ_F_TIMEOUT_NOSEQ = BIT(REQ_F_TIMEOUT_NOSEQ_BIT), | |
545 | /* completion under lock */ | |
546 | REQ_F_COMP_LOCKED = BIT(REQ_F_COMP_LOCKED_BIT), | |
99bc4c38 PB |
547 | /* needs cleanup */ |
548 | REQ_F_NEED_CLEANUP = BIT(REQ_F_NEED_CLEANUP_BIT), | |
2ca10259 JA |
549 | /* in overflow list */ |
550 | REQ_F_OVERFLOW = BIT(REQ_F_OVERFLOW_BIT), | |
d7718a9d JA |
551 | /* already went through poll handler */ |
552 | REQ_F_POLLED = BIT(REQ_F_POLLED_BIT), | |
553 | }; | |
554 | ||
555 | struct async_poll { | |
556 | struct io_poll_iocb poll; | |
557 | struct io_wq_work work; | |
6b47ee6e PB |
558 | }; |
559 | ||
09bb8394 JA |
560 | /* |
561 | * NOTE! Each of the iocb union members has the file pointer | |
562 | * as the first entry in their struct definition. So you can | |
563 | * access the file pointer through any of the sub-structs, | |
564 | * or directly as just 'ki_filp' in this struct. | |
565 | */ | |
2b188cc1 | 566 | struct io_kiocb { |
221c5eb2 | 567 | union { |
09bb8394 | 568 | struct file *file; |
9adbd45d | 569 | struct io_rw rw; |
221c5eb2 | 570 | struct io_poll_iocb poll; |
8ed8d3c3 JA |
571 | struct io_accept accept; |
572 | struct io_sync sync; | |
fbf23849 | 573 | struct io_cancel cancel; |
b29472ee | 574 | struct io_timeout timeout; |
3fbb51c1 | 575 | struct io_connect connect; |
e47293fd | 576 | struct io_sr_msg sr_msg; |
15b71abe | 577 | struct io_open open; |
b5dba59e | 578 | struct io_close close; |
05f3fb3c | 579 | struct io_files_update files_update; |
4840e418 | 580 | struct io_fadvise fadvise; |
c1ca757b | 581 | struct io_madvise madvise; |
3e4827b0 | 582 | struct io_epoll epoll; |
7d67af2c | 583 | struct io_splice splice; |
ddf0322d | 584 | struct io_provide_buf pbuf; |
221c5eb2 | 585 | }; |
2b188cc1 | 586 | |
1a6b74fc | 587 | struct io_async_ctx *io; |
cf6fd4bd | 588 | bool needs_fixed_file; |
d625c6ee | 589 | u8 opcode; |
2b188cc1 JA |
590 | |
591 | struct io_ring_ctx *ctx; | |
d7718a9d | 592 | struct list_head list; |
2b188cc1 | 593 | unsigned int flags; |
c16361c1 | 594 | refcount_t refs; |
d7718a9d | 595 | struct task_struct *task; |
2b188cc1 | 596 | u64 user_data; |
9e645e11 | 597 | u32 result; |
de0617e4 | 598 | u32 sequence; |
2b188cc1 | 599 | |
d7718a9d JA |
600 | struct list_head link_list; |
601 | ||
fcb323cc JA |
602 | struct list_head inflight_entry; |
603 | ||
b41e9852 JA |
604 | union { |
605 | /* | |
606 | * Only commands that never go async can use the below fields, | |
d7718a9d JA |
607 | * obviously. Right now only IORING_OP_POLL_ADD uses them, and |
608 | * async armed poll handlers for regular commands. The latter | |
609 | * restore the work, if needed. | |
b41e9852 JA |
610 | */ |
611 | struct { | |
b41e9852 | 612 | struct callback_head task_work; |
d7718a9d JA |
613 | struct hlist_node hash_node; |
614 | struct async_poll *apoll; | |
b41e9852 JA |
615 | }; |
616 | struct io_wq_work work; | |
617 | }; | |
2b188cc1 JA |
618 | }; |
619 | ||
620 | #define IO_PLUG_THRESHOLD 2 | |
def596e9 | 621 | #define IO_IOPOLL_BATCH 8 |
2b188cc1 | 622 | |
9a56a232 JA |
623 | struct io_submit_state { |
624 | struct blk_plug plug; | |
625 | ||
2579f913 JA |
626 | /* |
627 | * io_kiocb alloc cache | |
628 | */ | |
629 | void *reqs[IO_IOPOLL_BATCH]; | |
6c8a3134 | 630 | unsigned int free_reqs; |
2579f913 | 631 | |
9a56a232 JA |
632 | /* |
633 | * File reference cache | |
634 | */ | |
635 | struct file *file; | |
636 | unsigned int fd; | |
637 | unsigned int has_refs; | |
638 | unsigned int used_refs; | |
639 | unsigned int ios_left; | |
640 | }; | |
641 | ||
d3656344 JA |
642 | struct io_op_def { |
643 | /* needs req->io allocated for deferral/async */ | |
644 | unsigned async_ctx : 1; | |
645 | /* needs current->mm setup, does mm access */ | |
646 | unsigned needs_mm : 1; | |
647 | /* needs req->file assigned */ | |
648 | unsigned needs_file : 1; | |
649 | /* needs req->file assigned IFF fd is >= 0 */ | |
650 | unsigned fd_non_neg : 1; | |
651 | /* hash wq insertion if file is a regular file */ | |
652 | unsigned hash_reg_file : 1; | |
653 | /* unbound wq insertion if file is a non-regular file */ | |
654 | unsigned unbound_nonreg_file : 1; | |
66f4af93 JA |
655 | /* opcode is not supported by this kernel */ |
656 | unsigned not_supported : 1; | |
f86cd20c JA |
657 | /* needs file table */ |
658 | unsigned file_table : 1; | |
ff002b30 JA |
659 | /* needs ->fs */ |
660 | unsigned needs_fs : 1; | |
8a72758c JA |
661 | /* set if opcode supports polled "wait" */ |
662 | unsigned pollin : 1; | |
663 | unsigned pollout : 1; | |
d3656344 JA |
664 | }; |
665 | ||
666 | static const struct io_op_def io_op_defs[] = { | |
0463b6c5 PB |
667 | [IORING_OP_NOP] = {}, |
668 | [IORING_OP_READV] = { | |
d3656344 JA |
669 | .async_ctx = 1, |
670 | .needs_mm = 1, | |
671 | .needs_file = 1, | |
672 | .unbound_nonreg_file = 1, | |
8a72758c | 673 | .pollin = 1, |
d3656344 | 674 | }, |
0463b6c5 | 675 | [IORING_OP_WRITEV] = { |
d3656344 JA |
676 | .async_ctx = 1, |
677 | .needs_mm = 1, | |
678 | .needs_file = 1, | |
679 | .hash_reg_file = 1, | |
680 | .unbound_nonreg_file = 1, | |
8a72758c | 681 | .pollout = 1, |
d3656344 | 682 | }, |
0463b6c5 | 683 | [IORING_OP_FSYNC] = { |
d3656344 JA |
684 | .needs_file = 1, |
685 | }, | |
0463b6c5 | 686 | [IORING_OP_READ_FIXED] = { |
d3656344 JA |
687 | .needs_file = 1, |
688 | .unbound_nonreg_file = 1, | |
8a72758c | 689 | .pollin = 1, |
d3656344 | 690 | }, |
0463b6c5 | 691 | [IORING_OP_WRITE_FIXED] = { |
d3656344 JA |
692 | .needs_file = 1, |
693 | .hash_reg_file = 1, | |
694 | .unbound_nonreg_file = 1, | |
8a72758c | 695 | .pollout = 1, |
d3656344 | 696 | }, |
0463b6c5 | 697 | [IORING_OP_POLL_ADD] = { |
d3656344 JA |
698 | .needs_file = 1, |
699 | .unbound_nonreg_file = 1, | |
700 | }, | |
0463b6c5 PB |
701 | [IORING_OP_POLL_REMOVE] = {}, |
702 | [IORING_OP_SYNC_FILE_RANGE] = { | |
d3656344 JA |
703 | .needs_file = 1, |
704 | }, | |
0463b6c5 | 705 | [IORING_OP_SENDMSG] = { |
d3656344 JA |
706 | .async_ctx = 1, |
707 | .needs_mm = 1, | |
708 | .needs_file = 1, | |
709 | .unbound_nonreg_file = 1, | |
ff002b30 | 710 | .needs_fs = 1, |
8a72758c | 711 | .pollout = 1, |
d3656344 | 712 | }, |
0463b6c5 | 713 | [IORING_OP_RECVMSG] = { |
d3656344 JA |
714 | .async_ctx = 1, |
715 | .needs_mm = 1, | |
716 | .needs_file = 1, | |
717 | .unbound_nonreg_file = 1, | |
ff002b30 | 718 | .needs_fs = 1, |
8a72758c | 719 | .pollin = 1, |
d3656344 | 720 | }, |
0463b6c5 | 721 | [IORING_OP_TIMEOUT] = { |
d3656344 JA |
722 | .async_ctx = 1, |
723 | .needs_mm = 1, | |
724 | }, | |
0463b6c5 PB |
725 | [IORING_OP_TIMEOUT_REMOVE] = {}, |
726 | [IORING_OP_ACCEPT] = { | |
d3656344 JA |
727 | .needs_mm = 1, |
728 | .needs_file = 1, | |
729 | .unbound_nonreg_file = 1, | |
f86cd20c | 730 | .file_table = 1, |
8a72758c | 731 | .pollin = 1, |
d3656344 | 732 | }, |
0463b6c5 PB |
733 | [IORING_OP_ASYNC_CANCEL] = {}, |
734 | [IORING_OP_LINK_TIMEOUT] = { | |
d3656344 JA |
735 | .async_ctx = 1, |
736 | .needs_mm = 1, | |
737 | }, | |
0463b6c5 | 738 | [IORING_OP_CONNECT] = { |
d3656344 JA |
739 | .async_ctx = 1, |
740 | .needs_mm = 1, | |
741 | .needs_file = 1, | |
742 | .unbound_nonreg_file = 1, | |
8a72758c | 743 | .pollout = 1, |
d3656344 | 744 | }, |
0463b6c5 | 745 | [IORING_OP_FALLOCATE] = { |
d3656344 JA |
746 | .needs_file = 1, |
747 | }, | |
0463b6c5 | 748 | [IORING_OP_OPENAT] = { |
d3656344 JA |
749 | .needs_file = 1, |
750 | .fd_non_neg = 1, | |
f86cd20c | 751 | .file_table = 1, |
ff002b30 | 752 | .needs_fs = 1, |
d3656344 | 753 | }, |
0463b6c5 | 754 | [IORING_OP_CLOSE] = { |
d3656344 | 755 | .needs_file = 1, |
f86cd20c | 756 | .file_table = 1, |
d3656344 | 757 | }, |
0463b6c5 | 758 | [IORING_OP_FILES_UPDATE] = { |
d3656344 | 759 | .needs_mm = 1, |
f86cd20c | 760 | .file_table = 1, |
d3656344 | 761 | }, |
0463b6c5 | 762 | [IORING_OP_STATX] = { |
d3656344 JA |
763 | .needs_mm = 1, |
764 | .needs_file = 1, | |
765 | .fd_non_neg = 1, | |
ff002b30 | 766 | .needs_fs = 1, |
d3656344 | 767 | }, |
0463b6c5 | 768 | [IORING_OP_READ] = { |
3a6820f2 JA |
769 | .needs_mm = 1, |
770 | .needs_file = 1, | |
771 | .unbound_nonreg_file = 1, | |
8a72758c | 772 | .pollin = 1, |
3a6820f2 | 773 | }, |
0463b6c5 | 774 | [IORING_OP_WRITE] = { |
3a6820f2 JA |
775 | .needs_mm = 1, |
776 | .needs_file = 1, | |
777 | .unbound_nonreg_file = 1, | |
8a72758c | 778 | .pollout = 1, |
3a6820f2 | 779 | }, |
0463b6c5 | 780 | [IORING_OP_FADVISE] = { |
4840e418 JA |
781 | .needs_file = 1, |
782 | }, | |
0463b6c5 | 783 | [IORING_OP_MADVISE] = { |
c1ca757b JA |
784 | .needs_mm = 1, |
785 | }, | |
0463b6c5 | 786 | [IORING_OP_SEND] = { |
fddaface JA |
787 | .needs_mm = 1, |
788 | .needs_file = 1, | |
789 | .unbound_nonreg_file = 1, | |
8a72758c | 790 | .pollout = 1, |
fddaface | 791 | }, |
0463b6c5 | 792 | [IORING_OP_RECV] = { |
fddaface JA |
793 | .needs_mm = 1, |
794 | .needs_file = 1, | |
795 | .unbound_nonreg_file = 1, | |
8a72758c | 796 | .pollin = 1, |
fddaface | 797 | }, |
0463b6c5 | 798 | [IORING_OP_OPENAT2] = { |
cebdb986 JA |
799 | .needs_file = 1, |
800 | .fd_non_neg = 1, | |
f86cd20c | 801 | .file_table = 1, |
ff002b30 | 802 | .needs_fs = 1, |
cebdb986 | 803 | }, |
3e4827b0 JA |
804 | [IORING_OP_EPOLL_CTL] = { |
805 | .unbound_nonreg_file = 1, | |
806 | .file_table = 1, | |
807 | }, | |
7d67af2c PB |
808 | [IORING_OP_SPLICE] = { |
809 | .needs_file = 1, | |
810 | .hash_reg_file = 1, | |
811 | .unbound_nonreg_file = 1, | |
ddf0322d JA |
812 | }, |
813 | [IORING_OP_PROVIDE_BUFFERS] = {}, | |
d3656344 JA |
814 | }; |
815 | ||
561fb04a | 816 | static void io_wq_submit_work(struct io_wq_work **workptr); |
78e19bbe | 817 | static void io_cqring_fill_event(struct io_kiocb *req, long res); |
ec9c02ad | 818 | static void io_put_req(struct io_kiocb *req); |
978db57e | 819 | static void __io_double_put_req(struct io_kiocb *req); |
94ae5e77 JA |
820 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req); |
821 | static void io_queue_linked_timeout(struct io_kiocb *req); | |
05f3fb3c JA |
822 | static int __io_sqe_files_update(struct io_ring_ctx *ctx, |
823 | struct io_uring_files_update *ip, | |
824 | unsigned nr_args); | |
f86cd20c | 825 | static int io_grab_files(struct io_kiocb *req); |
2faf852d | 826 | static void io_ring_file_ref_flush(struct fixed_file_data *data); |
99bc4c38 | 827 | static void io_cleanup_req(struct io_kiocb *req); |
b41e9852 JA |
828 | static int io_file_get(struct io_submit_state *state, struct io_kiocb *req, |
829 | int fd, struct file **out_file, bool fixed); | |
830 | static void __io_queue_sqe(struct io_kiocb *req, | |
831 | const struct io_uring_sqe *sqe); | |
de0617e4 | 832 | |
2b188cc1 JA |
833 | static struct kmem_cache *req_cachep; |
834 | ||
835 | static const struct file_operations io_uring_fops; | |
836 | ||
837 | struct sock *io_uring_get_socket(struct file *file) | |
838 | { | |
839 | #if defined(CONFIG_UNIX) | |
840 | if (file->f_op == &io_uring_fops) { | |
841 | struct io_ring_ctx *ctx = file->private_data; | |
842 | ||
843 | return ctx->ring_sock->sk; | |
844 | } | |
845 | #endif | |
846 | return NULL; | |
847 | } | |
848 | EXPORT_SYMBOL(io_uring_get_socket); | |
849 | ||
850 | static void io_ring_ctx_ref_free(struct percpu_ref *ref) | |
851 | { | |
852 | struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs); | |
853 | ||
206aefde | 854 | complete(&ctx->completions[0]); |
2b188cc1 JA |
855 | } |
856 | ||
857 | static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p) | |
858 | { | |
859 | struct io_ring_ctx *ctx; | |
78076bb6 | 860 | int hash_bits; |
2b188cc1 JA |
861 | |
862 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); | |
863 | if (!ctx) | |
864 | return NULL; | |
865 | ||
0ddf92e8 JA |
866 | ctx->fallback_req = kmem_cache_alloc(req_cachep, GFP_KERNEL); |
867 | if (!ctx->fallback_req) | |
868 | goto err; | |
869 | ||
206aefde JA |
870 | ctx->completions = kmalloc(2 * sizeof(struct completion), GFP_KERNEL); |
871 | if (!ctx->completions) | |
872 | goto err; | |
873 | ||
78076bb6 JA |
874 | /* |
875 | * Use 5 bits less than the max cq entries, that should give us around | |
876 | * 32 entries per hash list if totally full and uniformly spread. | |
877 | */ | |
878 | hash_bits = ilog2(p->cq_entries); | |
879 | hash_bits -= 5; | |
880 | if (hash_bits <= 0) | |
881 | hash_bits = 1; | |
882 | ctx->cancel_hash_bits = hash_bits; | |
883 | ctx->cancel_hash = kmalloc((1U << hash_bits) * sizeof(struct hlist_head), | |
884 | GFP_KERNEL); | |
885 | if (!ctx->cancel_hash) | |
886 | goto err; | |
887 | __hash_init(ctx->cancel_hash, 1U << hash_bits); | |
888 | ||
21482896 | 889 | if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, |
206aefde JA |
890 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) |
891 | goto err; | |
2b188cc1 JA |
892 | |
893 | ctx->flags = p->flags; | |
894 | init_waitqueue_head(&ctx->cq_wait); | |
1d7bb1d5 | 895 | INIT_LIST_HEAD(&ctx->cq_overflow_list); |
206aefde JA |
896 | init_completion(&ctx->completions[0]); |
897 | init_completion(&ctx->completions[1]); | |
5a2e745d | 898 | idr_init(&ctx->io_buffer_idr); |
071698e1 | 899 | idr_init(&ctx->personality_idr); |
2b188cc1 JA |
900 | mutex_init(&ctx->uring_lock); |
901 | init_waitqueue_head(&ctx->wait); | |
902 | spin_lock_init(&ctx->completion_lock); | |
def596e9 | 903 | INIT_LIST_HEAD(&ctx->poll_list); |
de0617e4 | 904 | INIT_LIST_HEAD(&ctx->defer_list); |
5262f567 | 905 | INIT_LIST_HEAD(&ctx->timeout_list); |
fcb323cc JA |
906 | init_waitqueue_head(&ctx->inflight_wait); |
907 | spin_lock_init(&ctx->inflight_lock); | |
908 | INIT_LIST_HEAD(&ctx->inflight_list); | |
2b188cc1 | 909 | return ctx; |
206aefde | 910 | err: |
0ddf92e8 JA |
911 | if (ctx->fallback_req) |
912 | kmem_cache_free(req_cachep, ctx->fallback_req); | |
206aefde | 913 | kfree(ctx->completions); |
78076bb6 | 914 | kfree(ctx->cancel_hash); |
206aefde JA |
915 | kfree(ctx); |
916 | return NULL; | |
2b188cc1 JA |
917 | } |
918 | ||
9d858b21 | 919 | static inline bool __req_need_defer(struct io_kiocb *req) |
7adf4eaf | 920 | { |
a197f664 JL |
921 | struct io_ring_ctx *ctx = req->ctx; |
922 | ||
498ccd9e JA |
923 | return req->sequence != ctx->cached_cq_tail + ctx->cached_sq_dropped |
924 | + atomic_read(&ctx->cached_cq_overflow); | |
7adf4eaf JA |
925 | } |
926 | ||
9d858b21 | 927 | static inline bool req_need_defer(struct io_kiocb *req) |
de0617e4 | 928 | { |
87987898 | 929 | if (unlikely(req->flags & REQ_F_IO_DRAIN)) |
9d858b21 | 930 | return __req_need_defer(req); |
de0617e4 | 931 | |
9d858b21 | 932 | return false; |
de0617e4 JA |
933 | } |
934 | ||
7adf4eaf | 935 | static struct io_kiocb *io_get_deferred_req(struct io_ring_ctx *ctx) |
de0617e4 JA |
936 | { |
937 | struct io_kiocb *req; | |
938 | ||
7adf4eaf | 939 | req = list_first_entry_or_null(&ctx->defer_list, struct io_kiocb, list); |
9d858b21 | 940 | if (req && !req_need_defer(req)) { |
de0617e4 JA |
941 | list_del_init(&req->list); |
942 | return req; | |
943 | } | |
944 | ||
945 | return NULL; | |
946 | } | |
947 | ||
5262f567 JA |
948 | static struct io_kiocb *io_get_timeout_req(struct io_ring_ctx *ctx) |
949 | { | |
7adf4eaf JA |
950 | struct io_kiocb *req; |
951 | ||
952 | req = list_first_entry_or_null(&ctx->timeout_list, struct io_kiocb, list); | |
93bd25bb JA |
953 | if (req) { |
954 | if (req->flags & REQ_F_TIMEOUT_NOSEQ) | |
955 | return NULL; | |
fb4b3d3f | 956 | if (!__req_need_defer(req)) { |
93bd25bb JA |
957 | list_del_init(&req->list); |
958 | return req; | |
959 | } | |
7adf4eaf JA |
960 | } |
961 | ||
962 | return NULL; | |
5262f567 JA |
963 | } |
964 | ||
de0617e4 | 965 | static void __io_commit_cqring(struct io_ring_ctx *ctx) |
2b188cc1 | 966 | { |
75b28aff | 967 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 968 | |
07910158 PB |
969 | /* order cqe stores with ring update */ |
970 | smp_store_release(&rings->cq.tail, ctx->cached_cq_tail); | |
2b188cc1 | 971 | |
07910158 PB |
972 | if (wq_has_sleeper(&ctx->cq_wait)) { |
973 | wake_up_interruptible(&ctx->cq_wait); | |
974 | kill_fasync(&ctx->cq_fasync, SIGIO, POLL_IN); | |
2b188cc1 JA |
975 | } |
976 | } | |
977 | ||
cccf0ee8 JA |
978 | static inline void io_req_work_grab_env(struct io_kiocb *req, |
979 | const struct io_op_def *def) | |
980 | { | |
981 | if (!req->work.mm && def->needs_mm) { | |
982 | mmgrab(current->mm); | |
983 | req->work.mm = current->mm; | |
2b188cc1 | 984 | } |
cccf0ee8 JA |
985 | if (!req->work.creds) |
986 | req->work.creds = get_current_cred(); | |
ff002b30 JA |
987 | if (!req->work.fs && def->needs_fs) { |
988 | spin_lock(¤t->fs->lock); | |
989 | if (!current->fs->in_exec) { | |
990 | req->work.fs = current->fs; | |
991 | req->work.fs->users++; | |
992 | } else { | |
993 | req->work.flags |= IO_WQ_WORK_CANCEL; | |
994 | } | |
995 | spin_unlock(¤t->fs->lock); | |
996 | } | |
6ab23144 JA |
997 | if (!req->work.task_pid) |
998 | req->work.task_pid = task_pid_vnr(current); | |
2b188cc1 JA |
999 | } |
1000 | ||
cccf0ee8 | 1001 | static inline void io_req_work_drop_env(struct io_kiocb *req) |
18d9be1a | 1002 | { |
cccf0ee8 JA |
1003 | if (req->work.mm) { |
1004 | mmdrop(req->work.mm); | |
1005 | req->work.mm = NULL; | |
1006 | } | |
1007 | if (req->work.creds) { | |
1008 | put_cred(req->work.creds); | |
1009 | req->work.creds = NULL; | |
1010 | } | |
ff002b30 JA |
1011 | if (req->work.fs) { |
1012 | struct fs_struct *fs = req->work.fs; | |
1013 | ||
1014 | spin_lock(&req->work.fs->lock); | |
1015 | if (--fs->users) | |
1016 | fs = NULL; | |
1017 | spin_unlock(&req->work.fs->lock); | |
1018 | if (fs) | |
1019 | free_fs_struct(fs); | |
1020 | } | |
561fb04a JA |
1021 | } |
1022 | ||
94ae5e77 JA |
1023 | static inline bool io_prep_async_work(struct io_kiocb *req, |
1024 | struct io_kiocb **link) | |
18d9be1a | 1025 | { |
d3656344 | 1026 | const struct io_op_def *def = &io_op_defs[req->opcode]; |
561fb04a | 1027 | bool do_hashed = false; |
54a91f3b | 1028 | |
d3656344 JA |
1029 | if (req->flags & REQ_F_ISREG) { |
1030 | if (def->hash_reg_file) | |
3529d8c2 | 1031 | do_hashed = true; |
d3656344 JA |
1032 | } else { |
1033 | if (def->unbound_nonreg_file) | |
3529d8c2 | 1034 | req->work.flags |= IO_WQ_WORK_UNBOUND; |
54a91f3b | 1035 | } |
cccf0ee8 JA |
1036 | |
1037 | io_req_work_grab_env(req, def); | |
54a91f3b | 1038 | |
94ae5e77 | 1039 | *link = io_prep_linked_timeout(req); |
561fb04a JA |
1040 | return do_hashed; |
1041 | } | |
1042 | ||
a197f664 | 1043 | static inline void io_queue_async_work(struct io_kiocb *req) |
561fb04a | 1044 | { |
a197f664 | 1045 | struct io_ring_ctx *ctx = req->ctx; |
94ae5e77 JA |
1046 | struct io_kiocb *link; |
1047 | bool do_hashed; | |
1048 | ||
1049 | do_hashed = io_prep_async_work(req, &link); | |
561fb04a JA |
1050 | |
1051 | trace_io_uring_queue_async_work(ctx, do_hashed, req, &req->work, | |
1052 | req->flags); | |
1053 | if (!do_hashed) { | |
1054 | io_wq_enqueue(ctx->io_wq, &req->work); | |
1055 | } else { | |
1056 | io_wq_enqueue_hashed(ctx->io_wq, &req->work, | |
1057 | file_inode(req->file)); | |
1058 | } | |
94ae5e77 JA |
1059 | |
1060 | if (link) | |
1061 | io_queue_linked_timeout(link); | |
18d9be1a JA |
1062 | } |
1063 | ||
5262f567 JA |
1064 | static void io_kill_timeout(struct io_kiocb *req) |
1065 | { | |
1066 | int ret; | |
1067 | ||
2d28390a | 1068 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
5262f567 JA |
1069 | if (ret != -1) { |
1070 | atomic_inc(&req->ctx->cq_timeouts); | |
842f9612 | 1071 | list_del_init(&req->list); |
78e19bbe | 1072 | io_cqring_fill_event(req, 0); |
ec9c02ad | 1073 | io_put_req(req); |
5262f567 JA |
1074 | } |
1075 | } | |
1076 | ||
1077 | static void io_kill_timeouts(struct io_ring_ctx *ctx) | |
1078 | { | |
1079 | struct io_kiocb *req, *tmp; | |
1080 | ||
1081 | spin_lock_irq(&ctx->completion_lock); | |
1082 | list_for_each_entry_safe(req, tmp, &ctx->timeout_list, list) | |
1083 | io_kill_timeout(req); | |
1084 | spin_unlock_irq(&ctx->completion_lock); | |
1085 | } | |
1086 | ||
de0617e4 JA |
1087 | static void io_commit_cqring(struct io_ring_ctx *ctx) |
1088 | { | |
1089 | struct io_kiocb *req; | |
1090 | ||
5262f567 JA |
1091 | while ((req = io_get_timeout_req(ctx)) != NULL) |
1092 | io_kill_timeout(req); | |
1093 | ||
de0617e4 JA |
1094 | __io_commit_cqring(ctx); |
1095 | ||
87987898 | 1096 | while ((req = io_get_deferred_req(ctx)) != NULL) |
a197f664 | 1097 | io_queue_async_work(req); |
de0617e4 JA |
1098 | } |
1099 | ||
2b188cc1 JA |
1100 | static struct io_uring_cqe *io_get_cqring(struct io_ring_ctx *ctx) |
1101 | { | |
75b28aff | 1102 | struct io_rings *rings = ctx->rings; |
2b188cc1 JA |
1103 | unsigned tail; |
1104 | ||
1105 | tail = ctx->cached_cq_tail; | |
115e12e5 SB |
1106 | /* |
1107 | * writes to the cq entry need to come after reading head; the | |
1108 | * control dependency is enough as we're using WRITE_ONCE to | |
1109 | * fill the cq entry | |
1110 | */ | |
75b28aff | 1111 | if (tail - READ_ONCE(rings->cq.head) == rings->cq_ring_entries) |
2b188cc1 JA |
1112 | return NULL; |
1113 | ||
1114 | ctx->cached_cq_tail++; | |
75b28aff | 1115 | return &rings->cqes[tail & ctx->cq_mask]; |
2b188cc1 JA |
1116 | } |
1117 | ||
f2842ab5 JA |
1118 | static inline bool io_should_trigger_evfd(struct io_ring_ctx *ctx) |
1119 | { | |
f0b493e6 JA |
1120 | if (!ctx->cq_ev_fd) |
1121 | return false; | |
f2842ab5 JA |
1122 | if (!ctx->eventfd_async) |
1123 | return true; | |
b41e9852 | 1124 | return io_wq_current_is_worker(); |
f2842ab5 JA |
1125 | } |
1126 | ||
b41e9852 | 1127 | static void io_cqring_ev_posted(struct io_ring_ctx *ctx) |
1d7bb1d5 JA |
1128 | { |
1129 | if (waitqueue_active(&ctx->wait)) | |
1130 | wake_up(&ctx->wait); | |
1131 | if (waitqueue_active(&ctx->sqo_wait)) | |
1132 | wake_up(&ctx->sqo_wait); | |
b41e9852 | 1133 | if (io_should_trigger_evfd(ctx)) |
1d7bb1d5 JA |
1134 | eventfd_signal(ctx->cq_ev_fd, 1); |
1135 | } | |
1136 | ||
c4a2ed72 JA |
1137 | /* Returns true if there are no backlogged entries after the flush */ |
1138 | static bool io_cqring_overflow_flush(struct io_ring_ctx *ctx, bool force) | |
1d7bb1d5 JA |
1139 | { |
1140 | struct io_rings *rings = ctx->rings; | |
1141 | struct io_uring_cqe *cqe; | |
1142 | struct io_kiocb *req; | |
1143 | unsigned long flags; | |
1144 | LIST_HEAD(list); | |
1145 | ||
1146 | if (!force) { | |
1147 | if (list_empty_careful(&ctx->cq_overflow_list)) | |
c4a2ed72 | 1148 | return true; |
1d7bb1d5 JA |
1149 | if ((ctx->cached_cq_tail - READ_ONCE(rings->cq.head) == |
1150 | rings->cq_ring_entries)) | |
c4a2ed72 | 1151 | return false; |
1d7bb1d5 JA |
1152 | } |
1153 | ||
1154 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
1155 | ||
1156 | /* if force is set, the ring is going away. always drop after that */ | |
1157 | if (force) | |
69b3e546 | 1158 | ctx->cq_overflow_flushed = 1; |
1d7bb1d5 | 1159 | |
c4a2ed72 | 1160 | cqe = NULL; |
1d7bb1d5 JA |
1161 | while (!list_empty(&ctx->cq_overflow_list)) { |
1162 | cqe = io_get_cqring(ctx); | |
1163 | if (!cqe && !force) | |
1164 | break; | |
1165 | ||
1166 | req = list_first_entry(&ctx->cq_overflow_list, struct io_kiocb, | |
1167 | list); | |
1168 | list_move(&req->list, &list); | |
2ca10259 | 1169 | req->flags &= ~REQ_F_OVERFLOW; |
1d7bb1d5 JA |
1170 | if (cqe) { |
1171 | WRITE_ONCE(cqe->user_data, req->user_data); | |
1172 | WRITE_ONCE(cqe->res, req->result); | |
1173 | WRITE_ONCE(cqe->flags, 0); | |
1174 | } else { | |
1175 | WRITE_ONCE(ctx->rings->cq_overflow, | |
1176 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
1177 | } | |
1178 | } | |
1179 | ||
1180 | io_commit_cqring(ctx); | |
ad3eb2c8 JA |
1181 | if (cqe) { |
1182 | clear_bit(0, &ctx->sq_check_overflow); | |
1183 | clear_bit(0, &ctx->cq_check_overflow); | |
1184 | } | |
1d7bb1d5 JA |
1185 | spin_unlock_irqrestore(&ctx->completion_lock, flags); |
1186 | io_cqring_ev_posted(ctx); | |
1187 | ||
1188 | while (!list_empty(&list)) { | |
1189 | req = list_first_entry(&list, struct io_kiocb, list); | |
1190 | list_del(&req->list); | |
ec9c02ad | 1191 | io_put_req(req); |
1d7bb1d5 | 1192 | } |
c4a2ed72 JA |
1193 | |
1194 | return cqe != NULL; | |
1d7bb1d5 JA |
1195 | } |
1196 | ||
78e19bbe | 1197 | static void io_cqring_fill_event(struct io_kiocb *req, long res) |
2b188cc1 | 1198 | { |
78e19bbe | 1199 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
1200 | struct io_uring_cqe *cqe; |
1201 | ||
78e19bbe | 1202 | trace_io_uring_complete(ctx, req->user_data, res); |
51c3ff62 | 1203 | |
2b188cc1 JA |
1204 | /* |
1205 | * If we can't get a cq entry, userspace overflowed the | |
1206 | * submission (by quite a lot). Increment the overflow count in | |
1207 | * the ring. | |
1208 | */ | |
1209 | cqe = io_get_cqring(ctx); | |
1d7bb1d5 | 1210 | if (likely(cqe)) { |
78e19bbe | 1211 | WRITE_ONCE(cqe->user_data, req->user_data); |
2b188cc1 | 1212 | WRITE_ONCE(cqe->res, res); |
c71ffb67 | 1213 | WRITE_ONCE(cqe->flags, 0); |
1d7bb1d5 | 1214 | } else if (ctx->cq_overflow_flushed) { |
498ccd9e JA |
1215 | WRITE_ONCE(ctx->rings->cq_overflow, |
1216 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
1d7bb1d5 | 1217 | } else { |
ad3eb2c8 JA |
1218 | if (list_empty(&ctx->cq_overflow_list)) { |
1219 | set_bit(0, &ctx->sq_check_overflow); | |
1220 | set_bit(0, &ctx->cq_check_overflow); | |
1221 | } | |
2ca10259 | 1222 | req->flags |= REQ_F_OVERFLOW; |
1d7bb1d5 JA |
1223 | refcount_inc(&req->refs); |
1224 | req->result = res; | |
1225 | list_add_tail(&req->list, &ctx->cq_overflow_list); | |
2b188cc1 JA |
1226 | } |
1227 | } | |
1228 | ||
78e19bbe | 1229 | static void io_cqring_add_event(struct io_kiocb *req, long res) |
2b188cc1 | 1230 | { |
78e19bbe | 1231 | struct io_ring_ctx *ctx = req->ctx; |
2b188cc1 JA |
1232 | unsigned long flags; |
1233 | ||
1234 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
78e19bbe | 1235 | io_cqring_fill_event(req, res); |
2b188cc1 JA |
1236 | io_commit_cqring(ctx); |
1237 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1238 | ||
8c838788 | 1239 | io_cqring_ev_posted(ctx); |
2b188cc1 JA |
1240 | } |
1241 | ||
0ddf92e8 JA |
1242 | static inline bool io_is_fallback_req(struct io_kiocb *req) |
1243 | { | |
1244 | return req == (struct io_kiocb *) | |
1245 | ((unsigned long) req->ctx->fallback_req & ~1UL); | |
1246 | } | |
1247 | ||
1248 | static struct io_kiocb *io_get_fallback_req(struct io_ring_ctx *ctx) | |
1249 | { | |
1250 | struct io_kiocb *req; | |
1251 | ||
1252 | req = ctx->fallback_req; | |
1253 | if (!test_and_set_bit_lock(0, (unsigned long *) ctx->fallback_req)) | |
1254 | return req; | |
1255 | ||
1256 | return NULL; | |
1257 | } | |
1258 | ||
2579f913 JA |
1259 | static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx, |
1260 | struct io_submit_state *state) | |
2b188cc1 | 1261 | { |
fd6fab2c | 1262 | gfp_t gfp = GFP_KERNEL | __GFP_NOWARN; |
2b188cc1 JA |
1263 | struct io_kiocb *req; |
1264 | ||
2579f913 | 1265 | if (!state) { |
fd6fab2c | 1266 | req = kmem_cache_alloc(req_cachep, gfp); |
2579f913 | 1267 | if (unlikely(!req)) |
0ddf92e8 | 1268 | goto fallback; |
2579f913 JA |
1269 | } else if (!state->free_reqs) { |
1270 | size_t sz; | |
1271 | int ret; | |
1272 | ||
1273 | sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->reqs)); | |
fd6fab2c JA |
1274 | ret = kmem_cache_alloc_bulk(req_cachep, gfp, sz, state->reqs); |
1275 | ||
1276 | /* | |
1277 | * Bulk alloc is all-or-nothing. If we fail to get a batch, | |
1278 | * retry single alloc to be on the safe side. | |
1279 | */ | |
1280 | if (unlikely(ret <= 0)) { | |
1281 | state->reqs[0] = kmem_cache_alloc(req_cachep, gfp); | |
1282 | if (!state->reqs[0]) | |
0ddf92e8 | 1283 | goto fallback; |
fd6fab2c JA |
1284 | ret = 1; |
1285 | } | |
2579f913 | 1286 | state->free_reqs = ret - 1; |
6c8a3134 | 1287 | req = state->reqs[ret - 1]; |
2579f913 | 1288 | } else { |
2579f913 | 1289 | state->free_reqs--; |
6c8a3134 | 1290 | req = state->reqs[state->free_reqs]; |
2b188cc1 JA |
1291 | } |
1292 | ||
0ddf92e8 | 1293 | got_it: |
1a6b74fc | 1294 | req->io = NULL; |
60c112b0 | 1295 | req->file = NULL; |
2579f913 JA |
1296 | req->ctx = ctx; |
1297 | req->flags = 0; | |
e65ef56d JA |
1298 | /* one is dropped after submission, the other at completion */ |
1299 | refcount_set(&req->refs, 2); | |
9e645e11 | 1300 | req->result = 0; |
561fb04a | 1301 | INIT_IO_WORK(&req->work, io_wq_submit_work); |
2579f913 | 1302 | return req; |
0ddf92e8 JA |
1303 | fallback: |
1304 | req = io_get_fallback_req(ctx); | |
1305 | if (req) | |
1306 | goto got_it; | |
6805b32e | 1307 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
1308 | return NULL; |
1309 | } | |
1310 | ||
8da11c19 PB |
1311 | static inline void io_put_file(struct io_kiocb *req, struct file *file, |
1312 | bool fixed) | |
1313 | { | |
1314 | if (fixed) | |
1315 | percpu_ref_put(&req->ctx->file_data->refs); | |
1316 | else | |
1317 | fput(file); | |
1318 | } | |
1319 | ||
2b85edfc | 1320 | static void __io_req_do_free(struct io_kiocb *req) |
def596e9 | 1321 | { |
2b85edfc PB |
1322 | if (likely(!io_is_fallback_req(req))) |
1323 | kmem_cache_free(req_cachep, req); | |
1324 | else | |
1325 | clear_bit_unlock(0, (unsigned long *) req->ctx->fallback_req); | |
1326 | } | |
1327 | ||
c6ca97b3 | 1328 | static void __io_req_aux_free(struct io_kiocb *req) |
2b188cc1 | 1329 | { |
929a3af9 PB |
1330 | if (req->flags & REQ_F_NEED_CLEANUP) |
1331 | io_cleanup_req(req); | |
1332 | ||
96fd84d8 | 1333 | kfree(req->io); |
8da11c19 PB |
1334 | if (req->file) |
1335 | io_put_file(req, req->file, (req->flags & REQ_F_FIXED_FILE)); | |
cccf0ee8 JA |
1336 | |
1337 | io_req_work_drop_env(req); | |
def596e9 JA |
1338 | } |
1339 | ||
9e645e11 | 1340 | static void __io_free_req(struct io_kiocb *req) |
2b188cc1 | 1341 | { |
c6ca97b3 | 1342 | __io_req_aux_free(req); |
fcb323cc | 1343 | |
fcb323cc | 1344 | if (req->flags & REQ_F_INFLIGHT) { |
c6ca97b3 | 1345 | struct io_ring_ctx *ctx = req->ctx; |
fcb323cc JA |
1346 | unsigned long flags; |
1347 | ||
1348 | spin_lock_irqsave(&ctx->inflight_lock, flags); | |
1349 | list_del(&req->inflight_entry); | |
1350 | if (waitqueue_active(&ctx->inflight_wait)) | |
1351 | wake_up(&ctx->inflight_wait); | |
1352 | spin_unlock_irqrestore(&ctx->inflight_lock, flags); | |
1353 | } | |
2b85edfc PB |
1354 | |
1355 | percpu_ref_put(&req->ctx->refs); | |
1356 | __io_req_do_free(req); | |
e65ef56d JA |
1357 | } |
1358 | ||
c6ca97b3 JA |
1359 | struct req_batch { |
1360 | void *reqs[IO_IOPOLL_BATCH]; | |
1361 | int to_free; | |
1362 | int need_iter; | |
1363 | }; | |
1364 | ||
1365 | static void io_free_req_many(struct io_ring_ctx *ctx, struct req_batch *rb) | |
1366 | { | |
10fef4be JA |
1367 | int fixed_refs = rb->to_free; |
1368 | ||
c6ca97b3 JA |
1369 | if (!rb->to_free) |
1370 | return; | |
1371 | if (rb->need_iter) { | |
1372 | int i, inflight = 0; | |
1373 | unsigned long flags; | |
1374 | ||
10fef4be | 1375 | fixed_refs = 0; |
c6ca97b3 JA |
1376 | for (i = 0; i < rb->to_free; i++) { |
1377 | struct io_kiocb *req = rb->reqs[i]; | |
1378 | ||
10fef4be | 1379 | if (req->flags & REQ_F_FIXED_FILE) { |
c6ca97b3 | 1380 | req->file = NULL; |
10fef4be JA |
1381 | fixed_refs++; |
1382 | } | |
c6ca97b3 JA |
1383 | if (req->flags & REQ_F_INFLIGHT) |
1384 | inflight++; | |
c6ca97b3 JA |
1385 | __io_req_aux_free(req); |
1386 | } | |
1387 | if (!inflight) | |
1388 | goto do_free; | |
1389 | ||
1390 | spin_lock_irqsave(&ctx->inflight_lock, flags); | |
1391 | for (i = 0; i < rb->to_free; i++) { | |
1392 | struct io_kiocb *req = rb->reqs[i]; | |
1393 | ||
10fef4be | 1394 | if (req->flags & REQ_F_INFLIGHT) { |
c6ca97b3 JA |
1395 | list_del(&req->inflight_entry); |
1396 | if (!--inflight) | |
1397 | break; | |
1398 | } | |
1399 | } | |
1400 | spin_unlock_irqrestore(&ctx->inflight_lock, flags); | |
1401 | ||
1402 | if (waitqueue_active(&ctx->inflight_wait)) | |
1403 | wake_up(&ctx->inflight_wait); | |
1404 | } | |
1405 | do_free: | |
1406 | kmem_cache_free_bulk(req_cachep, rb->to_free, rb->reqs); | |
10fef4be JA |
1407 | if (fixed_refs) |
1408 | percpu_ref_put_many(&ctx->file_data->refs, fixed_refs); | |
c6ca97b3 | 1409 | percpu_ref_put_many(&ctx->refs, rb->to_free); |
c6ca97b3 | 1410 | rb->to_free = rb->need_iter = 0; |
e65ef56d JA |
1411 | } |
1412 | ||
a197f664 | 1413 | static bool io_link_cancel_timeout(struct io_kiocb *req) |
2665abfd | 1414 | { |
a197f664 | 1415 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd JA |
1416 | int ret; |
1417 | ||
2d28390a | 1418 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
2665abfd | 1419 | if (ret != -1) { |
78e19bbe | 1420 | io_cqring_fill_event(req, -ECANCELED); |
2665abfd JA |
1421 | io_commit_cqring(ctx); |
1422 | req->flags &= ~REQ_F_LINK; | |
ec9c02ad | 1423 | io_put_req(req); |
2665abfd JA |
1424 | return true; |
1425 | } | |
1426 | ||
1427 | return false; | |
e65ef56d JA |
1428 | } |
1429 | ||
ba816ad6 | 1430 | static void io_req_link_next(struct io_kiocb *req, struct io_kiocb **nxtptr) |
9e645e11 | 1431 | { |
2665abfd | 1432 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd | 1433 | bool wake_ev = false; |
9e645e11 | 1434 | |
4d7dd462 JA |
1435 | /* Already got next link */ |
1436 | if (req->flags & REQ_F_LINK_NEXT) | |
1437 | return; | |
1438 | ||
9e645e11 JA |
1439 | /* |
1440 | * The list should never be empty when we are called here. But could | |
1441 | * potentially happen if the chain is messed up, check to be on the | |
1442 | * safe side. | |
1443 | */ | |
4493233e PB |
1444 | while (!list_empty(&req->link_list)) { |
1445 | struct io_kiocb *nxt = list_first_entry(&req->link_list, | |
1446 | struct io_kiocb, link_list); | |
94ae5e77 | 1447 | |
4493233e PB |
1448 | if (unlikely((req->flags & REQ_F_LINK_TIMEOUT) && |
1449 | (nxt->flags & REQ_F_TIMEOUT))) { | |
1450 | list_del_init(&nxt->link_list); | |
94ae5e77 | 1451 | wake_ev |= io_link_cancel_timeout(nxt); |
94ae5e77 JA |
1452 | req->flags &= ~REQ_F_LINK_TIMEOUT; |
1453 | continue; | |
1454 | } | |
9e645e11 | 1455 | |
4493233e PB |
1456 | list_del_init(&req->link_list); |
1457 | if (!list_empty(&nxt->link_list)) | |
1458 | nxt->flags |= REQ_F_LINK; | |
b18fdf71 | 1459 | *nxtptr = nxt; |
94ae5e77 | 1460 | break; |
9e645e11 | 1461 | } |
2665abfd | 1462 | |
4d7dd462 | 1463 | req->flags |= REQ_F_LINK_NEXT; |
2665abfd JA |
1464 | if (wake_ev) |
1465 | io_cqring_ev_posted(ctx); | |
9e645e11 JA |
1466 | } |
1467 | ||
1468 | /* | |
1469 | * Called if REQ_F_LINK is set, and we fail the head request | |
1470 | */ | |
1471 | static void io_fail_links(struct io_kiocb *req) | |
1472 | { | |
2665abfd | 1473 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd JA |
1474 | unsigned long flags; |
1475 | ||
1476 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
9e645e11 JA |
1477 | |
1478 | while (!list_empty(&req->link_list)) { | |
4493233e PB |
1479 | struct io_kiocb *link = list_first_entry(&req->link_list, |
1480 | struct io_kiocb, link_list); | |
9e645e11 | 1481 | |
4493233e | 1482 | list_del_init(&link->link_list); |
c826bd7a | 1483 | trace_io_uring_fail_link(req, link); |
2665abfd JA |
1484 | |
1485 | if ((req->flags & REQ_F_LINK_TIMEOUT) && | |
d625c6ee | 1486 | link->opcode == IORING_OP_LINK_TIMEOUT) { |
a197f664 | 1487 | io_link_cancel_timeout(link); |
2665abfd | 1488 | } else { |
78e19bbe | 1489 | io_cqring_fill_event(link, -ECANCELED); |
978db57e | 1490 | __io_double_put_req(link); |
2665abfd | 1491 | } |
5d960724 | 1492 | req->flags &= ~REQ_F_LINK_TIMEOUT; |
9e645e11 | 1493 | } |
2665abfd JA |
1494 | |
1495 | io_commit_cqring(ctx); | |
1496 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1497 | io_cqring_ev_posted(ctx); | |
9e645e11 JA |
1498 | } |
1499 | ||
4d7dd462 | 1500 | static void io_req_find_next(struct io_kiocb *req, struct io_kiocb **nxt) |
9e645e11 | 1501 | { |
4d7dd462 | 1502 | if (likely(!(req->flags & REQ_F_LINK))) |
2665abfd | 1503 | return; |
2665abfd | 1504 | |
9e645e11 JA |
1505 | /* |
1506 | * If LINK is set, we have dependent requests in this chain. If we | |
1507 | * didn't fail this request, queue the first one up, moving any other | |
1508 | * dependencies to the next request. In case of failure, fail the rest | |
1509 | * of the chain. | |
1510 | */ | |
2665abfd JA |
1511 | if (req->flags & REQ_F_FAIL_LINK) { |
1512 | io_fail_links(req); | |
7c9e7f0f JA |
1513 | } else if ((req->flags & (REQ_F_LINK_TIMEOUT | REQ_F_COMP_LOCKED)) == |
1514 | REQ_F_LINK_TIMEOUT) { | |
2665abfd JA |
1515 | struct io_ring_ctx *ctx = req->ctx; |
1516 | unsigned long flags; | |
1517 | ||
1518 | /* | |
1519 | * If this is a timeout link, we could be racing with the | |
1520 | * timeout timer. Grab the completion lock for this case to | |
7c9e7f0f | 1521 | * protect against that. |
2665abfd JA |
1522 | */ |
1523 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
1524 | io_req_link_next(req, nxt); | |
1525 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
1526 | } else { | |
1527 | io_req_link_next(req, nxt); | |
9e645e11 | 1528 | } |
4d7dd462 | 1529 | } |
9e645e11 | 1530 | |
c69f8dbe JL |
1531 | static void io_free_req(struct io_kiocb *req) |
1532 | { | |
944e58bf PB |
1533 | struct io_kiocb *nxt = NULL; |
1534 | ||
1535 | io_req_find_next(req, &nxt); | |
70cf9f32 | 1536 | __io_free_req(req); |
944e58bf PB |
1537 | |
1538 | if (nxt) | |
1539 | io_queue_async_work(nxt); | |
c69f8dbe JL |
1540 | } |
1541 | ||
7a743e22 PB |
1542 | static void io_link_work_cb(struct io_wq_work **workptr) |
1543 | { | |
1544 | struct io_wq_work *work = *workptr; | |
1545 | struct io_kiocb *link = work->data; | |
1546 | ||
1547 | io_queue_linked_timeout(link); | |
1548 | io_wq_submit_work(workptr); | |
1549 | } | |
1550 | ||
1551 | static void io_wq_assign_next(struct io_wq_work **workptr, struct io_kiocb *nxt) | |
1552 | { | |
1553 | struct io_kiocb *link; | |
1554 | ||
1555 | *workptr = &nxt->work; | |
1556 | link = io_prep_linked_timeout(nxt); | |
1557 | if (link) { | |
1558 | nxt->work.func = io_link_work_cb; | |
1559 | nxt->work.data = link; | |
1560 | } | |
1561 | } | |
1562 | ||
ba816ad6 JA |
1563 | /* |
1564 | * Drop reference to request, return next in chain (if there is one) if this | |
1565 | * was the last reference to this request. | |
1566 | */ | |
f9bd67f6 | 1567 | __attribute__((nonnull)) |
ec9c02ad | 1568 | static void io_put_req_find_next(struct io_kiocb *req, struct io_kiocb **nxtptr) |
e65ef56d | 1569 | { |
2a44f467 JA |
1570 | if (refcount_dec_and_test(&req->refs)) { |
1571 | io_req_find_next(req, nxtptr); | |
4d7dd462 | 1572 | __io_free_req(req); |
2a44f467 | 1573 | } |
2b188cc1 JA |
1574 | } |
1575 | ||
e65ef56d JA |
1576 | static void io_put_req(struct io_kiocb *req) |
1577 | { | |
1578 | if (refcount_dec_and_test(&req->refs)) | |
1579 | io_free_req(req); | |
2b188cc1 JA |
1580 | } |
1581 | ||
e9fd9396 PB |
1582 | static void io_steal_work(struct io_kiocb *req, |
1583 | struct io_wq_work **workptr) | |
7a743e22 PB |
1584 | { |
1585 | /* | |
1586 | * It's in an io-wq worker, so there always should be at least | |
1587 | * one reference, which will be dropped in io_put_work() just | |
1588 | * after the current handler returns. | |
1589 | * | |
1590 | * It also means, that if the counter dropped to 1, then there is | |
1591 | * no asynchronous users left, so it's safe to steal the next work. | |
1592 | */ | |
7a743e22 PB |
1593 | if (refcount_read(&req->refs) == 1) { |
1594 | struct io_kiocb *nxt = NULL; | |
1595 | ||
1596 | io_req_find_next(req, &nxt); | |
1597 | if (nxt) | |
1598 | io_wq_assign_next(workptr, nxt); | |
1599 | } | |
1600 | } | |
1601 | ||
978db57e JA |
1602 | /* |
1603 | * Must only be used if we don't need to care about links, usually from | |
1604 | * within the completion handling itself. | |
1605 | */ | |
1606 | static void __io_double_put_req(struct io_kiocb *req) | |
78e19bbe JA |
1607 | { |
1608 | /* drop both submit and complete references */ | |
1609 | if (refcount_sub_and_test(2, &req->refs)) | |
1610 | __io_free_req(req); | |
1611 | } | |
1612 | ||
978db57e JA |
1613 | static void io_double_put_req(struct io_kiocb *req) |
1614 | { | |
1615 | /* drop both submit and complete references */ | |
1616 | if (refcount_sub_and_test(2, &req->refs)) | |
1617 | io_free_req(req); | |
1618 | } | |
1619 | ||
1d7bb1d5 | 1620 | static unsigned io_cqring_events(struct io_ring_ctx *ctx, bool noflush) |
a3a0e43f | 1621 | { |
84f97dc2 JA |
1622 | struct io_rings *rings = ctx->rings; |
1623 | ||
ad3eb2c8 JA |
1624 | if (test_bit(0, &ctx->cq_check_overflow)) { |
1625 | /* | |
1626 | * noflush == true is from the waitqueue handler, just ensure | |
1627 | * we wake up the task, and the next invocation will flush the | |
1628 | * entries. We cannot safely to it from here. | |
1629 | */ | |
1630 | if (noflush && !list_empty(&ctx->cq_overflow_list)) | |
1631 | return -1U; | |
1d7bb1d5 | 1632 | |
ad3eb2c8 JA |
1633 | io_cqring_overflow_flush(ctx, false); |
1634 | } | |
1d7bb1d5 | 1635 | |
a3a0e43f JA |
1636 | /* See comment at the top of this file */ |
1637 | smp_rmb(); | |
ad3eb2c8 | 1638 | return ctx->cached_cq_tail - READ_ONCE(rings->cq.head); |
a3a0e43f JA |
1639 | } |
1640 | ||
fb5ccc98 PB |
1641 | static inline unsigned int io_sqring_entries(struct io_ring_ctx *ctx) |
1642 | { | |
1643 | struct io_rings *rings = ctx->rings; | |
1644 | ||
1645 | /* make sure SQ entry isn't read before tail */ | |
1646 | return smp_load_acquire(&rings->sq.tail) - ctx->cached_sq_head; | |
1647 | } | |
1648 | ||
8237e045 | 1649 | static inline bool io_req_multi_free(struct req_batch *rb, struct io_kiocb *req) |
e94f141b | 1650 | { |
c6ca97b3 JA |
1651 | if ((req->flags & REQ_F_LINK) || io_is_fallback_req(req)) |
1652 | return false; | |
e94f141b | 1653 | |
c6ca97b3 JA |
1654 | if (!(req->flags & REQ_F_FIXED_FILE) || req->io) |
1655 | rb->need_iter++; | |
1656 | ||
1657 | rb->reqs[rb->to_free++] = req; | |
1658 | if (unlikely(rb->to_free == ARRAY_SIZE(rb->reqs))) | |
1659 | io_free_req_many(req->ctx, rb); | |
1660 | return true; | |
e94f141b JA |
1661 | } |
1662 | ||
def596e9 JA |
1663 | /* |
1664 | * Find and free completed poll iocbs | |
1665 | */ | |
1666 | static void io_iopoll_complete(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1667 | struct list_head *done) | |
1668 | { | |
8237e045 | 1669 | struct req_batch rb; |
def596e9 | 1670 | struct io_kiocb *req; |
def596e9 | 1671 | |
c6ca97b3 | 1672 | rb.to_free = rb.need_iter = 0; |
def596e9 JA |
1673 | while (!list_empty(done)) { |
1674 | req = list_first_entry(done, struct io_kiocb, list); | |
1675 | list_del(&req->list); | |
1676 | ||
78e19bbe | 1677 | io_cqring_fill_event(req, req->result); |
def596e9 JA |
1678 | (*nr_events)++; |
1679 | ||
8237e045 JA |
1680 | if (refcount_dec_and_test(&req->refs) && |
1681 | !io_req_multi_free(&rb, req)) | |
1682 | io_free_req(req); | |
def596e9 | 1683 | } |
def596e9 | 1684 | |
09bb8394 | 1685 | io_commit_cqring(ctx); |
8237e045 | 1686 | io_free_req_many(ctx, &rb); |
def596e9 JA |
1687 | } |
1688 | ||
1689 | static int io_do_iopoll(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1690 | long min) | |
1691 | { | |
1692 | struct io_kiocb *req, *tmp; | |
1693 | LIST_HEAD(done); | |
1694 | bool spin; | |
1695 | int ret; | |
1696 | ||
1697 | /* | |
1698 | * Only spin for completions if we don't have multiple devices hanging | |
1699 | * off our complete list, and we're under the requested amount. | |
1700 | */ | |
1701 | spin = !ctx->poll_multi_file && *nr_events < min; | |
1702 | ||
1703 | ret = 0; | |
1704 | list_for_each_entry_safe(req, tmp, &ctx->poll_list, list) { | |
9adbd45d | 1705 | struct kiocb *kiocb = &req->rw.kiocb; |
def596e9 JA |
1706 | |
1707 | /* | |
1708 | * Move completed entries to our local list. If we find a | |
1709 | * request that requires polling, break out and complete | |
1710 | * the done list first, if we have entries there. | |
1711 | */ | |
1712 | if (req->flags & REQ_F_IOPOLL_COMPLETED) { | |
1713 | list_move_tail(&req->list, &done); | |
1714 | continue; | |
1715 | } | |
1716 | if (!list_empty(&done)) | |
1717 | break; | |
1718 | ||
1719 | ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin); | |
1720 | if (ret < 0) | |
1721 | break; | |
1722 | ||
1723 | if (ret && spin) | |
1724 | spin = false; | |
1725 | ret = 0; | |
1726 | } | |
1727 | ||
1728 | if (!list_empty(&done)) | |
1729 | io_iopoll_complete(ctx, nr_events, &done); | |
1730 | ||
1731 | return ret; | |
1732 | } | |
1733 | ||
1734 | /* | |
d195a66e | 1735 | * Poll for a minimum of 'min' events. Note that if min == 0 we consider that a |
def596e9 JA |
1736 | * non-spinning poll check - we'll still enter the driver poll loop, but only |
1737 | * as a non-spinning completion check. | |
1738 | */ | |
1739 | static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events, | |
1740 | long min) | |
1741 | { | |
08f5439f | 1742 | while (!list_empty(&ctx->poll_list) && !need_resched()) { |
def596e9 JA |
1743 | int ret; |
1744 | ||
1745 | ret = io_do_iopoll(ctx, nr_events, min); | |
1746 | if (ret < 0) | |
1747 | return ret; | |
1748 | if (!min || *nr_events >= min) | |
1749 | return 0; | |
1750 | } | |
1751 | ||
1752 | return 1; | |
1753 | } | |
1754 | ||
1755 | /* | |
1756 | * We can't just wait for polled events to come to us, we have to actively | |
1757 | * find and complete them. | |
1758 | */ | |
1759 | static void io_iopoll_reap_events(struct io_ring_ctx *ctx) | |
1760 | { | |
1761 | if (!(ctx->flags & IORING_SETUP_IOPOLL)) | |
1762 | return; | |
1763 | ||
1764 | mutex_lock(&ctx->uring_lock); | |
1765 | while (!list_empty(&ctx->poll_list)) { | |
1766 | unsigned int nr_events = 0; | |
1767 | ||
1768 | io_iopoll_getevents(ctx, &nr_events, 1); | |
08f5439f JA |
1769 | |
1770 | /* | |
1771 | * Ensure we allow local-to-the-cpu processing to take place, | |
1772 | * in this case we need to ensure that we reap all events. | |
1773 | */ | |
1774 | cond_resched(); | |
def596e9 JA |
1775 | } |
1776 | mutex_unlock(&ctx->uring_lock); | |
1777 | } | |
1778 | ||
c7849be9 XW |
1779 | static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events, |
1780 | long min) | |
def596e9 | 1781 | { |
2b2ed975 | 1782 | int iters = 0, ret = 0; |
500f9fba | 1783 | |
c7849be9 XW |
1784 | /* |
1785 | * We disallow the app entering submit/complete with polling, but we | |
1786 | * still need to lock the ring to prevent racing with polled issue | |
1787 | * that got punted to a workqueue. | |
1788 | */ | |
1789 | mutex_lock(&ctx->uring_lock); | |
def596e9 JA |
1790 | do { |
1791 | int tmin = 0; | |
1792 | ||
a3a0e43f JA |
1793 | /* |
1794 | * Don't enter poll loop if we already have events pending. | |
1795 | * If we do, we can potentially be spinning for commands that | |
1796 | * already triggered a CQE (eg in error). | |
1797 | */ | |
1d7bb1d5 | 1798 | if (io_cqring_events(ctx, false)) |
a3a0e43f JA |
1799 | break; |
1800 | ||
500f9fba JA |
1801 | /* |
1802 | * If a submit got punted to a workqueue, we can have the | |
1803 | * application entering polling for a command before it gets | |
1804 | * issued. That app will hold the uring_lock for the duration | |
1805 | * of the poll right here, so we need to take a breather every | |
1806 | * now and then to ensure that the issue has a chance to add | |
1807 | * the poll to the issued list. Otherwise we can spin here | |
1808 | * forever, while the workqueue is stuck trying to acquire the | |
1809 | * very same mutex. | |
1810 | */ | |
1811 | if (!(++iters & 7)) { | |
1812 | mutex_unlock(&ctx->uring_lock); | |
1813 | mutex_lock(&ctx->uring_lock); | |
1814 | } | |
1815 | ||
def596e9 JA |
1816 | if (*nr_events < min) |
1817 | tmin = min - *nr_events; | |
1818 | ||
1819 | ret = io_iopoll_getevents(ctx, nr_events, tmin); | |
1820 | if (ret <= 0) | |
1821 | break; | |
1822 | ret = 0; | |
1823 | } while (min && !*nr_events && !need_resched()); | |
1824 | ||
500f9fba | 1825 | mutex_unlock(&ctx->uring_lock); |
def596e9 JA |
1826 | return ret; |
1827 | } | |
1828 | ||
491381ce | 1829 | static void kiocb_end_write(struct io_kiocb *req) |
2b188cc1 | 1830 | { |
491381ce JA |
1831 | /* |
1832 | * Tell lockdep we inherited freeze protection from submission | |
1833 | * thread. | |
1834 | */ | |
1835 | if (req->flags & REQ_F_ISREG) { | |
1836 | struct inode *inode = file_inode(req->file); | |
2b188cc1 | 1837 | |
491381ce | 1838 | __sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE); |
2b188cc1 | 1839 | } |
491381ce | 1840 | file_end_write(req->file); |
2b188cc1 JA |
1841 | } |
1842 | ||
4e88d6e7 JA |
1843 | static inline void req_set_fail_links(struct io_kiocb *req) |
1844 | { | |
1845 | if ((req->flags & (REQ_F_LINK | REQ_F_HARDLINK)) == REQ_F_LINK) | |
1846 | req->flags |= REQ_F_FAIL_LINK; | |
1847 | } | |
1848 | ||
ba816ad6 | 1849 | static void io_complete_rw_common(struct kiocb *kiocb, long res) |
2b188cc1 | 1850 | { |
9adbd45d | 1851 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
2b188cc1 | 1852 | |
491381ce JA |
1853 | if (kiocb->ki_flags & IOCB_WRITE) |
1854 | kiocb_end_write(req); | |
2b188cc1 | 1855 | |
4e88d6e7 JA |
1856 | if (res != req->result) |
1857 | req_set_fail_links(req); | |
78e19bbe | 1858 | io_cqring_add_event(req, res); |
ba816ad6 JA |
1859 | } |
1860 | ||
1861 | static void io_complete_rw(struct kiocb *kiocb, long res, long res2) | |
1862 | { | |
9adbd45d | 1863 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
ba816ad6 JA |
1864 | |
1865 | io_complete_rw_common(kiocb, res); | |
e65ef56d | 1866 | io_put_req(req); |
2b188cc1 JA |
1867 | } |
1868 | ||
def596e9 JA |
1869 | static void io_complete_rw_iopoll(struct kiocb *kiocb, long res, long res2) |
1870 | { | |
9adbd45d | 1871 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
def596e9 | 1872 | |
491381ce JA |
1873 | if (kiocb->ki_flags & IOCB_WRITE) |
1874 | kiocb_end_write(req); | |
def596e9 | 1875 | |
4e88d6e7 JA |
1876 | if (res != req->result) |
1877 | req_set_fail_links(req); | |
9e645e11 | 1878 | req->result = res; |
def596e9 JA |
1879 | if (res != -EAGAIN) |
1880 | req->flags |= REQ_F_IOPOLL_COMPLETED; | |
1881 | } | |
1882 | ||
1883 | /* | |
1884 | * After the iocb has been issued, it's safe to be found on the poll list. | |
1885 | * Adding the kiocb to the list AFTER submission ensures that we don't | |
1886 | * find it from a io_iopoll_getevents() thread before the issuer is done | |
1887 | * accessing the kiocb cookie. | |
1888 | */ | |
1889 | static void io_iopoll_req_issued(struct io_kiocb *req) | |
1890 | { | |
1891 | struct io_ring_ctx *ctx = req->ctx; | |
1892 | ||
1893 | /* | |
1894 | * Track whether we have multiple files in our lists. This will impact | |
1895 | * how we do polling eventually, not spinning if we're on potentially | |
1896 | * different devices. | |
1897 | */ | |
1898 | if (list_empty(&ctx->poll_list)) { | |
1899 | ctx->poll_multi_file = false; | |
1900 | } else if (!ctx->poll_multi_file) { | |
1901 | struct io_kiocb *list_req; | |
1902 | ||
1903 | list_req = list_first_entry(&ctx->poll_list, struct io_kiocb, | |
1904 | list); | |
9adbd45d | 1905 | if (list_req->file != req->file) |
def596e9 JA |
1906 | ctx->poll_multi_file = true; |
1907 | } | |
1908 | ||
1909 | /* | |
1910 | * For fast devices, IO may have already completed. If it has, add | |
1911 | * it to the front so we find it first. | |
1912 | */ | |
1913 | if (req->flags & REQ_F_IOPOLL_COMPLETED) | |
1914 | list_add(&req->list, &ctx->poll_list); | |
1915 | else | |
1916 | list_add_tail(&req->list, &ctx->poll_list); | |
bdcd3eab XW |
1917 | |
1918 | if ((ctx->flags & IORING_SETUP_SQPOLL) && | |
1919 | wq_has_sleeper(&ctx->sqo_wait)) | |
1920 | wake_up(&ctx->sqo_wait); | |
def596e9 JA |
1921 | } |
1922 | ||
3d6770fb | 1923 | static void io_file_put(struct io_submit_state *state) |
9a56a232 | 1924 | { |
3d6770fb | 1925 | if (state->file) { |
9a56a232 JA |
1926 | int diff = state->has_refs - state->used_refs; |
1927 | ||
1928 | if (diff) | |
1929 | fput_many(state->file, diff); | |
1930 | state->file = NULL; | |
1931 | } | |
1932 | } | |
1933 | ||
1934 | /* | |
1935 | * Get as many references to a file as we have IOs left in this submission, | |
1936 | * assuming most submissions are for one file, or at least that each file | |
1937 | * has more than one submission. | |
1938 | */ | |
8da11c19 | 1939 | static struct file *__io_file_get(struct io_submit_state *state, int fd) |
9a56a232 JA |
1940 | { |
1941 | if (!state) | |
1942 | return fget(fd); | |
1943 | ||
1944 | if (state->file) { | |
1945 | if (state->fd == fd) { | |
1946 | state->used_refs++; | |
1947 | state->ios_left--; | |
1948 | return state->file; | |
1949 | } | |
3d6770fb | 1950 | io_file_put(state); |
9a56a232 JA |
1951 | } |
1952 | state->file = fget_many(fd, state->ios_left); | |
1953 | if (!state->file) | |
1954 | return NULL; | |
1955 | ||
1956 | state->fd = fd; | |
1957 | state->has_refs = state->ios_left; | |
1958 | state->used_refs = 1; | |
1959 | state->ios_left--; | |
1960 | return state->file; | |
1961 | } | |
1962 | ||
2b188cc1 JA |
1963 | /* |
1964 | * If we tracked the file through the SCM inflight mechanism, we could support | |
1965 | * any file. For now, just ensure that anything potentially problematic is done | |
1966 | * inline. | |
1967 | */ | |
1968 | static bool io_file_supports_async(struct file *file) | |
1969 | { | |
1970 | umode_t mode = file_inode(file)->i_mode; | |
1971 | ||
10d59345 | 1972 | if (S_ISBLK(mode) || S_ISCHR(mode) || S_ISSOCK(mode)) |
2b188cc1 JA |
1973 | return true; |
1974 | if (S_ISREG(mode) && file->f_op != &io_uring_fops) | |
1975 | return true; | |
1976 | ||
1977 | return false; | |
1978 | } | |
1979 | ||
3529d8c2 JA |
1980 | static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
1981 | bool force_nonblock) | |
2b188cc1 | 1982 | { |
def596e9 | 1983 | struct io_ring_ctx *ctx = req->ctx; |
9adbd45d | 1984 | struct kiocb *kiocb = &req->rw.kiocb; |
09bb8394 JA |
1985 | unsigned ioprio; |
1986 | int ret; | |
2b188cc1 | 1987 | |
491381ce JA |
1988 | if (S_ISREG(file_inode(req->file)->i_mode)) |
1989 | req->flags |= REQ_F_ISREG; | |
1990 | ||
2b188cc1 | 1991 | kiocb->ki_pos = READ_ONCE(sqe->off); |
ba04291e JA |
1992 | if (kiocb->ki_pos == -1 && !(req->file->f_mode & FMODE_STREAM)) { |
1993 | req->flags |= REQ_F_CUR_POS; | |
1994 | kiocb->ki_pos = req->file->f_pos; | |
1995 | } | |
2b188cc1 | 1996 | kiocb->ki_hint = ki_hint_validate(file_write_hint(kiocb->ki_filp)); |
3e577dcd PB |
1997 | kiocb->ki_flags = iocb_flags(kiocb->ki_filp); |
1998 | ret = kiocb_set_rw_flags(kiocb, READ_ONCE(sqe->rw_flags)); | |
1999 | if (unlikely(ret)) | |
2000 | return ret; | |
2b188cc1 JA |
2001 | |
2002 | ioprio = READ_ONCE(sqe->ioprio); | |
2003 | if (ioprio) { | |
2004 | ret = ioprio_check_cap(ioprio); | |
2005 | if (ret) | |
09bb8394 | 2006 | return ret; |
2b188cc1 JA |
2007 | |
2008 | kiocb->ki_ioprio = ioprio; | |
2009 | } else | |
2010 | kiocb->ki_ioprio = get_current_ioprio(); | |
2011 | ||
8449eeda | 2012 | /* don't allow async punt if RWF_NOWAIT was requested */ |
491381ce JA |
2013 | if ((kiocb->ki_flags & IOCB_NOWAIT) || |
2014 | (req->file->f_flags & O_NONBLOCK)) | |
8449eeda SB |
2015 | req->flags |= REQ_F_NOWAIT; |
2016 | ||
2017 | if (force_nonblock) | |
2b188cc1 | 2018 | kiocb->ki_flags |= IOCB_NOWAIT; |
8449eeda | 2019 | |
def596e9 | 2020 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 JA |
2021 | if (!(kiocb->ki_flags & IOCB_DIRECT) || |
2022 | !kiocb->ki_filp->f_op->iopoll) | |
09bb8394 | 2023 | return -EOPNOTSUPP; |
2b188cc1 | 2024 | |
def596e9 JA |
2025 | kiocb->ki_flags |= IOCB_HIPRI; |
2026 | kiocb->ki_complete = io_complete_rw_iopoll; | |
6873e0bd | 2027 | req->result = 0; |
def596e9 | 2028 | } else { |
09bb8394 JA |
2029 | if (kiocb->ki_flags & IOCB_HIPRI) |
2030 | return -EINVAL; | |
def596e9 JA |
2031 | kiocb->ki_complete = io_complete_rw; |
2032 | } | |
9adbd45d | 2033 | |
3529d8c2 JA |
2034 | req->rw.addr = READ_ONCE(sqe->addr); |
2035 | req->rw.len = READ_ONCE(sqe->len); | |
9adbd45d JA |
2036 | /* we own ->private, reuse it for the buffer index */ |
2037 | req->rw.kiocb.private = (void *) (unsigned long) | |
3529d8c2 | 2038 | READ_ONCE(sqe->buf_index); |
2b188cc1 | 2039 | return 0; |
2b188cc1 JA |
2040 | } |
2041 | ||
2042 | static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) | |
2043 | { | |
2044 | switch (ret) { | |
2045 | case -EIOCBQUEUED: | |
2046 | break; | |
2047 | case -ERESTARTSYS: | |
2048 | case -ERESTARTNOINTR: | |
2049 | case -ERESTARTNOHAND: | |
2050 | case -ERESTART_RESTARTBLOCK: | |
2051 | /* | |
2052 | * We can't just restart the syscall, since previously | |
2053 | * submitted sqes may already be in progress. Just fail this | |
2054 | * IO with EINTR. | |
2055 | */ | |
2056 | ret = -EINTR; | |
2057 | /* fall through */ | |
2058 | default: | |
2059 | kiocb->ki_complete(kiocb, ret, 0); | |
2060 | } | |
2061 | } | |
2062 | ||
014db007 | 2063 | static void kiocb_done(struct kiocb *kiocb, ssize_t ret) |
ba816ad6 | 2064 | { |
ba04291e JA |
2065 | struct io_kiocb *req = container_of(kiocb, struct io_kiocb, rw.kiocb); |
2066 | ||
2067 | if (req->flags & REQ_F_CUR_POS) | |
2068 | req->file->f_pos = kiocb->ki_pos; | |
bcaec089 | 2069 | if (ret >= 0 && kiocb->ki_complete == io_complete_rw) |
014db007 | 2070 | io_complete_rw(kiocb, ret, 0); |
ba816ad6 JA |
2071 | else |
2072 | io_rw_done(kiocb, ret); | |
2073 | } | |
2074 | ||
9adbd45d | 2075 | static ssize_t io_import_fixed(struct io_kiocb *req, int rw, |
7d009165 | 2076 | struct iov_iter *iter) |
edafccee | 2077 | { |
9adbd45d JA |
2078 | struct io_ring_ctx *ctx = req->ctx; |
2079 | size_t len = req->rw.len; | |
edafccee JA |
2080 | struct io_mapped_ubuf *imu; |
2081 | unsigned index, buf_index; | |
2082 | size_t offset; | |
2083 | u64 buf_addr; | |
2084 | ||
2085 | /* attempt to use fixed buffers without having provided iovecs */ | |
2086 | if (unlikely(!ctx->user_bufs)) | |
2087 | return -EFAULT; | |
2088 | ||
9adbd45d | 2089 | buf_index = (unsigned long) req->rw.kiocb.private; |
edafccee JA |
2090 | if (unlikely(buf_index >= ctx->nr_user_bufs)) |
2091 | return -EFAULT; | |
2092 | ||
2093 | index = array_index_nospec(buf_index, ctx->nr_user_bufs); | |
2094 | imu = &ctx->user_bufs[index]; | |
9adbd45d | 2095 | buf_addr = req->rw.addr; |
edafccee JA |
2096 | |
2097 | /* overflow */ | |
2098 | if (buf_addr + len < buf_addr) | |
2099 | return -EFAULT; | |
2100 | /* not inside the mapped region */ | |
2101 | if (buf_addr < imu->ubuf || buf_addr + len > imu->ubuf + imu->len) | |
2102 | return -EFAULT; | |
2103 | ||
2104 | /* | |
2105 | * May not be a start of buffer, set size appropriately | |
2106 | * and advance us to the beginning. | |
2107 | */ | |
2108 | offset = buf_addr - imu->ubuf; | |
2109 | iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len); | |
bd11b3a3 JA |
2110 | |
2111 | if (offset) { | |
2112 | /* | |
2113 | * Don't use iov_iter_advance() here, as it's really slow for | |
2114 | * using the latter parts of a big fixed buffer - it iterates | |
2115 | * over each segment manually. We can cheat a bit here, because | |
2116 | * we know that: | |
2117 | * | |
2118 | * 1) it's a BVEC iter, we set it up | |
2119 | * 2) all bvecs are PAGE_SIZE in size, except potentially the | |
2120 | * first and last bvec | |
2121 | * | |
2122 | * So just find our index, and adjust the iterator afterwards. | |
2123 | * If the offset is within the first bvec (or the whole first | |
2124 | * bvec, just use iov_iter_advance(). This makes it easier | |
2125 | * since we can just skip the first segment, which may not | |
2126 | * be PAGE_SIZE aligned. | |
2127 | */ | |
2128 | const struct bio_vec *bvec = imu->bvec; | |
2129 | ||
2130 | if (offset <= bvec->bv_len) { | |
2131 | iov_iter_advance(iter, offset); | |
2132 | } else { | |
2133 | unsigned long seg_skip; | |
2134 | ||
2135 | /* skip first vec */ | |
2136 | offset -= bvec->bv_len; | |
2137 | seg_skip = 1 + (offset >> PAGE_SHIFT); | |
2138 | ||
2139 | iter->bvec = bvec + seg_skip; | |
2140 | iter->nr_segs -= seg_skip; | |
99c79f66 | 2141 | iter->count -= bvec->bv_len + offset; |
bd11b3a3 | 2142 | iter->iov_offset = offset & ~PAGE_MASK; |
bd11b3a3 JA |
2143 | } |
2144 | } | |
2145 | ||
5e559561 | 2146 | return len; |
edafccee JA |
2147 | } |
2148 | ||
cf6fd4bd PB |
2149 | static ssize_t io_import_iovec(int rw, struct io_kiocb *req, |
2150 | struct iovec **iovec, struct iov_iter *iter) | |
2b188cc1 | 2151 | { |
9adbd45d JA |
2152 | void __user *buf = u64_to_user_ptr(req->rw.addr); |
2153 | size_t sqe_len = req->rw.len; | |
edafccee JA |
2154 | u8 opcode; |
2155 | ||
d625c6ee | 2156 | opcode = req->opcode; |
7d009165 | 2157 | if (opcode == IORING_OP_READ_FIXED || opcode == IORING_OP_WRITE_FIXED) { |
edafccee | 2158 | *iovec = NULL; |
9adbd45d | 2159 | return io_import_fixed(req, rw, iter); |
edafccee | 2160 | } |
2b188cc1 | 2161 | |
9adbd45d JA |
2162 | /* buffer index only valid with fixed read/write */ |
2163 | if (req->rw.kiocb.private) | |
2164 | return -EINVAL; | |
2165 | ||
3a6820f2 JA |
2166 | if (opcode == IORING_OP_READ || opcode == IORING_OP_WRITE) { |
2167 | ssize_t ret; | |
2168 | ret = import_single_range(rw, buf, sqe_len, *iovec, iter); | |
2169 | *iovec = NULL; | |
3a901598 | 2170 | return ret < 0 ? ret : sqe_len; |
3a6820f2 JA |
2171 | } |
2172 | ||
f67676d1 JA |
2173 | if (req->io) { |
2174 | struct io_async_rw *iorw = &req->io->rw; | |
2175 | ||
2176 | *iovec = iorw->iov; | |
2177 | iov_iter_init(iter, rw, *iovec, iorw->nr_segs, iorw->size); | |
2178 | if (iorw->iov == iorw->fast_iov) | |
2179 | *iovec = NULL; | |
2180 | return iorw->size; | |
2181 | } | |
2182 | ||
2b188cc1 | 2183 | #ifdef CONFIG_COMPAT |
cf6fd4bd | 2184 | if (req->ctx->compat) |
2b188cc1 JA |
2185 | return compat_import_iovec(rw, buf, sqe_len, UIO_FASTIOV, |
2186 | iovec, iter); | |
2187 | #endif | |
2188 | ||
2189 | return import_iovec(rw, buf, sqe_len, UIO_FASTIOV, iovec, iter); | |
2190 | } | |
2191 | ||
31b51510 | 2192 | /* |
32960613 JA |
2193 | * For files that don't have ->read_iter() and ->write_iter(), handle them |
2194 | * by looping over ->read() or ->write() manually. | |
31b51510 | 2195 | */ |
32960613 JA |
2196 | static ssize_t loop_rw_iter(int rw, struct file *file, struct kiocb *kiocb, |
2197 | struct iov_iter *iter) | |
2198 | { | |
2199 | ssize_t ret = 0; | |
2200 | ||
2201 | /* | |
2202 | * Don't support polled IO through this interface, and we can't | |
2203 | * support non-blocking either. For the latter, this just causes | |
2204 | * the kiocb to be handled from an async context. | |
2205 | */ | |
2206 | if (kiocb->ki_flags & IOCB_HIPRI) | |
2207 | return -EOPNOTSUPP; | |
2208 | if (kiocb->ki_flags & IOCB_NOWAIT) | |
2209 | return -EAGAIN; | |
2210 | ||
2211 | while (iov_iter_count(iter)) { | |
311ae9e1 | 2212 | struct iovec iovec; |
32960613 JA |
2213 | ssize_t nr; |
2214 | ||
311ae9e1 PB |
2215 | if (!iov_iter_is_bvec(iter)) { |
2216 | iovec = iov_iter_iovec(iter); | |
2217 | } else { | |
2218 | /* fixed buffers import bvec */ | |
2219 | iovec.iov_base = kmap(iter->bvec->bv_page) | |
2220 | + iter->iov_offset; | |
2221 | iovec.iov_len = min(iter->count, | |
2222 | iter->bvec->bv_len - iter->iov_offset); | |
2223 | } | |
2224 | ||
32960613 JA |
2225 | if (rw == READ) { |
2226 | nr = file->f_op->read(file, iovec.iov_base, | |
2227 | iovec.iov_len, &kiocb->ki_pos); | |
2228 | } else { | |
2229 | nr = file->f_op->write(file, iovec.iov_base, | |
2230 | iovec.iov_len, &kiocb->ki_pos); | |
2231 | } | |
2232 | ||
311ae9e1 PB |
2233 | if (iov_iter_is_bvec(iter)) |
2234 | kunmap(iter->bvec->bv_page); | |
2235 | ||
32960613 JA |
2236 | if (nr < 0) { |
2237 | if (!ret) | |
2238 | ret = nr; | |
2239 | break; | |
2240 | } | |
2241 | ret += nr; | |
2242 | if (nr != iovec.iov_len) | |
2243 | break; | |
2244 | iov_iter_advance(iter, nr); | |
2245 | } | |
2246 | ||
2247 | return ret; | |
2248 | } | |
2249 | ||
b7bb4f7d | 2250 | static void io_req_map_rw(struct io_kiocb *req, ssize_t io_size, |
f67676d1 JA |
2251 | struct iovec *iovec, struct iovec *fast_iov, |
2252 | struct iov_iter *iter) | |
2253 | { | |
2254 | req->io->rw.nr_segs = iter->nr_segs; | |
2255 | req->io->rw.size = io_size; | |
2256 | req->io->rw.iov = iovec; | |
2257 | if (!req->io->rw.iov) { | |
2258 | req->io->rw.iov = req->io->rw.fast_iov; | |
2259 | memcpy(req->io->rw.iov, fast_iov, | |
2260 | sizeof(struct iovec) * iter->nr_segs); | |
99bc4c38 PB |
2261 | } else { |
2262 | req->flags |= REQ_F_NEED_CLEANUP; | |
f67676d1 JA |
2263 | } |
2264 | } | |
2265 | ||
b7bb4f7d | 2266 | static int io_alloc_async_ctx(struct io_kiocb *req) |
f67676d1 | 2267 | { |
d3656344 JA |
2268 | if (!io_op_defs[req->opcode].async_ctx) |
2269 | return 0; | |
f67676d1 | 2270 | req->io = kmalloc(sizeof(*req->io), GFP_KERNEL); |
06b76d44 | 2271 | return req->io == NULL; |
b7bb4f7d JA |
2272 | } |
2273 | ||
b7bb4f7d JA |
2274 | static int io_setup_async_rw(struct io_kiocb *req, ssize_t io_size, |
2275 | struct iovec *iovec, struct iovec *fast_iov, | |
2276 | struct iov_iter *iter) | |
2277 | { | |
980ad263 | 2278 | if (!io_op_defs[req->opcode].async_ctx) |
74566df3 | 2279 | return 0; |
5d204bcf JA |
2280 | if (!req->io) { |
2281 | if (io_alloc_async_ctx(req)) | |
2282 | return -ENOMEM; | |
b7bb4f7d | 2283 | |
5d204bcf JA |
2284 | io_req_map_rw(req, io_size, iovec, fast_iov, iter); |
2285 | } | |
b7bb4f7d | 2286 | return 0; |
f67676d1 JA |
2287 | } |
2288 | ||
3529d8c2 JA |
2289 | static int io_read_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2290 | bool force_nonblock) | |
f67676d1 | 2291 | { |
3529d8c2 JA |
2292 | struct io_async_ctx *io; |
2293 | struct iov_iter iter; | |
f67676d1 JA |
2294 | ssize_t ret; |
2295 | ||
3529d8c2 JA |
2296 | ret = io_prep_rw(req, sqe, force_nonblock); |
2297 | if (ret) | |
2298 | return ret; | |
f67676d1 | 2299 | |
3529d8c2 JA |
2300 | if (unlikely(!(req->file->f_mode & FMODE_READ))) |
2301 | return -EBADF; | |
f67676d1 | 2302 | |
5f798bea PB |
2303 | /* either don't need iovec imported or already have it */ |
2304 | if (!req->io || req->flags & REQ_F_NEED_CLEANUP) | |
3529d8c2 JA |
2305 | return 0; |
2306 | ||
2307 | io = req->io; | |
2308 | io->rw.iov = io->rw.fast_iov; | |
2309 | req->io = NULL; | |
2310 | ret = io_import_iovec(READ, req, &io->rw.iov, &iter); | |
2311 | req->io = io; | |
2312 | if (ret < 0) | |
2313 | return ret; | |
2314 | ||
2315 | io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter); | |
2316 | return 0; | |
f67676d1 JA |
2317 | } |
2318 | ||
014db007 | 2319 | static int io_read(struct io_kiocb *req, bool force_nonblock) |
2b188cc1 JA |
2320 | { |
2321 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
9adbd45d | 2322 | struct kiocb *kiocb = &req->rw.kiocb; |
2b188cc1 | 2323 | struct iov_iter iter; |
31b51510 | 2324 | size_t iov_count; |
f67676d1 | 2325 | ssize_t io_size, ret; |
2b188cc1 | 2326 | |
3529d8c2 | 2327 | ret = io_import_iovec(READ, req, &iovec, &iter); |
06b76d44 JA |
2328 | if (ret < 0) |
2329 | return ret; | |
2b188cc1 | 2330 | |
fd6c2e4c JA |
2331 | /* Ensure we clear previously set non-block flag */ |
2332 | if (!force_nonblock) | |
29de5f6a | 2333 | kiocb->ki_flags &= ~IOCB_NOWAIT; |
fd6c2e4c | 2334 | |
797f3f53 | 2335 | req->result = 0; |
f67676d1 | 2336 | io_size = ret; |
9e645e11 | 2337 | if (req->flags & REQ_F_LINK) |
f67676d1 JA |
2338 | req->result = io_size; |
2339 | ||
2340 | /* | |
2341 | * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so | |
2342 | * we know to async punt it even if it was opened O_NONBLOCK | |
2343 | */ | |
29de5f6a | 2344 | if (force_nonblock && !io_file_supports_async(req->file)) |
f67676d1 | 2345 | goto copy_iov; |
9e645e11 | 2346 | |
31b51510 | 2347 | iov_count = iov_iter_count(&iter); |
9adbd45d | 2348 | ret = rw_verify_area(READ, req->file, &kiocb->ki_pos, iov_count); |
2b188cc1 JA |
2349 | if (!ret) { |
2350 | ssize_t ret2; | |
2351 | ||
9adbd45d JA |
2352 | if (req->file->f_op->read_iter) |
2353 | ret2 = call_read_iter(req->file, kiocb, &iter); | |
32960613 | 2354 | else |
9adbd45d | 2355 | ret2 = loop_rw_iter(READ, req->file, kiocb, &iter); |
32960613 | 2356 | |
9d93a3f5 | 2357 | /* Catch -EAGAIN return for forced non-blocking submission */ |
f67676d1 | 2358 | if (!force_nonblock || ret2 != -EAGAIN) { |
014db007 | 2359 | kiocb_done(kiocb, ret2); |
f67676d1 JA |
2360 | } else { |
2361 | copy_iov: | |
b7bb4f7d | 2362 | ret = io_setup_async_rw(req, io_size, iovec, |
f67676d1 JA |
2363 | inline_vecs, &iter); |
2364 | if (ret) | |
2365 | goto out_free; | |
29de5f6a JA |
2366 | /* any defer here is final, must blocking retry */ |
2367 | if (!(req->flags & REQ_F_NOWAIT)) | |
2368 | req->flags |= REQ_F_MUST_PUNT; | |
f67676d1 JA |
2369 | return -EAGAIN; |
2370 | } | |
2b188cc1 | 2371 | } |
f67676d1 | 2372 | out_free: |
1e95081c | 2373 | kfree(iovec); |
99bc4c38 | 2374 | req->flags &= ~REQ_F_NEED_CLEANUP; |
2b188cc1 JA |
2375 | return ret; |
2376 | } | |
2377 | ||
3529d8c2 JA |
2378 | static int io_write_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2379 | bool force_nonblock) | |
f67676d1 | 2380 | { |
3529d8c2 JA |
2381 | struct io_async_ctx *io; |
2382 | struct iov_iter iter; | |
f67676d1 JA |
2383 | ssize_t ret; |
2384 | ||
3529d8c2 JA |
2385 | ret = io_prep_rw(req, sqe, force_nonblock); |
2386 | if (ret) | |
2387 | return ret; | |
f67676d1 | 2388 | |
3529d8c2 JA |
2389 | if (unlikely(!(req->file->f_mode & FMODE_WRITE))) |
2390 | return -EBADF; | |
f67676d1 | 2391 | |
5f798bea PB |
2392 | /* either don't need iovec imported or already have it */ |
2393 | if (!req->io || req->flags & REQ_F_NEED_CLEANUP) | |
3529d8c2 JA |
2394 | return 0; |
2395 | ||
2396 | io = req->io; | |
2397 | io->rw.iov = io->rw.fast_iov; | |
2398 | req->io = NULL; | |
2399 | ret = io_import_iovec(WRITE, req, &io->rw.iov, &iter); | |
2400 | req->io = io; | |
2401 | if (ret < 0) | |
2402 | return ret; | |
2403 | ||
2404 | io_req_map_rw(req, ret, io->rw.iov, io->rw.fast_iov, &iter); | |
2405 | return 0; | |
f67676d1 JA |
2406 | } |
2407 | ||
014db007 | 2408 | static int io_write(struct io_kiocb *req, bool force_nonblock) |
2b188cc1 JA |
2409 | { |
2410 | struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs; | |
9adbd45d | 2411 | struct kiocb *kiocb = &req->rw.kiocb; |
2b188cc1 | 2412 | struct iov_iter iter; |
31b51510 | 2413 | size_t iov_count; |
f67676d1 | 2414 | ssize_t ret, io_size; |
2b188cc1 | 2415 | |
3529d8c2 | 2416 | ret = io_import_iovec(WRITE, req, &iovec, &iter); |
06b76d44 JA |
2417 | if (ret < 0) |
2418 | return ret; | |
2b188cc1 | 2419 | |
fd6c2e4c JA |
2420 | /* Ensure we clear previously set non-block flag */ |
2421 | if (!force_nonblock) | |
9adbd45d | 2422 | req->rw.kiocb.ki_flags &= ~IOCB_NOWAIT; |
fd6c2e4c | 2423 | |
797f3f53 | 2424 | req->result = 0; |
f67676d1 | 2425 | io_size = ret; |
9e645e11 | 2426 | if (req->flags & REQ_F_LINK) |
f67676d1 | 2427 | req->result = io_size; |
9e645e11 | 2428 | |
f67676d1 JA |
2429 | /* |
2430 | * If the file doesn't support async, mark it as REQ_F_MUST_PUNT so | |
2431 | * we know to async punt it even if it was opened O_NONBLOCK | |
2432 | */ | |
29de5f6a | 2433 | if (force_nonblock && !io_file_supports_async(req->file)) |
f67676d1 | 2434 | goto copy_iov; |
31b51510 | 2435 | |
10d59345 JA |
2436 | /* file path doesn't support NOWAIT for non-direct_IO */ |
2437 | if (force_nonblock && !(kiocb->ki_flags & IOCB_DIRECT) && | |
2438 | (req->flags & REQ_F_ISREG)) | |
f67676d1 | 2439 | goto copy_iov; |
31b51510 | 2440 | |
f67676d1 | 2441 | iov_count = iov_iter_count(&iter); |
9adbd45d | 2442 | ret = rw_verify_area(WRITE, req->file, &kiocb->ki_pos, iov_count); |
2b188cc1 | 2443 | if (!ret) { |
9bf7933f RP |
2444 | ssize_t ret2; |
2445 | ||
2b188cc1 JA |
2446 | /* |
2447 | * Open-code file_start_write here to grab freeze protection, | |
2448 | * which will be released by another thread in | |
2449 | * io_complete_rw(). Fool lockdep by telling it the lock got | |
2450 | * released so that it doesn't complain about the held lock when | |
2451 | * we return to userspace. | |
2452 | */ | |
491381ce | 2453 | if (req->flags & REQ_F_ISREG) { |
9adbd45d | 2454 | __sb_start_write(file_inode(req->file)->i_sb, |
2b188cc1 | 2455 | SB_FREEZE_WRITE, true); |
9adbd45d | 2456 | __sb_writers_release(file_inode(req->file)->i_sb, |
2b188cc1 JA |
2457 | SB_FREEZE_WRITE); |
2458 | } | |
2459 | kiocb->ki_flags |= IOCB_WRITE; | |
9bf7933f | 2460 | |
9adbd45d JA |
2461 | if (req->file->f_op->write_iter) |
2462 | ret2 = call_write_iter(req->file, kiocb, &iter); | |
32960613 | 2463 | else |
9adbd45d | 2464 | ret2 = loop_rw_iter(WRITE, req->file, kiocb, &iter); |
faac996c JA |
2465 | /* |
2466 | * Raw bdev writes will -EOPNOTSUPP for IOCB_NOWAIT. Just | |
2467 | * retry them without IOCB_NOWAIT. | |
2468 | */ | |
2469 | if (ret2 == -EOPNOTSUPP && (kiocb->ki_flags & IOCB_NOWAIT)) | |
2470 | ret2 = -EAGAIN; | |
f67676d1 | 2471 | if (!force_nonblock || ret2 != -EAGAIN) { |
014db007 | 2472 | kiocb_done(kiocb, ret2); |
f67676d1 JA |
2473 | } else { |
2474 | copy_iov: | |
b7bb4f7d | 2475 | ret = io_setup_async_rw(req, io_size, iovec, |
f67676d1 JA |
2476 | inline_vecs, &iter); |
2477 | if (ret) | |
2478 | goto out_free; | |
29de5f6a JA |
2479 | /* any defer here is final, must blocking retry */ |
2480 | req->flags |= REQ_F_MUST_PUNT; | |
f67676d1 JA |
2481 | return -EAGAIN; |
2482 | } | |
2b188cc1 | 2483 | } |
31b51510 | 2484 | out_free: |
99bc4c38 | 2485 | req->flags &= ~REQ_F_NEED_CLEANUP; |
1e95081c | 2486 | kfree(iovec); |
2b188cc1 JA |
2487 | return ret; |
2488 | } | |
2489 | ||
7d67af2c PB |
2490 | static int io_splice_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2491 | { | |
2492 | struct io_splice* sp = &req->splice; | |
2493 | unsigned int valid_flags = SPLICE_F_FD_IN_FIXED | SPLICE_F_ALL; | |
2494 | int ret; | |
2495 | ||
2496 | if (req->flags & REQ_F_NEED_CLEANUP) | |
2497 | return 0; | |
2498 | ||
2499 | sp->file_in = NULL; | |
2500 | sp->off_in = READ_ONCE(sqe->splice_off_in); | |
2501 | sp->off_out = READ_ONCE(sqe->off); | |
2502 | sp->len = READ_ONCE(sqe->len); | |
2503 | sp->flags = READ_ONCE(sqe->splice_flags); | |
2504 | ||
2505 | if (unlikely(sp->flags & ~valid_flags)) | |
2506 | return -EINVAL; | |
2507 | ||
2508 | ret = io_file_get(NULL, req, READ_ONCE(sqe->splice_fd_in), &sp->file_in, | |
2509 | (sp->flags & SPLICE_F_FD_IN_FIXED)); | |
2510 | if (ret) | |
2511 | return ret; | |
2512 | req->flags |= REQ_F_NEED_CLEANUP; | |
2513 | ||
2514 | if (!S_ISREG(file_inode(sp->file_in)->i_mode)) | |
2515 | req->work.flags |= IO_WQ_WORK_UNBOUND; | |
2516 | ||
2517 | return 0; | |
2518 | } | |
2519 | ||
2520 | static bool io_splice_punt(struct file *file) | |
2521 | { | |
2522 | if (get_pipe_info(file)) | |
2523 | return false; | |
2524 | if (!io_file_supports_async(file)) | |
2525 | return true; | |
2526 | return !(file->f_mode & O_NONBLOCK); | |
2527 | } | |
2528 | ||
014db007 | 2529 | static int io_splice(struct io_kiocb *req, bool force_nonblock) |
7d67af2c PB |
2530 | { |
2531 | struct io_splice *sp = &req->splice; | |
2532 | struct file *in = sp->file_in; | |
2533 | struct file *out = sp->file_out; | |
2534 | unsigned int flags = sp->flags & ~SPLICE_F_FD_IN_FIXED; | |
2535 | loff_t *poff_in, *poff_out; | |
2536 | long ret; | |
2537 | ||
2538 | if (force_nonblock) { | |
2539 | if (io_splice_punt(in) || io_splice_punt(out)) | |
2540 | return -EAGAIN; | |
2541 | flags |= SPLICE_F_NONBLOCK; | |
2542 | } | |
2543 | ||
2544 | poff_in = (sp->off_in == -1) ? NULL : &sp->off_in; | |
2545 | poff_out = (sp->off_out == -1) ? NULL : &sp->off_out; | |
2546 | ret = do_splice(in, poff_in, out, poff_out, sp->len, flags); | |
2547 | if (force_nonblock && ret == -EAGAIN) | |
2548 | return -EAGAIN; | |
2549 | ||
2550 | io_put_file(req, in, (sp->flags & SPLICE_F_FD_IN_FIXED)); | |
2551 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
2552 | ||
2553 | io_cqring_add_event(req, ret); | |
2554 | if (ret != sp->len) | |
2555 | req_set_fail_links(req); | |
014db007 | 2556 | io_put_req(req); |
7d67af2c PB |
2557 | return 0; |
2558 | } | |
2559 | ||
2b188cc1 JA |
2560 | /* |
2561 | * IORING_OP_NOP just posts a completion event, nothing else. | |
2562 | */ | |
78e19bbe | 2563 | static int io_nop(struct io_kiocb *req) |
2b188cc1 JA |
2564 | { |
2565 | struct io_ring_ctx *ctx = req->ctx; | |
2b188cc1 | 2566 | |
def596e9 JA |
2567 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
2568 | return -EINVAL; | |
2569 | ||
78e19bbe | 2570 | io_cqring_add_event(req, 0); |
e65ef56d | 2571 | io_put_req(req); |
2b188cc1 JA |
2572 | return 0; |
2573 | } | |
2574 | ||
3529d8c2 | 2575 | static int io_prep_fsync(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
c992fe29 | 2576 | { |
6b06314c | 2577 | struct io_ring_ctx *ctx = req->ctx; |
c992fe29 | 2578 | |
09bb8394 JA |
2579 | if (!req->file) |
2580 | return -EBADF; | |
c992fe29 | 2581 | |
6b06314c | 2582 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) |
def596e9 | 2583 | return -EINVAL; |
edafccee | 2584 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) |
c992fe29 CH |
2585 | return -EINVAL; |
2586 | ||
8ed8d3c3 JA |
2587 | req->sync.flags = READ_ONCE(sqe->fsync_flags); |
2588 | if (unlikely(req->sync.flags & ~IORING_FSYNC_DATASYNC)) | |
2589 | return -EINVAL; | |
2590 | ||
2591 | req->sync.off = READ_ONCE(sqe->off); | |
2592 | req->sync.len = READ_ONCE(sqe->len); | |
c992fe29 CH |
2593 | return 0; |
2594 | } | |
2595 | ||
8ed8d3c3 JA |
2596 | static bool io_req_cancelled(struct io_kiocb *req) |
2597 | { | |
2598 | if (req->work.flags & IO_WQ_WORK_CANCEL) { | |
2599 | req_set_fail_links(req); | |
2600 | io_cqring_add_event(req, -ECANCELED); | |
e9fd9396 | 2601 | io_put_req(req); |
8ed8d3c3 JA |
2602 | return true; |
2603 | } | |
2604 | ||
2605 | return false; | |
2606 | } | |
2607 | ||
014db007 | 2608 | static void __io_fsync(struct io_kiocb *req) |
8ed8d3c3 | 2609 | { |
8ed8d3c3 | 2610 | loff_t end = req->sync.off + req->sync.len; |
8ed8d3c3 JA |
2611 | int ret; |
2612 | ||
9adbd45d | 2613 | ret = vfs_fsync_range(req->file, req->sync.off, |
8ed8d3c3 JA |
2614 | end > 0 ? end : LLONG_MAX, |
2615 | req->sync.flags & IORING_FSYNC_DATASYNC); | |
2616 | if (ret < 0) | |
2617 | req_set_fail_links(req); | |
2618 | io_cqring_add_event(req, ret); | |
014db007 | 2619 | io_put_req(req); |
5ea62161 PB |
2620 | } |
2621 | ||
2622 | static void io_fsync_finish(struct io_wq_work **workptr) | |
2623 | { | |
2624 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
5ea62161 PB |
2625 | |
2626 | if (io_req_cancelled(req)) | |
2627 | return; | |
014db007 | 2628 | __io_fsync(req); |
e9fd9396 | 2629 | io_steal_work(req, workptr); |
8ed8d3c3 JA |
2630 | } |
2631 | ||
014db007 | 2632 | static int io_fsync(struct io_kiocb *req, bool force_nonblock) |
c992fe29 | 2633 | { |
c992fe29 | 2634 | /* fsync always requires a blocking context */ |
8ed8d3c3 | 2635 | if (force_nonblock) { |
8ed8d3c3 | 2636 | req->work.func = io_fsync_finish; |
c992fe29 | 2637 | return -EAGAIN; |
8ed8d3c3 | 2638 | } |
014db007 | 2639 | __io_fsync(req); |
c992fe29 CH |
2640 | return 0; |
2641 | } | |
2642 | ||
014db007 | 2643 | static void __io_fallocate(struct io_kiocb *req) |
8ed8d3c3 | 2644 | { |
8ed8d3c3 JA |
2645 | int ret; |
2646 | ||
d63d1b5e JA |
2647 | ret = vfs_fallocate(req->file, req->sync.mode, req->sync.off, |
2648 | req->sync.len); | |
8ed8d3c3 JA |
2649 | if (ret < 0) |
2650 | req_set_fail_links(req); | |
2651 | io_cqring_add_event(req, ret); | |
014db007 | 2652 | io_put_req(req); |
5ea62161 PB |
2653 | } |
2654 | ||
2655 | static void io_fallocate_finish(struct io_wq_work **workptr) | |
2656 | { | |
2657 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
5ea62161 | 2658 | |
594506fe PB |
2659 | if (io_req_cancelled(req)) |
2660 | return; | |
014db007 | 2661 | __io_fallocate(req); |
e9fd9396 | 2662 | io_steal_work(req, workptr); |
5d17b4a4 JA |
2663 | } |
2664 | ||
d63d1b5e JA |
2665 | static int io_fallocate_prep(struct io_kiocb *req, |
2666 | const struct io_uring_sqe *sqe) | |
2667 | { | |
2668 | if (sqe->ioprio || sqe->buf_index || sqe->rw_flags) | |
2669 | return -EINVAL; | |
2670 | ||
2671 | req->sync.off = READ_ONCE(sqe->off); | |
2672 | req->sync.len = READ_ONCE(sqe->addr); | |
2673 | req->sync.mode = READ_ONCE(sqe->len); | |
2674 | return 0; | |
2675 | } | |
2676 | ||
014db007 | 2677 | static int io_fallocate(struct io_kiocb *req, bool force_nonblock) |
5d17b4a4 | 2678 | { |
d63d1b5e | 2679 | /* fallocate always requiring blocking context */ |
8ed8d3c3 | 2680 | if (force_nonblock) { |
d63d1b5e | 2681 | req->work.func = io_fallocate_finish; |
5d17b4a4 | 2682 | return -EAGAIN; |
8ed8d3c3 | 2683 | } |
5d17b4a4 | 2684 | |
014db007 | 2685 | __io_fallocate(req); |
5d17b4a4 JA |
2686 | return 0; |
2687 | } | |
2688 | ||
15b71abe | 2689 | static int io_openat_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
b7bb4f7d | 2690 | { |
f8748881 | 2691 | const char __user *fname; |
15b71abe | 2692 | int ret; |
b7bb4f7d | 2693 | |
15b71abe JA |
2694 | if (sqe->ioprio || sqe->buf_index) |
2695 | return -EINVAL; | |
cf3040ca JA |
2696 | if (sqe->flags & IOSQE_FIXED_FILE) |
2697 | return -EBADF; | |
0bdbdd08 PB |
2698 | if (req->flags & REQ_F_NEED_CLEANUP) |
2699 | return 0; | |
03b1230c | 2700 | |
15b71abe | 2701 | req->open.dfd = READ_ONCE(sqe->fd); |
c12cedf2 | 2702 | req->open.how.mode = READ_ONCE(sqe->len); |
f8748881 | 2703 | fname = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
c12cedf2 | 2704 | req->open.how.flags = READ_ONCE(sqe->open_flags); |
3529d8c2 | 2705 | |
f8748881 | 2706 | req->open.filename = getname(fname); |
15b71abe JA |
2707 | if (IS_ERR(req->open.filename)) { |
2708 | ret = PTR_ERR(req->open.filename); | |
2709 | req->open.filename = NULL; | |
2710 | return ret; | |
2711 | } | |
3529d8c2 | 2712 | |
8fef80bf | 2713 | req->flags |= REQ_F_NEED_CLEANUP; |
15b71abe | 2714 | return 0; |
03b1230c JA |
2715 | } |
2716 | ||
cebdb986 | 2717 | static int io_openat2_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
aa1fa28f | 2718 | { |
cebdb986 JA |
2719 | struct open_how __user *how; |
2720 | const char __user *fname; | |
2721 | size_t len; | |
0fa03c62 JA |
2722 | int ret; |
2723 | ||
cebdb986 | 2724 | if (sqe->ioprio || sqe->buf_index) |
0fa03c62 | 2725 | return -EINVAL; |
cf3040ca JA |
2726 | if (sqe->flags & IOSQE_FIXED_FILE) |
2727 | return -EBADF; | |
0bdbdd08 PB |
2728 | if (req->flags & REQ_F_NEED_CLEANUP) |
2729 | return 0; | |
0fa03c62 | 2730 | |
cebdb986 JA |
2731 | req->open.dfd = READ_ONCE(sqe->fd); |
2732 | fname = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
2733 | how = u64_to_user_ptr(READ_ONCE(sqe->addr2)); | |
2734 | len = READ_ONCE(sqe->len); | |
0fa03c62 | 2735 | |
cebdb986 JA |
2736 | if (len < OPEN_HOW_SIZE_VER0) |
2737 | return -EINVAL; | |
3529d8c2 | 2738 | |
cebdb986 JA |
2739 | ret = copy_struct_from_user(&req->open.how, sizeof(req->open.how), how, |
2740 | len); | |
2741 | if (ret) | |
2742 | return ret; | |
3529d8c2 | 2743 | |
cebdb986 JA |
2744 | if (!(req->open.how.flags & O_PATH) && force_o_largefile()) |
2745 | req->open.how.flags |= O_LARGEFILE; | |
0fa03c62 | 2746 | |
cebdb986 JA |
2747 | req->open.filename = getname(fname); |
2748 | if (IS_ERR(req->open.filename)) { | |
2749 | ret = PTR_ERR(req->open.filename); | |
2750 | req->open.filename = NULL; | |
2751 | return ret; | |
2752 | } | |
2753 | ||
8fef80bf | 2754 | req->flags |= REQ_F_NEED_CLEANUP; |
cebdb986 JA |
2755 | return 0; |
2756 | } | |
2757 | ||
014db007 | 2758 | static int io_openat2(struct io_kiocb *req, bool force_nonblock) |
15b71abe JA |
2759 | { |
2760 | struct open_flags op; | |
15b71abe JA |
2761 | struct file *file; |
2762 | int ret; | |
2763 | ||
f86cd20c | 2764 | if (force_nonblock) |
15b71abe | 2765 | return -EAGAIN; |
15b71abe | 2766 | |
cebdb986 | 2767 | ret = build_open_flags(&req->open.how, &op); |
15b71abe JA |
2768 | if (ret) |
2769 | goto err; | |
2770 | ||
cebdb986 | 2771 | ret = get_unused_fd_flags(req->open.how.flags); |
15b71abe JA |
2772 | if (ret < 0) |
2773 | goto err; | |
2774 | ||
2775 | file = do_filp_open(req->open.dfd, req->open.filename, &op); | |
2776 | if (IS_ERR(file)) { | |
2777 | put_unused_fd(ret); | |
2778 | ret = PTR_ERR(file); | |
2779 | } else { | |
2780 | fsnotify_open(file); | |
2781 | fd_install(ret, file); | |
2782 | } | |
2783 | err: | |
2784 | putname(req->open.filename); | |
8fef80bf | 2785 | req->flags &= ~REQ_F_NEED_CLEANUP; |
15b71abe JA |
2786 | if (ret < 0) |
2787 | req_set_fail_links(req); | |
2788 | io_cqring_add_event(req, ret); | |
014db007 | 2789 | io_put_req(req); |
15b71abe JA |
2790 | return 0; |
2791 | } | |
2792 | ||
014db007 | 2793 | static int io_openat(struct io_kiocb *req, bool force_nonblock) |
cebdb986 JA |
2794 | { |
2795 | req->open.how = build_open_how(req->open.how.flags, req->open.how.mode); | |
014db007 | 2796 | return io_openat2(req, force_nonblock); |
cebdb986 JA |
2797 | } |
2798 | ||
ddf0322d JA |
2799 | static int io_provide_buffers_prep(struct io_kiocb *req, |
2800 | const struct io_uring_sqe *sqe) | |
2801 | { | |
2802 | struct io_provide_buf *p = &req->pbuf; | |
2803 | u64 tmp; | |
2804 | ||
2805 | if (sqe->ioprio || sqe->rw_flags) | |
2806 | return -EINVAL; | |
2807 | ||
2808 | tmp = READ_ONCE(sqe->fd); | |
2809 | if (!tmp || tmp > USHRT_MAX) | |
2810 | return -E2BIG; | |
2811 | p->nbufs = tmp; | |
2812 | p->addr = READ_ONCE(sqe->addr); | |
2813 | p->len = READ_ONCE(sqe->len); | |
2814 | ||
2815 | if (!access_ok(u64_to_user_ptr(p->addr), p->len)) | |
2816 | return -EFAULT; | |
2817 | ||
2818 | p->bgid = READ_ONCE(sqe->buf_group); | |
2819 | tmp = READ_ONCE(sqe->off); | |
2820 | if (tmp > USHRT_MAX) | |
2821 | return -E2BIG; | |
2822 | p->bid = tmp; | |
2823 | return 0; | |
2824 | } | |
2825 | ||
2826 | static int io_add_buffers(struct io_provide_buf *pbuf, struct io_buffer **head) | |
2827 | { | |
2828 | struct io_buffer *buf; | |
2829 | u64 addr = pbuf->addr; | |
2830 | int i, bid = pbuf->bid; | |
2831 | ||
2832 | for (i = 0; i < pbuf->nbufs; i++) { | |
2833 | buf = kmalloc(sizeof(*buf), GFP_KERNEL); | |
2834 | if (!buf) | |
2835 | break; | |
2836 | ||
2837 | buf->addr = addr; | |
2838 | buf->len = pbuf->len; | |
2839 | buf->bid = bid; | |
2840 | addr += pbuf->len; | |
2841 | bid++; | |
2842 | if (!*head) { | |
2843 | INIT_LIST_HEAD(&buf->list); | |
2844 | *head = buf; | |
2845 | } else { | |
2846 | list_add_tail(&buf->list, &(*head)->list); | |
2847 | } | |
2848 | } | |
2849 | ||
2850 | return i ? i : -ENOMEM; | |
2851 | } | |
2852 | ||
2853 | static void io_ring_submit_unlock(struct io_ring_ctx *ctx, bool needs_lock) | |
2854 | { | |
2855 | if (needs_lock) | |
2856 | mutex_unlock(&ctx->uring_lock); | |
2857 | } | |
2858 | ||
2859 | static void io_ring_submit_lock(struct io_ring_ctx *ctx, bool needs_lock) | |
2860 | { | |
2861 | /* | |
2862 | * "Normal" inline submissions always hold the uring_lock, since we | |
2863 | * grab it from the system call. Same is true for the SQPOLL offload. | |
2864 | * The only exception is when we've detached the request and issue it | |
2865 | * from an async worker thread, grab the lock for that case. | |
2866 | */ | |
2867 | if (needs_lock) | |
2868 | mutex_lock(&ctx->uring_lock); | |
2869 | } | |
2870 | ||
2871 | static int io_provide_buffers(struct io_kiocb *req, bool force_nonblock) | |
2872 | { | |
2873 | struct io_provide_buf *p = &req->pbuf; | |
2874 | struct io_ring_ctx *ctx = req->ctx; | |
2875 | struct io_buffer *head, *list; | |
2876 | int ret = 0; | |
2877 | ||
2878 | io_ring_submit_lock(ctx, !force_nonblock); | |
2879 | ||
2880 | lockdep_assert_held(&ctx->uring_lock); | |
2881 | ||
2882 | list = head = idr_find(&ctx->io_buffer_idr, p->bgid); | |
2883 | ||
2884 | ret = io_add_buffers(p, &head); | |
2885 | if (ret < 0) | |
2886 | goto out; | |
2887 | ||
2888 | if (!list) { | |
2889 | ret = idr_alloc(&ctx->io_buffer_idr, head, p->bgid, p->bgid + 1, | |
2890 | GFP_KERNEL); | |
2891 | if (ret < 0) { | |
2892 | while (!list_empty(&head->list)) { | |
2893 | struct io_buffer *buf; | |
2894 | ||
2895 | buf = list_first_entry(&head->list, | |
2896 | struct io_buffer, list); | |
2897 | list_del(&buf->list); | |
2898 | kfree(buf); | |
2899 | } | |
2900 | kfree(head); | |
2901 | goto out; | |
2902 | } | |
2903 | } | |
2904 | out: | |
2905 | io_ring_submit_unlock(ctx, !force_nonblock); | |
2906 | if (ret < 0) | |
2907 | req_set_fail_links(req); | |
2908 | io_cqring_add_event(req, ret); | |
2909 | io_put_req(req); | |
2910 | return 0; | |
2911 | } | |
2912 | ||
3e4827b0 JA |
2913 | static int io_epoll_ctl_prep(struct io_kiocb *req, |
2914 | const struct io_uring_sqe *sqe) | |
2915 | { | |
2916 | #if defined(CONFIG_EPOLL) | |
2917 | if (sqe->ioprio || sqe->buf_index) | |
2918 | return -EINVAL; | |
2919 | ||
2920 | req->epoll.epfd = READ_ONCE(sqe->fd); | |
2921 | req->epoll.op = READ_ONCE(sqe->len); | |
2922 | req->epoll.fd = READ_ONCE(sqe->off); | |
2923 | ||
2924 | if (ep_op_has_event(req->epoll.op)) { | |
2925 | struct epoll_event __user *ev; | |
2926 | ||
2927 | ev = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
2928 | if (copy_from_user(&req->epoll.event, ev, sizeof(*ev))) | |
2929 | return -EFAULT; | |
2930 | } | |
2931 | ||
2932 | return 0; | |
2933 | #else | |
2934 | return -EOPNOTSUPP; | |
2935 | #endif | |
2936 | } | |
2937 | ||
014db007 | 2938 | static int io_epoll_ctl(struct io_kiocb *req, bool force_nonblock) |
3e4827b0 JA |
2939 | { |
2940 | #if defined(CONFIG_EPOLL) | |
2941 | struct io_epoll *ie = &req->epoll; | |
2942 | int ret; | |
2943 | ||
2944 | ret = do_epoll_ctl(ie->epfd, ie->op, ie->fd, &ie->event, force_nonblock); | |
2945 | if (force_nonblock && ret == -EAGAIN) | |
2946 | return -EAGAIN; | |
2947 | ||
2948 | if (ret < 0) | |
2949 | req_set_fail_links(req); | |
2950 | io_cqring_add_event(req, ret); | |
014db007 | 2951 | io_put_req(req); |
3e4827b0 JA |
2952 | return 0; |
2953 | #else | |
2954 | return -EOPNOTSUPP; | |
2955 | #endif | |
2956 | } | |
2957 | ||
c1ca757b JA |
2958 | static int io_madvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2959 | { | |
2960 | #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU) | |
2961 | if (sqe->ioprio || sqe->buf_index || sqe->off) | |
2962 | return -EINVAL; | |
2963 | ||
2964 | req->madvise.addr = READ_ONCE(sqe->addr); | |
2965 | req->madvise.len = READ_ONCE(sqe->len); | |
2966 | req->madvise.advice = READ_ONCE(sqe->fadvise_advice); | |
2967 | return 0; | |
2968 | #else | |
2969 | return -EOPNOTSUPP; | |
2970 | #endif | |
2971 | } | |
2972 | ||
014db007 | 2973 | static int io_madvise(struct io_kiocb *req, bool force_nonblock) |
c1ca757b JA |
2974 | { |
2975 | #if defined(CONFIG_ADVISE_SYSCALLS) && defined(CONFIG_MMU) | |
2976 | struct io_madvise *ma = &req->madvise; | |
2977 | int ret; | |
2978 | ||
2979 | if (force_nonblock) | |
2980 | return -EAGAIN; | |
2981 | ||
2982 | ret = do_madvise(ma->addr, ma->len, ma->advice); | |
2983 | if (ret < 0) | |
2984 | req_set_fail_links(req); | |
2985 | io_cqring_add_event(req, ret); | |
014db007 | 2986 | io_put_req(req); |
c1ca757b JA |
2987 | return 0; |
2988 | #else | |
2989 | return -EOPNOTSUPP; | |
2990 | #endif | |
2991 | } | |
2992 | ||
4840e418 JA |
2993 | static int io_fadvise_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2994 | { | |
2995 | if (sqe->ioprio || sqe->buf_index || sqe->addr) | |
2996 | return -EINVAL; | |
2997 | ||
2998 | req->fadvise.offset = READ_ONCE(sqe->off); | |
2999 | req->fadvise.len = READ_ONCE(sqe->len); | |
3000 | req->fadvise.advice = READ_ONCE(sqe->fadvise_advice); | |
3001 | return 0; | |
3002 | } | |
3003 | ||
014db007 | 3004 | static int io_fadvise(struct io_kiocb *req, bool force_nonblock) |
4840e418 JA |
3005 | { |
3006 | struct io_fadvise *fa = &req->fadvise; | |
3007 | int ret; | |
3008 | ||
3e69426d JA |
3009 | if (force_nonblock) { |
3010 | switch (fa->advice) { | |
3011 | case POSIX_FADV_NORMAL: | |
3012 | case POSIX_FADV_RANDOM: | |
3013 | case POSIX_FADV_SEQUENTIAL: | |
3014 | break; | |
3015 | default: | |
3016 | return -EAGAIN; | |
3017 | } | |
3018 | } | |
4840e418 JA |
3019 | |
3020 | ret = vfs_fadvise(req->file, fa->offset, fa->len, fa->advice); | |
3021 | if (ret < 0) | |
3022 | req_set_fail_links(req); | |
3023 | io_cqring_add_event(req, ret); | |
014db007 | 3024 | io_put_req(req); |
4840e418 JA |
3025 | return 0; |
3026 | } | |
3027 | ||
eddc7ef5 JA |
3028 | static int io_statx_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
3029 | { | |
f8748881 | 3030 | const char __user *fname; |
eddc7ef5 JA |
3031 | unsigned lookup_flags; |
3032 | int ret; | |
3033 | ||
3034 | if (sqe->ioprio || sqe->buf_index) | |
3035 | return -EINVAL; | |
cf3040ca JA |
3036 | if (sqe->flags & IOSQE_FIXED_FILE) |
3037 | return -EBADF; | |
0bdbdd08 PB |
3038 | if (req->flags & REQ_F_NEED_CLEANUP) |
3039 | return 0; | |
eddc7ef5 JA |
3040 | |
3041 | req->open.dfd = READ_ONCE(sqe->fd); | |
3042 | req->open.mask = READ_ONCE(sqe->len); | |
f8748881 | 3043 | fname = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
eddc7ef5 | 3044 | req->open.buffer = u64_to_user_ptr(READ_ONCE(sqe->addr2)); |
c12cedf2 | 3045 | req->open.how.flags = READ_ONCE(sqe->statx_flags); |
eddc7ef5 | 3046 | |
c12cedf2 | 3047 | if (vfs_stat_set_lookup_flags(&lookup_flags, req->open.how.flags)) |
eddc7ef5 JA |
3048 | return -EINVAL; |
3049 | ||
f8748881 | 3050 | req->open.filename = getname_flags(fname, lookup_flags, NULL); |
eddc7ef5 JA |
3051 | if (IS_ERR(req->open.filename)) { |
3052 | ret = PTR_ERR(req->open.filename); | |
3053 | req->open.filename = NULL; | |
3054 | return ret; | |
3055 | } | |
3056 | ||
8fef80bf | 3057 | req->flags |= REQ_F_NEED_CLEANUP; |
eddc7ef5 JA |
3058 | return 0; |
3059 | } | |
3060 | ||
014db007 | 3061 | static int io_statx(struct io_kiocb *req, bool force_nonblock) |
eddc7ef5 JA |
3062 | { |
3063 | struct io_open *ctx = &req->open; | |
3064 | unsigned lookup_flags; | |
3065 | struct path path; | |
3066 | struct kstat stat; | |
3067 | int ret; | |
3068 | ||
3069 | if (force_nonblock) | |
3070 | return -EAGAIN; | |
3071 | ||
c12cedf2 | 3072 | if (vfs_stat_set_lookup_flags(&lookup_flags, ctx->how.flags)) |
eddc7ef5 JA |
3073 | return -EINVAL; |
3074 | ||
3075 | retry: | |
3076 | /* filename_lookup() drops it, keep a reference */ | |
3077 | ctx->filename->refcnt++; | |
3078 | ||
3079 | ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags, &path, | |
3080 | NULL); | |
3081 | if (ret) | |
3082 | goto err; | |
3083 | ||
c12cedf2 | 3084 | ret = vfs_getattr(&path, &stat, ctx->mask, ctx->how.flags); |
eddc7ef5 JA |
3085 | path_put(&path); |
3086 | if (retry_estale(ret, lookup_flags)) { | |
3087 | lookup_flags |= LOOKUP_REVAL; | |
3088 | goto retry; | |
3089 | } | |
3090 | if (!ret) | |
3091 | ret = cp_statx(&stat, ctx->buffer); | |
3092 | err: | |
3093 | putname(ctx->filename); | |
8fef80bf | 3094 | req->flags &= ~REQ_F_NEED_CLEANUP; |
eddc7ef5 JA |
3095 | if (ret < 0) |
3096 | req_set_fail_links(req); | |
3097 | io_cqring_add_event(req, ret); | |
014db007 | 3098 | io_put_req(req); |
eddc7ef5 JA |
3099 | return 0; |
3100 | } | |
3101 | ||
b5dba59e JA |
3102 | static int io_close_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
3103 | { | |
3104 | /* | |
3105 | * If we queue this for async, it must not be cancellable. That would | |
3106 | * leave the 'file' in an undeterminate state. | |
3107 | */ | |
3108 | req->work.flags |= IO_WQ_WORK_NO_CANCEL; | |
3109 | ||
3110 | if (sqe->ioprio || sqe->off || sqe->addr || sqe->len || | |
3111 | sqe->rw_flags || sqe->buf_index) | |
3112 | return -EINVAL; | |
3113 | if (sqe->flags & IOSQE_FIXED_FILE) | |
cf3040ca | 3114 | return -EBADF; |
b5dba59e JA |
3115 | |
3116 | req->close.fd = READ_ONCE(sqe->fd); | |
3117 | if (req->file->f_op == &io_uring_fops || | |
b14cca0c | 3118 | req->close.fd == req->ctx->ring_fd) |
b5dba59e JA |
3119 | return -EBADF; |
3120 | ||
3121 | return 0; | |
3122 | } | |
3123 | ||
a93b3331 | 3124 | /* only called when __close_fd_get_file() is done */ |
014db007 | 3125 | static void __io_close_finish(struct io_kiocb *req) |
a93b3331 PB |
3126 | { |
3127 | int ret; | |
3128 | ||
3129 | ret = filp_close(req->close.put_file, req->work.files); | |
3130 | if (ret < 0) | |
3131 | req_set_fail_links(req); | |
3132 | io_cqring_add_event(req, ret); | |
3133 | fput(req->close.put_file); | |
014db007 | 3134 | io_put_req(req); |
a93b3331 PB |
3135 | } |
3136 | ||
b5dba59e JA |
3137 | static void io_close_finish(struct io_wq_work **workptr) |
3138 | { | |
3139 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
b5dba59e | 3140 | |
7fbeb95d | 3141 | /* not cancellable, don't do io_req_cancelled() */ |
014db007 | 3142 | __io_close_finish(req); |
e9fd9396 | 3143 | io_steal_work(req, workptr); |
b5dba59e JA |
3144 | } |
3145 | ||
014db007 | 3146 | static int io_close(struct io_kiocb *req, bool force_nonblock) |
b5dba59e JA |
3147 | { |
3148 | int ret; | |
3149 | ||
3150 | req->close.put_file = NULL; | |
3151 | ret = __close_fd_get_file(req->close.fd, &req->close.put_file); | |
3152 | if (ret < 0) | |
3153 | return ret; | |
3154 | ||
3155 | /* if the file has a flush method, be safe and punt to async */ | |
a2100672 | 3156 | if (req->close.put_file->f_op->flush && force_nonblock) { |
594506fe PB |
3157 | /* submission ref will be dropped, take it for async */ |
3158 | refcount_inc(&req->refs); | |
3159 | ||
a2100672 PB |
3160 | req->work.func = io_close_finish; |
3161 | /* | |
3162 | * Do manual async queue here to avoid grabbing files - we don't | |
3163 | * need the files, and it'll cause io_close_finish() to close | |
3164 | * the file again and cause a double CQE entry for this request | |
3165 | */ | |
3166 | io_queue_async_work(req); | |
3167 | return 0; | |
3168 | } | |
b5dba59e JA |
3169 | |
3170 | /* | |
3171 | * No ->flush(), safely close from here and just punt the | |
3172 | * fput() to async context. | |
3173 | */ | |
014db007 | 3174 | __io_close_finish(req); |
a93b3331 | 3175 | return 0; |
b5dba59e JA |
3176 | } |
3177 | ||
3529d8c2 | 3178 | static int io_prep_sfr(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
5d17b4a4 JA |
3179 | { |
3180 | struct io_ring_ctx *ctx = req->ctx; | |
5d17b4a4 JA |
3181 | |
3182 | if (!req->file) | |
3183 | return -EBADF; | |
5d17b4a4 JA |
3184 | |
3185 | if (unlikely(ctx->flags & IORING_SETUP_IOPOLL)) | |
3186 | return -EINVAL; | |
3187 | if (unlikely(sqe->addr || sqe->ioprio || sqe->buf_index)) | |
3188 | return -EINVAL; | |
3189 | ||
8ed8d3c3 JA |
3190 | req->sync.off = READ_ONCE(sqe->off); |
3191 | req->sync.len = READ_ONCE(sqe->len); | |
3192 | req->sync.flags = READ_ONCE(sqe->sync_range_flags); | |
8ed8d3c3 JA |
3193 | return 0; |
3194 | } | |
3195 | ||
014db007 | 3196 | static void __io_sync_file_range(struct io_kiocb *req) |
8ed8d3c3 | 3197 | { |
8ed8d3c3 JA |
3198 | int ret; |
3199 | ||
9adbd45d | 3200 | ret = sync_file_range(req->file, req->sync.off, req->sync.len, |
8ed8d3c3 JA |
3201 | req->sync.flags); |
3202 | if (ret < 0) | |
3203 | req_set_fail_links(req); | |
3204 | io_cqring_add_event(req, ret); | |
014db007 | 3205 | io_put_req(req); |
5ea62161 PB |
3206 | } |
3207 | ||
3208 | ||
3209 | static void io_sync_file_range_finish(struct io_wq_work **workptr) | |
3210 | { | |
3211 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
3212 | struct io_kiocb *nxt = NULL; | |
3213 | ||
3214 | if (io_req_cancelled(req)) | |
3215 | return; | |
014db007 | 3216 | __io_sync_file_range(req); |
594506fe | 3217 | io_put_req(req); /* put submission ref */ |
8ed8d3c3 | 3218 | if (nxt) |
78912934 | 3219 | io_wq_assign_next(workptr, nxt); |
5d17b4a4 JA |
3220 | } |
3221 | ||
014db007 | 3222 | static int io_sync_file_range(struct io_kiocb *req, bool force_nonblock) |
5d17b4a4 | 3223 | { |
5d17b4a4 | 3224 | /* sync_file_range always requires a blocking context */ |
8ed8d3c3 | 3225 | if (force_nonblock) { |
8ed8d3c3 | 3226 | req->work.func = io_sync_file_range_finish; |
5d17b4a4 | 3227 | return -EAGAIN; |
8ed8d3c3 | 3228 | } |
5d17b4a4 | 3229 | |
014db007 | 3230 | __io_sync_file_range(req); |
5d17b4a4 JA |
3231 | return 0; |
3232 | } | |
3233 | ||
02d27d89 PB |
3234 | static int io_setup_async_msg(struct io_kiocb *req, |
3235 | struct io_async_msghdr *kmsg) | |
3236 | { | |
3237 | if (req->io) | |
3238 | return -EAGAIN; | |
3239 | if (io_alloc_async_ctx(req)) { | |
3240 | if (kmsg->iov != kmsg->fast_iov) | |
3241 | kfree(kmsg->iov); | |
3242 | return -ENOMEM; | |
3243 | } | |
3244 | req->flags |= REQ_F_NEED_CLEANUP; | |
3245 | memcpy(&req->io->msg, kmsg, sizeof(*kmsg)); | |
3246 | return -EAGAIN; | |
3247 | } | |
3248 | ||
3529d8c2 | 3249 | static int io_sendmsg_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
03b1230c | 3250 | { |
0fa03c62 | 3251 | #if defined(CONFIG_NET) |
e47293fd | 3252 | struct io_sr_msg *sr = &req->sr_msg; |
3529d8c2 | 3253 | struct io_async_ctx *io = req->io; |
99bc4c38 | 3254 | int ret; |
03b1230c | 3255 | |
e47293fd JA |
3256 | sr->msg_flags = READ_ONCE(sqe->msg_flags); |
3257 | sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
fddaface | 3258 | sr->len = READ_ONCE(sqe->len); |
3529d8c2 | 3259 | |
d8768362 JA |
3260 | #ifdef CONFIG_COMPAT |
3261 | if (req->ctx->compat) | |
3262 | sr->msg_flags |= MSG_CMSG_COMPAT; | |
3263 | #endif | |
3264 | ||
fddaface | 3265 | if (!io || req->opcode == IORING_OP_SEND) |
3529d8c2 | 3266 | return 0; |
5f798bea PB |
3267 | /* iovec is already imported */ |
3268 | if (req->flags & REQ_F_NEED_CLEANUP) | |
3269 | return 0; | |
3529d8c2 | 3270 | |
d9688565 | 3271 | io->msg.iov = io->msg.fast_iov; |
99bc4c38 | 3272 | ret = sendmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags, |
e47293fd | 3273 | &io->msg.iov); |
99bc4c38 PB |
3274 | if (!ret) |
3275 | req->flags |= REQ_F_NEED_CLEANUP; | |
3276 | return ret; | |
03b1230c | 3277 | #else |
e47293fd | 3278 | return -EOPNOTSUPP; |
03b1230c JA |
3279 | #endif |
3280 | } | |
3281 | ||
014db007 | 3282 | static int io_sendmsg(struct io_kiocb *req, bool force_nonblock) |
aa1fa28f | 3283 | { |
03b1230c | 3284 | #if defined(CONFIG_NET) |
0b416c3e | 3285 | struct io_async_msghdr *kmsg = NULL; |
0fa03c62 JA |
3286 | struct socket *sock; |
3287 | int ret; | |
3288 | ||
3289 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
3290 | return -EINVAL; | |
3291 | ||
3292 | sock = sock_from_file(req->file, &ret); | |
3293 | if (sock) { | |
b7bb4f7d | 3294 | struct io_async_ctx io; |
0fa03c62 JA |
3295 | unsigned flags; |
3296 | ||
03b1230c | 3297 | if (req->io) { |
0b416c3e | 3298 | kmsg = &req->io->msg; |
b537916c | 3299 | kmsg->msg.msg_name = &req->io->msg.addr; |
0b416c3e JA |
3300 | /* if iov is set, it's allocated already */ |
3301 | if (!kmsg->iov) | |
3302 | kmsg->iov = kmsg->fast_iov; | |
3303 | kmsg->msg.msg_iter.iov = kmsg->iov; | |
03b1230c | 3304 | } else { |
3529d8c2 JA |
3305 | struct io_sr_msg *sr = &req->sr_msg; |
3306 | ||
0b416c3e | 3307 | kmsg = &io.msg; |
b537916c | 3308 | kmsg->msg.msg_name = &io.msg.addr; |
3529d8c2 JA |
3309 | |
3310 | io.msg.iov = io.msg.fast_iov; | |
3311 | ret = sendmsg_copy_msghdr(&io.msg.msg, sr->msg, | |
3312 | sr->msg_flags, &io.msg.iov); | |
03b1230c | 3313 | if (ret) |
3529d8c2 | 3314 | return ret; |
03b1230c | 3315 | } |
0fa03c62 | 3316 | |
e47293fd JA |
3317 | flags = req->sr_msg.msg_flags; |
3318 | if (flags & MSG_DONTWAIT) | |
3319 | req->flags |= REQ_F_NOWAIT; | |
3320 | else if (force_nonblock) | |
3321 | flags |= MSG_DONTWAIT; | |
3322 | ||
0b416c3e | 3323 | ret = __sys_sendmsg_sock(sock, &kmsg->msg, flags); |
02d27d89 PB |
3324 | if (force_nonblock && ret == -EAGAIN) |
3325 | return io_setup_async_msg(req, kmsg); | |
441cdbd5 JA |
3326 | if (ret == -ERESTARTSYS) |
3327 | ret = -EINTR; | |
0fa03c62 JA |
3328 | } |
3329 | ||
1e95081c | 3330 | if (kmsg && kmsg->iov != kmsg->fast_iov) |
0b416c3e | 3331 | kfree(kmsg->iov); |
99bc4c38 | 3332 | req->flags &= ~REQ_F_NEED_CLEANUP; |
78e19bbe | 3333 | io_cqring_add_event(req, ret); |
4e88d6e7 JA |
3334 | if (ret < 0) |
3335 | req_set_fail_links(req); | |
014db007 | 3336 | io_put_req(req); |
5d17b4a4 | 3337 | return 0; |
03b1230c JA |
3338 | #else |
3339 | return -EOPNOTSUPP; | |
aa1fa28f | 3340 | #endif |
03b1230c | 3341 | } |
aa1fa28f | 3342 | |
014db007 | 3343 | static int io_send(struct io_kiocb *req, bool force_nonblock) |
fddaface JA |
3344 | { |
3345 | #if defined(CONFIG_NET) | |
3346 | struct socket *sock; | |
3347 | int ret; | |
3348 | ||
3349 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
3350 | return -EINVAL; | |
3351 | ||
3352 | sock = sock_from_file(req->file, &ret); | |
3353 | if (sock) { | |
3354 | struct io_sr_msg *sr = &req->sr_msg; | |
3355 | struct msghdr msg; | |
3356 | struct iovec iov; | |
3357 | unsigned flags; | |
3358 | ||
3359 | ret = import_single_range(WRITE, sr->buf, sr->len, &iov, | |
3360 | &msg.msg_iter); | |
3361 | if (ret) | |
3362 | return ret; | |
3363 | ||
3364 | msg.msg_name = NULL; | |
3365 | msg.msg_control = NULL; | |
3366 | msg.msg_controllen = 0; | |
3367 | msg.msg_namelen = 0; | |
3368 | ||
3369 | flags = req->sr_msg.msg_flags; | |
3370 | if (flags & MSG_DONTWAIT) | |
3371 | req->flags |= REQ_F_NOWAIT; | |
3372 | else if (force_nonblock) | |
3373 | flags |= MSG_DONTWAIT; | |
3374 | ||
0b7b21e4 JA |
3375 | msg.msg_flags = flags; |
3376 | ret = sock_sendmsg(sock, &msg); | |
fddaface JA |
3377 | if (force_nonblock && ret == -EAGAIN) |
3378 | return -EAGAIN; | |
3379 | if (ret == -ERESTARTSYS) | |
3380 | ret = -EINTR; | |
3381 | } | |
3382 | ||
3383 | io_cqring_add_event(req, ret); | |
3384 | if (ret < 0) | |
3385 | req_set_fail_links(req); | |
014db007 | 3386 | io_put_req(req); |
fddaface JA |
3387 | return 0; |
3388 | #else | |
3389 | return -EOPNOTSUPP; | |
3390 | #endif | |
3391 | } | |
3392 | ||
3529d8c2 JA |
3393 | static int io_recvmsg_prep(struct io_kiocb *req, |
3394 | const struct io_uring_sqe *sqe) | |
aa1fa28f JA |
3395 | { |
3396 | #if defined(CONFIG_NET) | |
e47293fd | 3397 | struct io_sr_msg *sr = &req->sr_msg; |
3529d8c2 | 3398 | struct io_async_ctx *io = req->io; |
99bc4c38 | 3399 | int ret; |
3529d8c2 JA |
3400 | |
3401 | sr->msg_flags = READ_ONCE(sqe->msg_flags); | |
3402 | sr->msg = u64_to_user_ptr(READ_ONCE(sqe->addr)); | |
0b7b21e4 | 3403 | sr->len = READ_ONCE(sqe->len); |
06b76d44 | 3404 | |
d8768362 JA |
3405 | #ifdef CONFIG_COMPAT |
3406 | if (req->ctx->compat) | |
3407 | sr->msg_flags |= MSG_CMSG_COMPAT; | |
3408 | #endif | |
3409 | ||
fddaface | 3410 | if (!io || req->opcode == IORING_OP_RECV) |
06b76d44 | 3411 | return 0; |
5f798bea PB |
3412 | /* iovec is already imported */ |
3413 | if (req->flags & REQ_F_NEED_CLEANUP) | |
3414 | return 0; | |
03b1230c | 3415 | |
d9688565 | 3416 | io->msg.iov = io->msg.fast_iov; |
99bc4c38 | 3417 | ret = recvmsg_copy_msghdr(&io->msg.msg, sr->msg, sr->msg_flags, |
e47293fd | 3418 | &io->msg.uaddr, &io->msg.iov); |
99bc4c38 PB |
3419 | if (!ret) |
3420 | req->flags |= REQ_F_NEED_CLEANUP; | |
3421 | return ret; | |
aa1fa28f | 3422 | #else |
e47293fd | 3423 | return -EOPNOTSUPP; |
aa1fa28f JA |
3424 | #endif |
3425 | } | |
3426 | ||
014db007 | 3427 | static int io_recvmsg(struct io_kiocb *req, bool force_nonblock) |
aa1fa28f JA |
3428 | { |
3429 | #if defined(CONFIG_NET) | |
0b416c3e | 3430 | struct io_async_msghdr *kmsg = NULL; |
03b1230c JA |
3431 | struct socket *sock; |
3432 | int ret; | |
3433 | ||
3434 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
3435 | return -EINVAL; | |
3436 | ||
3437 | sock = sock_from_file(req->file, &ret); | |
3438 | if (sock) { | |
b7bb4f7d | 3439 | struct io_async_ctx io; |
03b1230c JA |
3440 | unsigned flags; |
3441 | ||
03b1230c | 3442 | if (req->io) { |
0b416c3e | 3443 | kmsg = &req->io->msg; |
b537916c | 3444 | kmsg->msg.msg_name = &req->io->msg.addr; |
0b416c3e JA |
3445 | /* if iov is set, it's allocated already */ |
3446 | if (!kmsg->iov) | |
3447 | kmsg->iov = kmsg->fast_iov; | |
3448 | kmsg->msg.msg_iter.iov = kmsg->iov; | |
03b1230c | 3449 | } else { |
3529d8c2 JA |
3450 | struct io_sr_msg *sr = &req->sr_msg; |
3451 | ||
0b416c3e | 3452 | kmsg = &io.msg; |
b537916c | 3453 | kmsg->msg.msg_name = &io.msg.addr; |
3529d8c2 JA |
3454 | |
3455 | io.msg.iov = io.msg.fast_iov; | |
3456 | ret = recvmsg_copy_msghdr(&io.msg.msg, sr->msg, | |
3457 | sr->msg_flags, &io.msg.uaddr, | |
3458 | &io.msg.iov); | |
03b1230c | 3459 | if (ret) |
3529d8c2 | 3460 | return ret; |
03b1230c JA |
3461 | } |
3462 | ||
e47293fd JA |
3463 | flags = req->sr_msg.msg_flags; |
3464 | if (flags & MSG_DONTWAIT) | |
3465 | req->flags |= REQ_F_NOWAIT; | |
3466 | else if (force_nonblock) | |
3467 | flags |= MSG_DONTWAIT; | |
3468 | ||
3469 | ret = __sys_recvmsg_sock(sock, &kmsg->msg, req->sr_msg.msg, | |
3470 | kmsg->uaddr, flags); | |
02d27d89 PB |
3471 | if (force_nonblock && ret == -EAGAIN) |
3472 | return io_setup_async_msg(req, kmsg); | |
03b1230c JA |
3473 | if (ret == -ERESTARTSYS) |
3474 | ret = -EINTR; | |
3475 | } | |
3476 | ||
1e95081c | 3477 | if (kmsg && kmsg->iov != kmsg->fast_iov) |
0b416c3e | 3478 | kfree(kmsg->iov); |
99bc4c38 | 3479 | req->flags &= ~REQ_F_NEED_CLEANUP; |
03b1230c | 3480 | io_cqring_add_event(req, ret); |
4e88d6e7 JA |
3481 | if (ret < 0) |
3482 | req_set_fail_links(req); | |
014db007 | 3483 | io_put_req(req); |
03b1230c | 3484 | return 0; |
0fa03c62 JA |
3485 | #else |
3486 | return -EOPNOTSUPP; | |
3487 | #endif | |
3488 | } | |
5d17b4a4 | 3489 | |
014db007 | 3490 | static int io_recv(struct io_kiocb *req, bool force_nonblock) |
fddaface JA |
3491 | { |
3492 | #if defined(CONFIG_NET) | |
3493 | struct socket *sock; | |
3494 | int ret; | |
3495 | ||
3496 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
3497 | return -EINVAL; | |
3498 | ||
3499 | sock = sock_from_file(req->file, &ret); | |
3500 | if (sock) { | |
3501 | struct io_sr_msg *sr = &req->sr_msg; | |
3502 | struct msghdr msg; | |
3503 | struct iovec iov; | |
3504 | unsigned flags; | |
3505 | ||
3506 | ret = import_single_range(READ, sr->buf, sr->len, &iov, | |
3507 | &msg.msg_iter); | |
3508 | if (ret) | |
3509 | return ret; | |
3510 | ||
3511 | msg.msg_name = NULL; | |
3512 | msg.msg_control = NULL; | |
3513 | msg.msg_controllen = 0; | |
3514 | msg.msg_namelen = 0; | |
3515 | msg.msg_iocb = NULL; | |
3516 | msg.msg_flags = 0; | |
3517 | ||
3518 | flags = req->sr_msg.msg_flags; | |
3519 | if (flags & MSG_DONTWAIT) | |
3520 | req->flags |= REQ_F_NOWAIT; | |
3521 | else if (force_nonblock) | |
3522 | flags |= MSG_DONTWAIT; | |
3523 | ||
0b7b21e4 | 3524 | ret = sock_recvmsg(sock, &msg, flags); |
fddaface JA |
3525 | if (force_nonblock && ret == -EAGAIN) |
3526 | return -EAGAIN; | |
3527 | if (ret == -ERESTARTSYS) | |
3528 | ret = -EINTR; | |
3529 | } | |
3530 | ||
3531 | io_cqring_add_event(req, ret); | |
3532 | if (ret < 0) | |
3533 | req_set_fail_links(req); | |
014db007 | 3534 | io_put_req(req); |
fddaface JA |
3535 | return 0; |
3536 | #else | |
3537 | return -EOPNOTSUPP; | |
3538 | #endif | |
3539 | } | |
3540 | ||
3541 | ||
3529d8c2 | 3542 | static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
17f2fe35 JA |
3543 | { |
3544 | #if defined(CONFIG_NET) | |
8ed8d3c3 JA |
3545 | struct io_accept *accept = &req->accept; |
3546 | ||
17f2fe35 JA |
3547 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) |
3548 | return -EINVAL; | |
8042d6ce | 3549 | if (sqe->ioprio || sqe->len || sqe->buf_index) |
17f2fe35 JA |
3550 | return -EINVAL; |
3551 | ||
d55e5f5b JA |
3552 | accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
3553 | accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2)); | |
8ed8d3c3 | 3554 | accept->flags = READ_ONCE(sqe->accept_flags); |
8ed8d3c3 JA |
3555 | return 0; |
3556 | #else | |
3557 | return -EOPNOTSUPP; | |
3558 | #endif | |
3559 | } | |
17f2fe35 | 3560 | |
8ed8d3c3 | 3561 | #if defined(CONFIG_NET) |
014db007 | 3562 | static int __io_accept(struct io_kiocb *req, bool force_nonblock) |
8ed8d3c3 JA |
3563 | { |
3564 | struct io_accept *accept = &req->accept; | |
3565 | unsigned file_flags; | |
3566 | int ret; | |
3567 | ||
3568 | file_flags = force_nonblock ? O_NONBLOCK : 0; | |
3569 | ret = __sys_accept4_file(req->file, file_flags, accept->addr, | |
3570 | accept->addr_len, accept->flags); | |
3571 | if (ret == -EAGAIN && force_nonblock) | |
17f2fe35 | 3572 | return -EAGAIN; |
8e3cca12 JA |
3573 | if (ret == -ERESTARTSYS) |
3574 | ret = -EINTR; | |
4e88d6e7 JA |
3575 | if (ret < 0) |
3576 | req_set_fail_links(req); | |
78e19bbe | 3577 | io_cqring_add_event(req, ret); |
014db007 | 3578 | io_put_req(req); |
17f2fe35 | 3579 | return 0; |
8ed8d3c3 JA |
3580 | } |
3581 | ||
3582 | static void io_accept_finish(struct io_wq_work **workptr) | |
3583 | { | |
3584 | struct io_kiocb *req = container_of(*workptr, struct io_kiocb, work); | |
8ed8d3c3 JA |
3585 | |
3586 | if (io_req_cancelled(req)) | |
3587 | return; | |
014db007 | 3588 | __io_accept(req, false); |
e9fd9396 | 3589 | io_steal_work(req, workptr); |
8ed8d3c3 JA |
3590 | } |
3591 | #endif | |
3592 | ||
014db007 | 3593 | static int io_accept(struct io_kiocb *req, bool force_nonblock) |
8ed8d3c3 JA |
3594 | { |
3595 | #if defined(CONFIG_NET) | |
3596 | int ret; | |
3597 | ||
014db007 | 3598 | ret = __io_accept(req, force_nonblock); |
8ed8d3c3 JA |
3599 | if (ret == -EAGAIN && force_nonblock) { |
3600 | req->work.func = io_accept_finish; | |
8ed8d3c3 JA |
3601 | return -EAGAIN; |
3602 | } | |
3603 | return 0; | |
0fa03c62 JA |
3604 | #else |
3605 | return -EOPNOTSUPP; | |
3606 | #endif | |
3607 | } | |
5d17b4a4 | 3608 | |
3529d8c2 | 3609 | static int io_connect_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
f499a021 JA |
3610 | { |
3611 | #if defined(CONFIG_NET) | |
3529d8c2 JA |
3612 | struct io_connect *conn = &req->connect; |
3613 | struct io_async_ctx *io = req->io; | |
f499a021 | 3614 | |
3fbb51c1 JA |
3615 | if (unlikely(req->ctx->flags & (IORING_SETUP_IOPOLL|IORING_SETUP_SQPOLL))) |
3616 | return -EINVAL; | |
3617 | if (sqe->ioprio || sqe->len || sqe->buf_index || sqe->rw_flags) | |
3618 | return -EINVAL; | |
3619 | ||
3529d8c2 JA |
3620 | conn->addr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
3621 | conn->addr_len = READ_ONCE(sqe->addr2); | |
3622 | ||
3623 | if (!io) | |
3624 | return 0; | |
3625 | ||
3626 | return move_addr_to_kernel(conn->addr, conn->addr_len, | |
3fbb51c1 | 3627 | &io->connect.address); |
f499a021 | 3628 | #else |
3fbb51c1 | 3629 | return -EOPNOTSUPP; |
f499a021 JA |
3630 | #endif |
3631 | } | |
3632 | ||
014db007 | 3633 | static int io_connect(struct io_kiocb *req, bool force_nonblock) |
f8e85cf2 JA |
3634 | { |
3635 | #if defined(CONFIG_NET) | |
f499a021 | 3636 | struct io_async_ctx __io, *io; |
f8e85cf2 | 3637 | unsigned file_flags; |
3fbb51c1 | 3638 | int ret; |
f8e85cf2 | 3639 | |
f499a021 JA |
3640 | if (req->io) { |
3641 | io = req->io; | |
3642 | } else { | |
3529d8c2 JA |
3643 | ret = move_addr_to_kernel(req->connect.addr, |
3644 | req->connect.addr_len, | |
3645 | &__io.connect.address); | |
f499a021 JA |
3646 | if (ret) |
3647 | goto out; | |
3648 | io = &__io; | |
3649 | } | |
3650 | ||
3fbb51c1 JA |
3651 | file_flags = force_nonblock ? O_NONBLOCK : 0; |
3652 | ||
3653 | ret = __sys_connect_file(req->file, &io->connect.address, | |
3654 | req->connect.addr_len, file_flags); | |
87f80d62 | 3655 | if ((ret == -EAGAIN || ret == -EINPROGRESS) && force_nonblock) { |
b7bb4f7d JA |
3656 | if (req->io) |
3657 | return -EAGAIN; | |
3658 | if (io_alloc_async_ctx(req)) { | |
f499a021 JA |
3659 | ret = -ENOMEM; |
3660 | goto out; | |
3661 | } | |
b7bb4f7d | 3662 | memcpy(&req->io->connect, &__io.connect, sizeof(__io.connect)); |
f8e85cf2 | 3663 | return -EAGAIN; |
f499a021 | 3664 | } |
f8e85cf2 JA |
3665 | if (ret == -ERESTARTSYS) |
3666 | ret = -EINTR; | |
f499a021 | 3667 | out: |
4e88d6e7 JA |
3668 | if (ret < 0) |
3669 | req_set_fail_links(req); | |
f8e85cf2 | 3670 | io_cqring_add_event(req, ret); |
014db007 | 3671 | io_put_req(req); |
f8e85cf2 JA |
3672 | return 0; |
3673 | #else | |
3674 | return -EOPNOTSUPP; | |
3675 | #endif | |
3676 | } | |
3677 | ||
d7718a9d JA |
3678 | struct io_poll_table { |
3679 | struct poll_table_struct pt; | |
3680 | struct io_kiocb *req; | |
3681 | int error; | |
3682 | }; | |
3683 | ||
3684 | static void __io_queue_proc(struct io_poll_iocb *poll, struct io_poll_table *pt, | |
3685 | struct wait_queue_head *head) | |
3686 | { | |
3687 | if (unlikely(poll->head)) { | |
3688 | pt->error = -EINVAL; | |
3689 | return; | |
3690 | } | |
3691 | ||
3692 | pt->error = 0; | |
3693 | poll->head = head; | |
3694 | add_wait_queue(head, &poll->wait); | |
3695 | } | |
3696 | ||
3697 | static void io_async_queue_proc(struct file *file, struct wait_queue_head *head, | |
3698 | struct poll_table_struct *p) | |
3699 | { | |
3700 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
3701 | ||
3702 | __io_queue_proc(&pt->req->apoll->poll, pt, head); | |
3703 | } | |
3704 | ||
3705 | static int __io_async_wake(struct io_kiocb *req, struct io_poll_iocb *poll, | |
3706 | __poll_t mask, task_work_func_t func) | |
3707 | { | |
3708 | struct task_struct *tsk; | |
3709 | ||
3710 | /* for instances that support it check for an event match first: */ | |
3711 | if (mask && !(mask & poll->events)) | |
3712 | return 0; | |
3713 | ||
3714 | trace_io_uring_task_add(req->ctx, req->opcode, req->user_data, mask); | |
3715 | ||
3716 | list_del_init(&poll->wait.entry); | |
3717 | ||
3718 | tsk = req->task; | |
3719 | req->result = mask; | |
3720 | init_task_work(&req->task_work, func); | |
3721 | /* | |
3722 | * If this fails, then the task is exiting. If that is the case, then | |
3723 | * the exit check will ultimately cancel these work items. Hence we | |
3724 | * don't need to check here and handle it specifically. | |
3725 | */ | |
3726 | task_work_add(tsk, &req->task_work, true); | |
3727 | wake_up_process(tsk); | |
3728 | return 1; | |
3729 | } | |
3730 | ||
3731 | static void io_async_task_func(struct callback_head *cb) | |
3732 | { | |
3733 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); | |
3734 | struct async_poll *apoll = req->apoll; | |
3735 | struct io_ring_ctx *ctx = req->ctx; | |
3736 | ||
3737 | trace_io_uring_task_run(req->ctx, req->opcode, req->user_data); | |
3738 | ||
3739 | WARN_ON_ONCE(!list_empty(&req->apoll->poll.wait.entry)); | |
3740 | ||
3741 | if (hash_hashed(&req->hash_node)) { | |
3742 | spin_lock_irq(&ctx->completion_lock); | |
3743 | hash_del(&req->hash_node); | |
3744 | spin_unlock_irq(&ctx->completion_lock); | |
3745 | } | |
3746 | ||
3747 | /* restore ->work in case we need to retry again */ | |
3748 | memcpy(&req->work, &apoll->work, sizeof(req->work)); | |
3749 | ||
3750 | __set_current_state(TASK_RUNNING); | |
3751 | mutex_lock(&ctx->uring_lock); | |
3752 | __io_queue_sqe(req, NULL); | |
3753 | mutex_unlock(&ctx->uring_lock); | |
3754 | ||
3755 | kfree(apoll); | |
3756 | } | |
3757 | ||
3758 | static int io_async_wake(struct wait_queue_entry *wait, unsigned mode, int sync, | |
3759 | void *key) | |
3760 | { | |
3761 | struct io_kiocb *req = wait->private; | |
3762 | struct io_poll_iocb *poll = &req->apoll->poll; | |
3763 | ||
3764 | trace_io_uring_poll_wake(req->ctx, req->opcode, req->user_data, | |
3765 | key_to_poll(key)); | |
3766 | ||
3767 | return __io_async_wake(req, poll, key_to_poll(key), io_async_task_func); | |
3768 | } | |
3769 | ||
3770 | static void io_poll_req_insert(struct io_kiocb *req) | |
3771 | { | |
3772 | struct io_ring_ctx *ctx = req->ctx; | |
3773 | struct hlist_head *list; | |
3774 | ||
3775 | list = &ctx->cancel_hash[hash_long(req->user_data, ctx->cancel_hash_bits)]; | |
3776 | hlist_add_head(&req->hash_node, list); | |
3777 | } | |
3778 | ||
3779 | static __poll_t __io_arm_poll_handler(struct io_kiocb *req, | |
3780 | struct io_poll_iocb *poll, | |
3781 | struct io_poll_table *ipt, __poll_t mask, | |
3782 | wait_queue_func_t wake_func) | |
3783 | __acquires(&ctx->completion_lock) | |
3784 | { | |
3785 | struct io_ring_ctx *ctx = req->ctx; | |
3786 | bool cancel = false; | |
3787 | ||
3788 | poll->file = req->file; | |
3789 | poll->head = NULL; | |
3790 | poll->done = poll->canceled = false; | |
3791 | poll->events = mask; | |
3792 | ||
3793 | ipt->pt._key = mask; | |
3794 | ipt->req = req; | |
3795 | ipt->error = -EINVAL; | |
3796 | ||
3797 | INIT_LIST_HEAD(&poll->wait.entry); | |
3798 | init_waitqueue_func_entry(&poll->wait, wake_func); | |
3799 | poll->wait.private = req; | |
3800 | ||
3801 | mask = vfs_poll(req->file, &ipt->pt) & poll->events; | |
3802 | ||
3803 | spin_lock_irq(&ctx->completion_lock); | |
3804 | if (likely(poll->head)) { | |
3805 | spin_lock(&poll->head->lock); | |
3806 | if (unlikely(list_empty(&poll->wait.entry))) { | |
3807 | if (ipt->error) | |
3808 | cancel = true; | |
3809 | ipt->error = 0; | |
3810 | mask = 0; | |
3811 | } | |
3812 | if (mask || ipt->error) | |
3813 | list_del_init(&poll->wait.entry); | |
3814 | else if (cancel) | |
3815 | WRITE_ONCE(poll->canceled, true); | |
3816 | else if (!poll->done) /* actually waiting for an event */ | |
3817 | io_poll_req_insert(req); | |
3818 | spin_unlock(&poll->head->lock); | |
3819 | } | |
3820 | ||
3821 | return mask; | |
3822 | } | |
3823 | ||
3824 | static bool io_arm_poll_handler(struct io_kiocb *req) | |
3825 | { | |
3826 | const struct io_op_def *def = &io_op_defs[req->opcode]; | |
3827 | struct io_ring_ctx *ctx = req->ctx; | |
3828 | struct async_poll *apoll; | |
3829 | struct io_poll_table ipt; | |
3830 | __poll_t mask, ret; | |
3831 | ||
3832 | if (!req->file || !file_can_poll(req->file)) | |
3833 | return false; | |
3834 | if (req->flags & (REQ_F_MUST_PUNT | REQ_F_POLLED)) | |
3835 | return false; | |
3836 | if (!def->pollin && !def->pollout) | |
3837 | return false; | |
3838 | ||
3839 | apoll = kmalloc(sizeof(*apoll), GFP_ATOMIC); | |
3840 | if (unlikely(!apoll)) | |
3841 | return false; | |
3842 | ||
3843 | req->flags |= REQ_F_POLLED; | |
3844 | memcpy(&apoll->work, &req->work, sizeof(req->work)); | |
3845 | ||
3846 | /* | |
3847 | * Don't need a reference here, as we're adding it to the task | |
3848 | * task_works list. If the task exits, the list is pruned. | |
3849 | */ | |
3850 | req->task = current; | |
3851 | req->apoll = apoll; | |
3852 | INIT_HLIST_NODE(&req->hash_node); | |
3853 | ||
8755d97a | 3854 | mask = 0; |
d7718a9d | 3855 | if (def->pollin) |
8755d97a | 3856 | mask |= POLLIN | POLLRDNORM; |
d7718a9d JA |
3857 | if (def->pollout) |
3858 | mask |= POLLOUT | POLLWRNORM; | |
3859 | mask |= POLLERR | POLLPRI; | |
3860 | ||
3861 | ipt.pt._qproc = io_async_queue_proc; | |
3862 | ||
3863 | ret = __io_arm_poll_handler(req, &apoll->poll, &ipt, mask, | |
3864 | io_async_wake); | |
3865 | if (ret) { | |
3866 | ipt.error = 0; | |
3867 | apoll->poll.done = true; | |
3868 | spin_unlock_irq(&ctx->completion_lock); | |
3869 | memcpy(&req->work, &apoll->work, sizeof(req->work)); | |
3870 | kfree(apoll); | |
3871 | return false; | |
3872 | } | |
3873 | spin_unlock_irq(&ctx->completion_lock); | |
3874 | trace_io_uring_poll_arm(ctx, req->opcode, req->user_data, mask, | |
3875 | apoll->poll.events); | |
3876 | return true; | |
3877 | } | |
3878 | ||
3879 | static bool __io_poll_remove_one(struct io_kiocb *req, | |
3880 | struct io_poll_iocb *poll) | |
221c5eb2 | 3881 | { |
b41e9852 | 3882 | bool do_complete = false; |
221c5eb2 JA |
3883 | |
3884 | spin_lock(&poll->head->lock); | |
3885 | WRITE_ONCE(poll->canceled, true); | |
392edb45 JA |
3886 | if (!list_empty(&poll->wait.entry)) { |
3887 | list_del_init(&poll->wait.entry); | |
b41e9852 | 3888 | do_complete = true; |
221c5eb2 JA |
3889 | } |
3890 | spin_unlock(&poll->head->lock); | |
d7718a9d JA |
3891 | return do_complete; |
3892 | } | |
3893 | ||
3894 | static bool io_poll_remove_one(struct io_kiocb *req) | |
3895 | { | |
3896 | bool do_complete; | |
3897 | ||
3898 | if (req->opcode == IORING_OP_POLL_ADD) { | |
3899 | do_complete = __io_poll_remove_one(req, &req->poll); | |
3900 | } else { | |
3901 | /* non-poll requests have submit ref still */ | |
3902 | do_complete = __io_poll_remove_one(req, &req->apoll->poll); | |
3903 | if (do_complete) | |
3904 | io_put_req(req); | |
3905 | } | |
3906 | ||
78076bb6 | 3907 | hash_del(&req->hash_node); |
d7718a9d | 3908 | |
b41e9852 JA |
3909 | if (do_complete) { |
3910 | io_cqring_fill_event(req, -ECANCELED); | |
3911 | io_commit_cqring(req->ctx); | |
3912 | req->flags |= REQ_F_COMP_LOCKED; | |
3913 | io_put_req(req); | |
3914 | } | |
3915 | ||
3916 | return do_complete; | |
221c5eb2 JA |
3917 | } |
3918 | ||
3919 | static void io_poll_remove_all(struct io_ring_ctx *ctx) | |
3920 | { | |
78076bb6 | 3921 | struct hlist_node *tmp; |
221c5eb2 | 3922 | struct io_kiocb *req; |
78076bb6 | 3923 | int i; |
221c5eb2 JA |
3924 | |
3925 | spin_lock_irq(&ctx->completion_lock); | |
78076bb6 JA |
3926 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { |
3927 | struct hlist_head *list; | |
3928 | ||
3929 | list = &ctx->cancel_hash[i]; | |
3930 | hlist_for_each_entry_safe(req, tmp, list, hash_node) | |
3931 | io_poll_remove_one(req); | |
221c5eb2 JA |
3932 | } |
3933 | spin_unlock_irq(&ctx->completion_lock); | |
b41e9852 JA |
3934 | |
3935 | io_cqring_ev_posted(ctx); | |
221c5eb2 JA |
3936 | } |
3937 | ||
47f46768 JA |
3938 | static int io_poll_cancel(struct io_ring_ctx *ctx, __u64 sqe_addr) |
3939 | { | |
78076bb6 | 3940 | struct hlist_head *list; |
47f46768 JA |
3941 | struct io_kiocb *req; |
3942 | ||
78076bb6 JA |
3943 | list = &ctx->cancel_hash[hash_long(sqe_addr, ctx->cancel_hash_bits)]; |
3944 | hlist_for_each_entry(req, list, hash_node) { | |
b41e9852 JA |
3945 | if (sqe_addr != req->user_data) |
3946 | continue; | |
3947 | if (io_poll_remove_one(req)) | |
eac406c6 | 3948 | return 0; |
b41e9852 | 3949 | return -EALREADY; |
47f46768 JA |
3950 | } |
3951 | ||
3952 | return -ENOENT; | |
3953 | } | |
3954 | ||
3529d8c2 JA |
3955 | static int io_poll_remove_prep(struct io_kiocb *req, |
3956 | const struct io_uring_sqe *sqe) | |
0969e783 | 3957 | { |
0969e783 JA |
3958 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
3959 | return -EINVAL; | |
3960 | if (sqe->ioprio || sqe->off || sqe->len || sqe->buf_index || | |
3961 | sqe->poll_events) | |
3962 | return -EINVAL; | |
3963 | ||
3964 | req->poll.addr = READ_ONCE(sqe->addr); | |
0969e783 JA |
3965 | return 0; |
3966 | } | |
3967 | ||
221c5eb2 JA |
3968 | /* |
3969 | * Find a running poll command that matches one specified in sqe->addr, | |
3970 | * and remove it if found. | |
3971 | */ | |
fc4df999 | 3972 | static int io_poll_remove(struct io_kiocb *req) |
221c5eb2 JA |
3973 | { |
3974 | struct io_ring_ctx *ctx = req->ctx; | |
0969e783 | 3975 | u64 addr; |
47f46768 | 3976 | int ret; |
221c5eb2 | 3977 | |
0969e783 | 3978 | addr = req->poll.addr; |
221c5eb2 | 3979 | spin_lock_irq(&ctx->completion_lock); |
0969e783 | 3980 | ret = io_poll_cancel(ctx, addr); |
221c5eb2 JA |
3981 | spin_unlock_irq(&ctx->completion_lock); |
3982 | ||
78e19bbe | 3983 | io_cqring_add_event(req, ret); |
4e88d6e7 JA |
3984 | if (ret < 0) |
3985 | req_set_fail_links(req); | |
e65ef56d | 3986 | io_put_req(req); |
221c5eb2 JA |
3987 | return 0; |
3988 | } | |
3989 | ||
b0dd8a41 | 3990 | static void io_poll_complete(struct io_kiocb *req, __poll_t mask, int error) |
221c5eb2 | 3991 | { |
a197f664 JL |
3992 | struct io_ring_ctx *ctx = req->ctx; |
3993 | ||
8c838788 | 3994 | req->poll.done = true; |
b0a20349 | 3995 | io_cqring_fill_event(req, error ? error : mangle_poll(mask)); |
8c838788 | 3996 | io_commit_cqring(ctx); |
221c5eb2 JA |
3997 | } |
3998 | ||
b41e9852 | 3999 | static void io_poll_task_handler(struct io_kiocb *req, struct io_kiocb **nxt) |
221c5eb2 | 4000 | { |
221c5eb2 | 4001 | struct io_ring_ctx *ctx = req->ctx; |
221c5eb2 | 4002 | |
221c5eb2 | 4003 | spin_lock_irq(&ctx->completion_lock); |
78076bb6 | 4004 | hash_del(&req->hash_node); |
b41e9852 JA |
4005 | io_poll_complete(req, req->result, 0); |
4006 | req->flags |= REQ_F_COMP_LOCKED; | |
4007 | io_put_req_find_next(req, nxt); | |
e94f141b JA |
4008 | spin_unlock_irq(&ctx->completion_lock); |
4009 | ||
4010 | io_cqring_ev_posted(ctx); | |
e94f141b JA |
4011 | } |
4012 | ||
b41e9852 | 4013 | static void io_poll_task_func(struct callback_head *cb) |
f0b493e6 | 4014 | { |
b41e9852 JA |
4015 | struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work); |
4016 | struct io_kiocb *nxt = NULL; | |
f0b493e6 | 4017 | |
b41e9852 | 4018 | io_poll_task_handler(req, &nxt); |
d7718a9d JA |
4019 | if (nxt) { |
4020 | struct io_ring_ctx *ctx = nxt->ctx; | |
4021 | ||
4022 | mutex_lock(&ctx->uring_lock); | |
b41e9852 | 4023 | __io_queue_sqe(nxt, NULL); |
d7718a9d JA |
4024 | mutex_unlock(&ctx->uring_lock); |
4025 | } | |
f0b493e6 JA |
4026 | } |
4027 | ||
221c5eb2 JA |
4028 | static int io_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync, |
4029 | void *key) | |
4030 | { | |
c2f2eb7d JA |
4031 | struct io_kiocb *req = wait->private; |
4032 | struct io_poll_iocb *poll = &req->poll; | |
221c5eb2 | 4033 | |
d7718a9d | 4034 | return __io_async_wake(req, poll, key_to_poll(key), io_poll_task_func); |
221c5eb2 JA |
4035 | } |
4036 | ||
221c5eb2 JA |
4037 | static void io_poll_queue_proc(struct file *file, struct wait_queue_head *head, |
4038 | struct poll_table_struct *p) | |
4039 | { | |
4040 | struct io_poll_table *pt = container_of(p, struct io_poll_table, pt); | |
4041 | ||
d7718a9d | 4042 | __io_queue_proc(&pt->req->poll, pt, head); |
eac406c6 JA |
4043 | } |
4044 | ||
3529d8c2 | 4045 | static int io_poll_add_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
221c5eb2 JA |
4046 | { |
4047 | struct io_poll_iocb *poll = &req->poll; | |
221c5eb2 | 4048 | u16 events; |
221c5eb2 JA |
4049 | |
4050 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) | |
4051 | return -EINVAL; | |
4052 | if (sqe->addr || sqe->ioprio || sqe->off || sqe->len || sqe->buf_index) | |
4053 | return -EINVAL; | |
09bb8394 JA |
4054 | if (!poll->file) |
4055 | return -EBADF; | |
221c5eb2 | 4056 | |
221c5eb2 JA |
4057 | events = READ_ONCE(sqe->poll_events); |
4058 | poll->events = demangle_poll(events) | EPOLLERR | EPOLLHUP; | |
b41e9852 | 4059 | |
d7718a9d JA |
4060 | /* |
4061 | * Don't need a reference here, as we're adding it to the task | |
4062 | * task_works list. If the task exits, the list is pruned. | |
4063 | */ | |
b41e9852 | 4064 | req->task = current; |
0969e783 JA |
4065 | return 0; |
4066 | } | |
4067 | ||
014db007 | 4068 | static int io_poll_add(struct io_kiocb *req) |
0969e783 JA |
4069 | { |
4070 | struct io_poll_iocb *poll = &req->poll; | |
4071 | struct io_ring_ctx *ctx = req->ctx; | |
4072 | struct io_poll_table ipt; | |
0969e783 | 4073 | __poll_t mask; |
0969e783 | 4074 | |
78076bb6 | 4075 | INIT_HLIST_NODE(&req->hash_node); |
36703247 | 4076 | INIT_LIST_HEAD(&req->list); |
d7718a9d | 4077 | ipt.pt._qproc = io_poll_queue_proc; |
36703247 | 4078 | |
d7718a9d JA |
4079 | mask = __io_arm_poll_handler(req, &req->poll, &ipt, poll->events, |
4080 | io_poll_wake); | |
221c5eb2 | 4081 | |
8c838788 | 4082 | if (mask) { /* no async, we'd stolen it */ |
221c5eb2 | 4083 | ipt.error = 0; |
b0dd8a41 | 4084 | io_poll_complete(req, mask, 0); |
221c5eb2 | 4085 | } |
221c5eb2 JA |
4086 | spin_unlock_irq(&ctx->completion_lock); |
4087 | ||
8c838788 JA |
4088 | if (mask) { |
4089 | io_cqring_ev_posted(ctx); | |
014db007 | 4090 | io_put_req(req); |
221c5eb2 | 4091 | } |
8c838788 | 4092 | return ipt.error; |
221c5eb2 JA |
4093 | } |
4094 | ||
5262f567 JA |
4095 | static enum hrtimer_restart io_timeout_fn(struct hrtimer *timer) |
4096 | { | |
ad8a48ac JA |
4097 | struct io_timeout_data *data = container_of(timer, |
4098 | struct io_timeout_data, timer); | |
4099 | struct io_kiocb *req = data->req; | |
4100 | struct io_ring_ctx *ctx = req->ctx; | |
5262f567 JA |
4101 | unsigned long flags; |
4102 | ||
5262f567 JA |
4103 | atomic_inc(&ctx->cq_timeouts); |
4104 | ||
4105 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
ef03681a | 4106 | /* |
11365043 JA |
4107 | * We could be racing with timeout deletion. If the list is empty, |
4108 | * then timeout lookup already found it and will be handling it. | |
ef03681a | 4109 | */ |
842f9612 | 4110 | if (!list_empty(&req->list)) { |
11365043 | 4111 | struct io_kiocb *prev; |
5262f567 | 4112 | |
11365043 JA |
4113 | /* |
4114 | * Adjust the reqs sequence before the current one because it | |
d195a66e | 4115 | * will consume a slot in the cq_ring and the cq_tail |
11365043 JA |
4116 | * pointer will be increased, otherwise other timeout reqs may |
4117 | * return in advance without waiting for enough wait_nr. | |
4118 | */ | |
4119 | prev = req; | |
4120 | list_for_each_entry_continue_reverse(prev, &ctx->timeout_list, list) | |
4121 | prev->sequence++; | |
11365043 | 4122 | list_del_init(&req->list); |
11365043 | 4123 | } |
5262f567 | 4124 | |
78e19bbe | 4125 | io_cqring_fill_event(req, -ETIME); |
5262f567 JA |
4126 | io_commit_cqring(ctx); |
4127 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
4128 | ||
4129 | io_cqring_ev_posted(ctx); | |
4e88d6e7 | 4130 | req_set_fail_links(req); |
5262f567 JA |
4131 | io_put_req(req); |
4132 | return HRTIMER_NORESTART; | |
4133 | } | |
4134 | ||
47f46768 JA |
4135 | static int io_timeout_cancel(struct io_ring_ctx *ctx, __u64 user_data) |
4136 | { | |
4137 | struct io_kiocb *req; | |
4138 | int ret = -ENOENT; | |
4139 | ||
4140 | list_for_each_entry(req, &ctx->timeout_list, list) { | |
4141 | if (user_data == req->user_data) { | |
4142 | list_del_init(&req->list); | |
4143 | ret = 0; | |
4144 | break; | |
4145 | } | |
4146 | } | |
4147 | ||
4148 | if (ret == -ENOENT) | |
4149 | return ret; | |
4150 | ||
2d28390a | 4151 | ret = hrtimer_try_to_cancel(&req->io->timeout.timer); |
47f46768 JA |
4152 | if (ret == -1) |
4153 | return -EALREADY; | |
4154 | ||
4e88d6e7 | 4155 | req_set_fail_links(req); |
47f46768 JA |
4156 | io_cqring_fill_event(req, -ECANCELED); |
4157 | io_put_req(req); | |
4158 | return 0; | |
4159 | } | |
4160 | ||
3529d8c2 JA |
4161 | static int io_timeout_remove_prep(struct io_kiocb *req, |
4162 | const struct io_uring_sqe *sqe) | |
b29472ee | 4163 | { |
b29472ee JA |
4164 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
4165 | return -EINVAL; | |
4166 | if (sqe->flags || sqe->ioprio || sqe->buf_index || sqe->len) | |
4167 | return -EINVAL; | |
4168 | ||
4169 | req->timeout.addr = READ_ONCE(sqe->addr); | |
4170 | req->timeout.flags = READ_ONCE(sqe->timeout_flags); | |
4171 | if (req->timeout.flags) | |
4172 | return -EINVAL; | |
4173 | ||
b29472ee JA |
4174 | return 0; |
4175 | } | |
4176 | ||
11365043 JA |
4177 | /* |
4178 | * Remove or update an existing timeout command | |
4179 | */ | |
fc4df999 | 4180 | static int io_timeout_remove(struct io_kiocb *req) |
11365043 JA |
4181 | { |
4182 | struct io_ring_ctx *ctx = req->ctx; | |
47f46768 | 4183 | int ret; |
11365043 | 4184 | |
11365043 | 4185 | spin_lock_irq(&ctx->completion_lock); |
b29472ee | 4186 | ret = io_timeout_cancel(ctx, req->timeout.addr); |
11365043 | 4187 | |
47f46768 | 4188 | io_cqring_fill_event(req, ret); |
11365043 JA |
4189 | io_commit_cqring(ctx); |
4190 | spin_unlock_irq(&ctx->completion_lock); | |
5262f567 | 4191 | io_cqring_ev_posted(ctx); |
4e88d6e7 JA |
4192 | if (ret < 0) |
4193 | req_set_fail_links(req); | |
ec9c02ad | 4194 | io_put_req(req); |
11365043 | 4195 | return 0; |
5262f567 JA |
4196 | } |
4197 | ||
3529d8c2 | 4198 | static int io_timeout_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
2d28390a | 4199 | bool is_timeout_link) |
5262f567 | 4200 | { |
ad8a48ac | 4201 | struct io_timeout_data *data; |
a41525ab | 4202 | unsigned flags; |
5262f567 | 4203 | |
ad8a48ac | 4204 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
5262f567 | 4205 | return -EINVAL; |
ad8a48ac | 4206 | if (sqe->ioprio || sqe->buf_index || sqe->len != 1) |
a41525ab | 4207 | return -EINVAL; |
2d28390a JA |
4208 | if (sqe->off && is_timeout_link) |
4209 | return -EINVAL; | |
a41525ab JA |
4210 | flags = READ_ONCE(sqe->timeout_flags); |
4211 | if (flags & ~IORING_TIMEOUT_ABS) | |
5262f567 | 4212 | return -EINVAL; |
bdf20073 | 4213 | |
26a61679 JA |
4214 | req->timeout.count = READ_ONCE(sqe->off); |
4215 | ||
3529d8c2 | 4216 | if (!req->io && io_alloc_async_ctx(req)) |
26a61679 JA |
4217 | return -ENOMEM; |
4218 | ||
4219 | data = &req->io->timeout; | |
ad8a48ac | 4220 | data->req = req; |
ad8a48ac JA |
4221 | req->flags |= REQ_F_TIMEOUT; |
4222 | ||
4223 | if (get_timespec64(&data->ts, u64_to_user_ptr(sqe->addr))) | |
5262f567 JA |
4224 | return -EFAULT; |
4225 | ||
11365043 | 4226 | if (flags & IORING_TIMEOUT_ABS) |
ad8a48ac | 4227 | data->mode = HRTIMER_MODE_ABS; |
11365043 | 4228 | else |
ad8a48ac | 4229 | data->mode = HRTIMER_MODE_REL; |
11365043 | 4230 | |
ad8a48ac JA |
4231 | hrtimer_init(&data->timer, CLOCK_MONOTONIC, data->mode); |
4232 | return 0; | |
4233 | } | |
4234 | ||
fc4df999 | 4235 | static int io_timeout(struct io_kiocb *req) |
ad8a48ac JA |
4236 | { |
4237 | unsigned count; | |
4238 | struct io_ring_ctx *ctx = req->ctx; | |
4239 | struct io_timeout_data *data; | |
4240 | struct list_head *entry; | |
4241 | unsigned span = 0; | |
ad8a48ac | 4242 | |
2d28390a | 4243 | data = &req->io->timeout; |
93bd25bb | 4244 | |
5262f567 JA |
4245 | /* |
4246 | * sqe->off holds how many events that need to occur for this | |
93bd25bb JA |
4247 | * timeout event to be satisfied. If it isn't set, then this is |
4248 | * a pure timeout request, sequence isn't used. | |
5262f567 | 4249 | */ |
26a61679 | 4250 | count = req->timeout.count; |
93bd25bb JA |
4251 | if (!count) { |
4252 | req->flags |= REQ_F_TIMEOUT_NOSEQ; | |
4253 | spin_lock_irq(&ctx->completion_lock); | |
4254 | entry = ctx->timeout_list.prev; | |
4255 | goto add; | |
4256 | } | |
5262f567 JA |
4257 | |
4258 | req->sequence = ctx->cached_sq_head + count - 1; | |
2d28390a | 4259 | data->seq_offset = count; |
5262f567 JA |
4260 | |
4261 | /* | |
4262 | * Insertion sort, ensuring the first entry in the list is always | |
4263 | * the one we need first. | |
4264 | */ | |
5262f567 JA |
4265 | spin_lock_irq(&ctx->completion_lock); |
4266 | list_for_each_prev(entry, &ctx->timeout_list) { | |
4267 | struct io_kiocb *nxt = list_entry(entry, struct io_kiocb, list); | |
5da0fb1a | 4268 | unsigned nxt_sq_head; |
4269 | long long tmp, tmp_nxt; | |
2d28390a | 4270 | u32 nxt_offset = nxt->io->timeout.seq_offset; |
5262f567 | 4271 | |
93bd25bb JA |
4272 | if (nxt->flags & REQ_F_TIMEOUT_NOSEQ) |
4273 | continue; | |
4274 | ||
5da0fb1a | 4275 | /* |
4276 | * Since cached_sq_head + count - 1 can overflow, use type long | |
4277 | * long to store it. | |
4278 | */ | |
4279 | tmp = (long long)ctx->cached_sq_head + count - 1; | |
cc42e0ac PB |
4280 | nxt_sq_head = nxt->sequence - nxt_offset + 1; |
4281 | tmp_nxt = (long long)nxt_sq_head + nxt_offset - 1; | |
5da0fb1a | 4282 | |
4283 | /* | |
4284 | * cached_sq_head may overflow, and it will never overflow twice | |
4285 | * once there is some timeout req still be valid. | |
4286 | */ | |
4287 | if (ctx->cached_sq_head < nxt_sq_head) | |
8b07a65a | 4288 | tmp += UINT_MAX; |
5da0fb1a | 4289 | |
a1f58ba4 | 4290 | if (tmp > tmp_nxt) |
5262f567 | 4291 | break; |
a1f58ba4 | 4292 | |
4293 | /* | |
4294 | * Sequence of reqs after the insert one and itself should | |
4295 | * be adjusted because each timeout req consumes a slot. | |
4296 | */ | |
4297 | span++; | |
4298 | nxt->sequence++; | |
5262f567 | 4299 | } |
a1f58ba4 | 4300 | req->sequence -= span; |
93bd25bb | 4301 | add: |
5262f567 | 4302 | list_add(&req->list, entry); |
ad8a48ac JA |
4303 | data->timer.function = io_timeout_fn; |
4304 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), data->mode); | |
5262f567 | 4305 | spin_unlock_irq(&ctx->completion_lock); |
5262f567 JA |
4306 | return 0; |
4307 | } | |
5262f567 | 4308 | |
62755e35 JA |
4309 | static bool io_cancel_cb(struct io_wq_work *work, void *data) |
4310 | { | |
4311 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
4312 | ||
4313 | return req->user_data == (unsigned long) data; | |
4314 | } | |
4315 | ||
e977d6d3 | 4316 | static int io_async_cancel_one(struct io_ring_ctx *ctx, void *sqe_addr) |
62755e35 | 4317 | { |
62755e35 | 4318 | enum io_wq_cancel cancel_ret; |
62755e35 JA |
4319 | int ret = 0; |
4320 | ||
62755e35 JA |
4321 | cancel_ret = io_wq_cancel_cb(ctx->io_wq, io_cancel_cb, sqe_addr); |
4322 | switch (cancel_ret) { | |
4323 | case IO_WQ_CANCEL_OK: | |
4324 | ret = 0; | |
4325 | break; | |
4326 | case IO_WQ_CANCEL_RUNNING: | |
4327 | ret = -EALREADY; | |
4328 | break; | |
4329 | case IO_WQ_CANCEL_NOTFOUND: | |
4330 | ret = -ENOENT; | |
4331 | break; | |
4332 | } | |
4333 | ||
e977d6d3 JA |
4334 | return ret; |
4335 | } | |
4336 | ||
47f46768 JA |
4337 | static void io_async_find_and_cancel(struct io_ring_ctx *ctx, |
4338 | struct io_kiocb *req, __u64 sqe_addr, | |
014db007 | 4339 | int success_ret) |
47f46768 JA |
4340 | { |
4341 | unsigned long flags; | |
4342 | int ret; | |
4343 | ||
4344 | ret = io_async_cancel_one(ctx, (void *) (unsigned long) sqe_addr); | |
4345 | if (ret != -ENOENT) { | |
4346 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
4347 | goto done; | |
4348 | } | |
4349 | ||
4350 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
4351 | ret = io_timeout_cancel(ctx, sqe_addr); | |
4352 | if (ret != -ENOENT) | |
4353 | goto done; | |
4354 | ret = io_poll_cancel(ctx, sqe_addr); | |
4355 | done: | |
b0dd8a41 JA |
4356 | if (!ret) |
4357 | ret = success_ret; | |
47f46768 JA |
4358 | io_cqring_fill_event(req, ret); |
4359 | io_commit_cqring(ctx); | |
4360 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
4361 | io_cqring_ev_posted(ctx); | |
4362 | ||
4e88d6e7 JA |
4363 | if (ret < 0) |
4364 | req_set_fail_links(req); | |
014db007 | 4365 | io_put_req(req); |
47f46768 JA |
4366 | } |
4367 | ||
3529d8c2 JA |
4368 | static int io_async_cancel_prep(struct io_kiocb *req, |
4369 | const struct io_uring_sqe *sqe) | |
e977d6d3 | 4370 | { |
fbf23849 | 4371 | if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL)) |
e977d6d3 JA |
4372 | return -EINVAL; |
4373 | if (sqe->flags || sqe->ioprio || sqe->off || sqe->len || | |
4374 | sqe->cancel_flags) | |
4375 | return -EINVAL; | |
4376 | ||
fbf23849 JA |
4377 | req->cancel.addr = READ_ONCE(sqe->addr); |
4378 | return 0; | |
4379 | } | |
4380 | ||
014db007 | 4381 | static int io_async_cancel(struct io_kiocb *req) |
fbf23849 JA |
4382 | { |
4383 | struct io_ring_ctx *ctx = req->ctx; | |
fbf23849 | 4384 | |
014db007 | 4385 | io_async_find_and_cancel(ctx, req, req->cancel.addr, 0); |
5262f567 JA |
4386 | return 0; |
4387 | } | |
4388 | ||
05f3fb3c JA |
4389 | static int io_files_update_prep(struct io_kiocb *req, |
4390 | const struct io_uring_sqe *sqe) | |
4391 | { | |
4392 | if (sqe->flags || sqe->ioprio || sqe->rw_flags) | |
4393 | return -EINVAL; | |
4394 | ||
4395 | req->files_update.offset = READ_ONCE(sqe->off); | |
4396 | req->files_update.nr_args = READ_ONCE(sqe->len); | |
4397 | if (!req->files_update.nr_args) | |
4398 | return -EINVAL; | |
4399 | req->files_update.arg = READ_ONCE(sqe->addr); | |
4400 | return 0; | |
4401 | } | |
4402 | ||
4403 | static int io_files_update(struct io_kiocb *req, bool force_nonblock) | |
fbf23849 JA |
4404 | { |
4405 | struct io_ring_ctx *ctx = req->ctx; | |
05f3fb3c JA |
4406 | struct io_uring_files_update up; |
4407 | int ret; | |
fbf23849 | 4408 | |
f86cd20c | 4409 | if (force_nonblock) |
05f3fb3c | 4410 | return -EAGAIN; |
05f3fb3c JA |
4411 | |
4412 | up.offset = req->files_update.offset; | |
4413 | up.fds = req->files_update.arg; | |
4414 | ||
4415 | mutex_lock(&ctx->uring_lock); | |
4416 | ret = __io_sqe_files_update(ctx, &up, req->files_update.nr_args); | |
4417 | mutex_unlock(&ctx->uring_lock); | |
4418 | ||
4419 | if (ret < 0) | |
4420 | req_set_fail_links(req); | |
4421 | io_cqring_add_event(req, ret); | |
4422 | io_put_req(req); | |
5262f567 JA |
4423 | return 0; |
4424 | } | |
4425 | ||
3529d8c2 JA |
4426 | static int io_req_defer_prep(struct io_kiocb *req, |
4427 | const struct io_uring_sqe *sqe) | |
f67676d1 | 4428 | { |
e781573e | 4429 | ssize_t ret = 0; |
f67676d1 | 4430 | |
f86cd20c JA |
4431 | if (io_op_defs[req->opcode].file_table) { |
4432 | ret = io_grab_files(req); | |
4433 | if (unlikely(ret)) | |
4434 | return ret; | |
4435 | } | |
4436 | ||
cccf0ee8 JA |
4437 | io_req_work_grab_env(req, &io_op_defs[req->opcode]); |
4438 | ||
d625c6ee | 4439 | switch (req->opcode) { |
e781573e JA |
4440 | case IORING_OP_NOP: |
4441 | break; | |
f67676d1 JA |
4442 | case IORING_OP_READV: |
4443 | case IORING_OP_READ_FIXED: | |
3a6820f2 | 4444 | case IORING_OP_READ: |
3529d8c2 | 4445 | ret = io_read_prep(req, sqe, true); |
f67676d1 JA |
4446 | break; |
4447 | case IORING_OP_WRITEV: | |
4448 | case IORING_OP_WRITE_FIXED: | |
3a6820f2 | 4449 | case IORING_OP_WRITE: |
3529d8c2 | 4450 | ret = io_write_prep(req, sqe, true); |
f67676d1 | 4451 | break; |
0969e783 | 4452 | case IORING_OP_POLL_ADD: |
3529d8c2 | 4453 | ret = io_poll_add_prep(req, sqe); |
0969e783 JA |
4454 | break; |
4455 | case IORING_OP_POLL_REMOVE: | |
3529d8c2 | 4456 | ret = io_poll_remove_prep(req, sqe); |
0969e783 | 4457 | break; |
8ed8d3c3 | 4458 | case IORING_OP_FSYNC: |
3529d8c2 | 4459 | ret = io_prep_fsync(req, sqe); |
8ed8d3c3 JA |
4460 | break; |
4461 | case IORING_OP_SYNC_FILE_RANGE: | |
3529d8c2 | 4462 | ret = io_prep_sfr(req, sqe); |
8ed8d3c3 | 4463 | break; |
03b1230c | 4464 | case IORING_OP_SENDMSG: |
fddaface | 4465 | case IORING_OP_SEND: |
3529d8c2 | 4466 | ret = io_sendmsg_prep(req, sqe); |
03b1230c JA |
4467 | break; |
4468 | case IORING_OP_RECVMSG: | |
fddaface | 4469 | case IORING_OP_RECV: |
3529d8c2 | 4470 | ret = io_recvmsg_prep(req, sqe); |
03b1230c | 4471 | break; |
f499a021 | 4472 | case IORING_OP_CONNECT: |
3529d8c2 | 4473 | ret = io_connect_prep(req, sqe); |
f499a021 | 4474 | break; |
2d28390a | 4475 | case IORING_OP_TIMEOUT: |
3529d8c2 | 4476 | ret = io_timeout_prep(req, sqe, false); |
b7bb4f7d | 4477 | break; |
b29472ee | 4478 | case IORING_OP_TIMEOUT_REMOVE: |
3529d8c2 | 4479 | ret = io_timeout_remove_prep(req, sqe); |
b29472ee | 4480 | break; |
fbf23849 | 4481 | case IORING_OP_ASYNC_CANCEL: |
3529d8c2 | 4482 | ret = io_async_cancel_prep(req, sqe); |
fbf23849 | 4483 | break; |
2d28390a | 4484 | case IORING_OP_LINK_TIMEOUT: |
3529d8c2 | 4485 | ret = io_timeout_prep(req, sqe, true); |
b7bb4f7d | 4486 | break; |
8ed8d3c3 | 4487 | case IORING_OP_ACCEPT: |
3529d8c2 | 4488 | ret = io_accept_prep(req, sqe); |
8ed8d3c3 | 4489 | break; |
d63d1b5e JA |
4490 | case IORING_OP_FALLOCATE: |
4491 | ret = io_fallocate_prep(req, sqe); | |
4492 | break; | |
15b71abe JA |
4493 | case IORING_OP_OPENAT: |
4494 | ret = io_openat_prep(req, sqe); | |
4495 | break; | |
b5dba59e JA |
4496 | case IORING_OP_CLOSE: |
4497 | ret = io_close_prep(req, sqe); | |
4498 | break; | |
05f3fb3c JA |
4499 | case IORING_OP_FILES_UPDATE: |
4500 | ret = io_files_update_prep(req, sqe); | |
4501 | break; | |
eddc7ef5 JA |
4502 | case IORING_OP_STATX: |
4503 | ret = io_statx_prep(req, sqe); | |
4504 | break; | |
4840e418 JA |
4505 | case IORING_OP_FADVISE: |
4506 | ret = io_fadvise_prep(req, sqe); | |
4507 | break; | |
c1ca757b JA |
4508 | case IORING_OP_MADVISE: |
4509 | ret = io_madvise_prep(req, sqe); | |
4510 | break; | |
cebdb986 JA |
4511 | case IORING_OP_OPENAT2: |
4512 | ret = io_openat2_prep(req, sqe); | |
4513 | break; | |
3e4827b0 JA |
4514 | case IORING_OP_EPOLL_CTL: |
4515 | ret = io_epoll_ctl_prep(req, sqe); | |
4516 | break; | |
7d67af2c PB |
4517 | case IORING_OP_SPLICE: |
4518 | ret = io_splice_prep(req, sqe); | |
4519 | break; | |
ddf0322d JA |
4520 | case IORING_OP_PROVIDE_BUFFERS: |
4521 | ret = io_provide_buffers_prep(req, sqe); | |
4522 | break; | |
f67676d1 | 4523 | default: |
e781573e JA |
4524 | printk_once(KERN_WARNING "io_uring: unhandled opcode %d\n", |
4525 | req->opcode); | |
4526 | ret = -EINVAL; | |
b7bb4f7d | 4527 | break; |
f67676d1 JA |
4528 | } |
4529 | ||
b7bb4f7d | 4530 | return ret; |
f67676d1 JA |
4531 | } |
4532 | ||
3529d8c2 | 4533 | static int io_req_defer(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
de0617e4 | 4534 | { |
a197f664 | 4535 | struct io_ring_ctx *ctx = req->ctx; |
f67676d1 | 4536 | int ret; |
de0617e4 | 4537 | |
9d858b21 BL |
4538 | /* Still need defer if there is pending req in defer list. */ |
4539 | if (!req_need_defer(req) && list_empty(&ctx->defer_list)) | |
de0617e4 JA |
4540 | return 0; |
4541 | ||
3529d8c2 | 4542 | if (!req->io && io_alloc_async_ctx(req)) |
de0617e4 JA |
4543 | return -EAGAIN; |
4544 | ||
3529d8c2 | 4545 | ret = io_req_defer_prep(req, sqe); |
b7bb4f7d | 4546 | if (ret < 0) |
2d28390a | 4547 | return ret; |
2d28390a | 4548 | |
de0617e4 | 4549 | spin_lock_irq(&ctx->completion_lock); |
9d858b21 | 4550 | if (!req_need_defer(req) && list_empty(&ctx->defer_list)) { |
de0617e4 | 4551 | spin_unlock_irq(&ctx->completion_lock); |
de0617e4 JA |
4552 | return 0; |
4553 | } | |
4554 | ||
915967f6 | 4555 | trace_io_uring_defer(ctx, req, req->user_data); |
de0617e4 JA |
4556 | list_add_tail(&req->list, &ctx->defer_list); |
4557 | spin_unlock_irq(&ctx->completion_lock); | |
4558 | return -EIOCBQUEUED; | |
4559 | } | |
4560 | ||
99bc4c38 PB |
4561 | static void io_cleanup_req(struct io_kiocb *req) |
4562 | { | |
4563 | struct io_async_ctx *io = req->io; | |
4564 | ||
4565 | switch (req->opcode) { | |
4566 | case IORING_OP_READV: | |
4567 | case IORING_OP_READ_FIXED: | |
4568 | case IORING_OP_READ: | |
4569 | case IORING_OP_WRITEV: | |
4570 | case IORING_OP_WRITE_FIXED: | |
4571 | case IORING_OP_WRITE: | |
4572 | if (io->rw.iov != io->rw.fast_iov) | |
4573 | kfree(io->rw.iov); | |
4574 | break; | |
4575 | case IORING_OP_SENDMSG: | |
4576 | case IORING_OP_RECVMSG: | |
4577 | if (io->msg.iov != io->msg.fast_iov) | |
4578 | kfree(io->msg.iov); | |
4579 | break; | |
8fef80bf PB |
4580 | case IORING_OP_OPENAT: |
4581 | case IORING_OP_OPENAT2: | |
4582 | case IORING_OP_STATX: | |
4583 | putname(req->open.filename); | |
4584 | break; | |
7d67af2c PB |
4585 | case IORING_OP_SPLICE: |
4586 | io_put_file(req, req->splice.file_in, | |
4587 | (req->splice.flags & SPLICE_F_FD_IN_FIXED)); | |
4588 | break; | |
99bc4c38 PB |
4589 | } |
4590 | ||
4591 | req->flags &= ~REQ_F_NEED_CLEANUP; | |
4592 | } | |
4593 | ||
3529d8c2 | 4594 | static int io_issue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
014db007 | 4595 | bool force_nonblock) |
2b188cc1 | 4596 | { |
a197f664 | 4597 | struct io_ring_ctx *ctx = req->ctx; |
d625c6ee | 4598 | int ret; |
2b188cc1 | 4599 | |
d625c6ee | 4600 | switch (req->opcode) { |
2b188cc1 | 4601 | case IORING_OP_NOP: |
78e19bbe | 4602 | ret = io_nop(req); |
2b188cc1 JA |
4603 | break; |
4604 | case IORING_OP_READV: | |
edafccee | 4605 | case IORING_OP_READ_FIXED: |
3a6820f2 | 4606 | case IORING_OP_READ: |
3529d8c2 JA |
4607 | if (sqe) { |
4608 | ret = io_read_prep(req, sqe, force_nonblock); | |
4609 | if (ret < 0) | |
4610 | break; | |
4611 | } | |
014db007 | 4612 | ret = io_read(req, force_nonblock); |
edafccee | 4613 | break; |
3529d8c2 | 4614 | case IORING_OP_WRITEV: |
edafccee | 4615 | case IORING_OP_WRITE_FIXED: |
3a6820f2 | 4616 | case IORING_OP_WRITE: |
3529d8c2 JA |
4617 | if (sqe) { |
4618 | ret = io_write_prep(req, sqe, force_nonblock); | |
4619 | if (ret < 0) | |
4620 | break; | |
4621 | } | |
014db007 | 4622 | ret = io_write(req, force_nonblock); |
2b188cc1 | 4623 | break; |
c992fe29 | 4624 | case IORING_OP_FSYNC: |
3529d8c2 JA |
4625 | if (sqe) { |
4626 | ret = io_prep_fsync(req, sqe); | |
4627 | if (ret < 0) | |
4628 | break; | |
4629 | } | |
014db007 | 4630 | ret = io_fsync(req, force_nonblock); |
c992fe29 | 4631 | break; |
221c5eb2 | 4632 | case IORING_OP_POLL_ADD: |
3529d8c2 JA |
4633 | if (sqe) { |
4634 | ret = io_poll_add_prep(req, sqe); | |
4635 | if (ret) | |
4636 | break; | |
4637 | } | |
014db007 | 4638 | ret = io_poll_add(req); |
221c5eb2 JA |
4639 | break; |
4640 | case IORING_OP_POLL_REMOVE: | |
3529d8c2 JA |
4641 | if (sqe) { |
4642 | ret = io_poll_remove_prep(req, sqe); | |
4643 | if (ret < 0) | |
4644 | break; | |
4645 | } | |
fc4df999 | 4646 | ret = io_poll_remove(req); |
221c5eb2 | 4647 | break; |
5d17b4a4 | 4648 | case IORING_OP_SYNC_FILE_RANGE: |
3529d8c2 JA |
4649 | if (sqe) { |
4650 | ret = io_prep_sfr(req, sqe); | |
4651 | if (ret < 0) | |
4652 | break; | |
4653 | } | |
014db007 | 4654 | ret = io_sync_file_range(req, force_nonblock); |
5d17b4a4 | 4655 | break; |
0fa03c62 | 4656 | case IORING_OP_SENDMSG: |
fddaface | 4657 | case IORING_OP_SEND: |
3529d8c2 JA |
4658 | if (sqe) { |
4659 | ret = io_sendmsg_prep(req, sqe); | |
4660 | if (ret < 0) | |
4661 | break; | |
4662 | } | |
fddaface | 4663 | if (req->opcode == IORING_OP_SENDMSG) |
014db007 | 4664 | ret = io_sendmsg(req, force_nonblock); |
fddaface | 4665 | else |
014db007 | 4666 | ret = io_send(req, force_nonblock); |
0fa03c62 | 4667 | break; |
aa1fa28f | 4668 | case IORING_OP_RECVMSG: |
fddaface | 4669 | case IORING_OP_RECV: |
3529d8c2 JA |
4670 | if (sqe) { |
4671 | ret = io_recvmsg_prep(req, sqe); | |
4672 | if (ret) | |
4673 | break; | |
4674 | } | |
fddaface | 4675 | if (req->opcode == IORING_OP_RECVMSG) |
014db007 | 4676 | ret = io_recvmsg(req, force_nonblock); |
fddaface | 4677 | else |
014db007 | 4678 | ret = io_recv(req, force_nonblock); |
aa1fa28f | 4679 | break; |
5262f567 | 4680 | case IORING_OP_TIMEOUT: |
3529d8c2 JA |
4681 | if (sqe) { |
4682 | ret = io_timeout_prep(req, sqe, false); | |
4683 | if (ret) | |
4684 | break; | |
4685 | } | |
fc4df999 | 4686 | ret = io_timeout(req); |
5262f567 | 4687 | break; |
11365043 | 4688 | case IORING_OP_TIMEOUT_REMOVE: |
3529d8c2 JA |
4689 | if (sqe) { |
4690 | ret = io_timeout_remove_prep(req, sqe); | |
4691 | if (ret) | |
4692 | break; | |
4693 | } | |
fc4df999 | 4694 | ret = io_timeout_remove(req); |
11365043 | 4695 | break; |
17f2fe35 | 4696 | case IORING_OP_ACCEPT: |
3529d8c2 JA |
4697 | if (sqe) { |
4698 | ret = io_accept_prep(req, sqe); | |
4699 | if (ret) | |
4700 | break; | |
4701 | } | |
014db007 | 4702 | ret = io_accept(req, force_nonblock); |
17f2fe35 | 4703 | break; |
f8e85cf2 | 4704 | case IORING_OP_CONNECT: |
3529d8c2 JA |
4705 | if (sqe) { |
4706 | ret = io_connect_prep(req, sqe); | |
4707 | if (ret) | |
4708 | break; | |
4709 | } | |
014db007 | 4710 | ret = io_connect(req, force_nonblock); |
f8e85cf2 | 4711 | break; |
62755e35 | 4712 | case IORING_OP_ASYNC_CANCEL: |
3529d8c2 JA |
4713 | if (sqe) { |
4714 | ret = io_async_cancel_prep(req, sqe); | |
4715 | if (ret) | |
4716 | break; | |
4717 | } | |
014db007 | 4718 | ret = io_async_cancel(req); |
62755e35 | 4719 | break; |
d63d1b5e JA |
4720 | case IORING_OP_FALLOCATE: |
4721 | if (sqe) { | |
4722 | ret = io_fallocate_prep(req, sqe); | |
4723 | if (ret) | |
4724 | break; | |
4725 | } | |
014db007 | 4726 | ret = io_fallocate(req, force_nonblock); |
d63d1b5e | 4727 | break; |
15b71abe JA |
4728 | case IORING_OP_OPENAT: |
4729 | if (sqe) { | |
4730 | ret = io_openat_prep(req, sqe); | |
4731 | if (ret) | |
4732 | break; | |
4733 | } | |
014db007 | 4734 | ret = io_openat(req, force_nonblock); |
15b71abe | 4735 | break; |
b5dba59e JA |
4736 | case IORING_OP_CLOSE: |
4737 | if (sqe) { | |
4738 | ret = io_close_prep(req, sqe); | |
4739 | if (ret) | |
4740 | break; | |
4741 | } | |
014db007 | 4742 | ret = io_close(req, force_nonblock); |
b5dba59e | 4743 | break; |
05f3fb3c JA |
4744 | case IORING_OP_FILES_UPDATE: |
4745 | if (sqe) { | |
4746 | ret = io_files_update_prep(req, sqe); | |
4747 | if (ret) | |
4748 | break; | |
4749 | } | |
4750 | ret = io_files_update(req, force_nonblock); | |
4751 | break; | |
eddc7ef5 JA |
4752 | case IORING_OP_STATX: |
4753 | if (sqe) { | |
4754 | ret = io_statx_prep(req, sqe); | |
4755 | if (ret) | |
4756 | break; | |
4757 | } | |
014db007 | 4758 | ret = io_statx(req, force_nonblock); |
eddc7ef5 | 4759 | break; |
4840e418 JA |
4760 | case IORING_OP_FADVISE: |
4761 | if (sqe) { | |
4762 | ret = io_fadvise_prep(req, sqe); | |
4763 | if (ret) | |
4764 | break; | |
4765 | } | |
014db007 | 4766 | ret = io_fadvise(req, force_nonblock); |
4840e418 | 4767 | break; |
c1ca757b JA |
4768 | case IORING_OP_MADVISE: |
4769 | if (sqe) { | |
4770 | ret = io_madvise_prep(req, sqe); | |
4771 | if (ret) | |
4772 | break; | |
4773 | } | |
014db007 | 4774 | ret = io_madvise(req, force_nonblock); |
c1ca757b | 4775 | break; |
cebdb986 JA |
4776 | case IORING_OP_OPENAT2: |
4777 | if (sqe) { | |
4778 | ret = io_openat2_prep(req, sqe); | |
4779 | if (ret) | |
4780 | break; | |
4781 | } | |
014db007 | 4782 | ret = io_openat2(req, force_nonblock); |
cebdb986 | 4783 | break; |
3e4827b0 JA |
4784 | case IORING_OP_EPOLL_CTL: |
4785 | if (sqe) { | |
4786 | ret = io_epoll_ctl_prep(req, sqe); | |
4787 | if (ret) | |
4788 | break; | |
4789 | } | |
014db007 | 4790 | ret = io_epoll_ctl(req, force_nonblock); |
3e4827b0 | 4791 | break; |
7d67af2c PB |
4792 | case IORING_OP_SPLICE: |
4793 | if (sqe) { | |
4794 | ret = io_splice_prep(req, sqe); | |
4795 | if (ret < 0) | |
4796 | break; | |
4797 | } | |
014db007 | 4798 | ret = io_splice(req, force_nonblock); |
7d67af2c | 4799 | break; |
ddf0322d JA |
4800 | case IORING_OP_PROVIDE_BUFFERS: |
4801 | if (sqe) { | |
4802 | ret = io_provide_buffers_prep(req, sqe); | |
4803 | if (ret) | |
4804 | break; | |
4805 | } | |
4806 | ret = io_provide_buffers(req, force_nonblock); | |
4807 | break; | |
2b188cc1 JA |
4808 | default: |
4809 | ret = -EINVAL; | |
4810 | break; | |
4811 | } | |
4812 | ||
def596e9 JA |
4813 | if (ret) |
4814 | return ret; | |
4815 | ||
4816 | if (ctx->flags & IORING_SETUP_IOPOLL) { | |
11ba820b JA |
4817 | const bool in_async = io_wq_current_is_worker(); |
4818 | ||
9e645e11 | 4819 | if (req->result == -EAGAIN) |
def596e9 JA |
4820 | return -EAGAIN; |
4821 | ||
11ba820b JA |
4822 | /* workqueue context doesn't hold uring_lock, grab it now */ |
4823 | if (in_async) | |
4824 | mutex_lock(&ctx->uring_lock); | |
4825 | ||
def596e9 | 4826 | io_iopoll_req_issued(req); |
11ba820b JA |
4827 | |
4828 | if (in_async) | |
4829 | mutex_unlock(&ctx->uring_lock); | |
def596e9 JA |
4830 | } |
4831 | ||
4832 | return 0; | |
2b188cc1 JA |
4833 | } |
4834 | ||
561fb04a | 4835 | static void io_wq_submit_work(struct io_wq_work **workptr) |
2b188cc1 | 4836 | { |
561fb04a | 4837 | struct io_wq_work *work = *workptr; |
2b188cc1 | 4838 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); |
561fb04a | 4839 | int ret = 0; |
2b188cc1 | 4840 | |
0c9d5ccd JA |
4841 | /* if NO_CANCEL is set, we must still run the work */ |
4842 | if ((work->flags & (IO_WQ_WORK_CANCEL|IO_WQ_WORK_NO_CANCEL)) == | |
4843 | IO_WQ_WORK_CANCEL) { | |
561fb04a | 4844 | ret = -ECANCELED; |
0c9d5ccd | 4845 | } |
31b51510 | 4846 | |
561fb04a | 4847 | if (!ret) { |
561fb04a | 4848 | do { |
014db007 | 4849 | ret = io_issue_sqe(req, NULL, false); |
561fb04a JA |
4850 | /* |
4851 | * We can get EAGAIN for polled IO even though we're | |
4852 | * forcing a sync submission from here, since we can't | |
4853 | * wait for request slots on the block side. | |
4854 | */ | |
4855 | if (ret != -EAGAIN) | |
4856 | break; | |
4857 | cond_resched(); | |
4858 | } while (1); | |
4859 | } | |
31b51510 | 4860 | |
561fb04a | 4861 | if (ret) { |
4e88d6e7 | 4862 | req_set_fail_links(req); |
78e19bbe | 4863 | io_cqring_add_event(req, ret); |
817869d2 | 4864 | io_put_req(req); |
edafccee | 4865 | } |
2b188cc1 | 4866 | |
e9fd9396 | 4867 | io_steal_work(req, workptr); |
2b188cc1 JA |
4868 | } |
4869 | ||
15b71abe | 4870 | static int io_req_needs_file(struct io_kiocb *req, int fd) |
9e3aa61a | 4871 | { |
d3656344 | 4872 | if (!io_op_defs[req->opcode].needs_file) |
9e3aa61a | 4873 | return 0; |
0b5faf6b | 4874 | if ((fd == -1 || fd == AT_FDCWD) && io_op_defs[req->opcode].fd_non_neg) |
d3656344 JA |
4875 | return 0; |
4876 | return 1; | |
09bb8394 JA |
4877 | } |
4878 | ||
65e19f54 JA |
4879 | static inline struct file *io_file_from_index(struct io_ring_ctx *ctx, |
4880 | int index) | |
4881 | { | |
4882 | struct fixed_file_table *table; | |
4883 | ||
05f3fb3c JA |
4884 | table = &ctx->file_data->table[index >> IORING_FILE_TABLE_SHIFT]; |
4885 | return table->files[index & IORING_FILE_TABLE_MASK];; | |
65e19f54 JA |
4886 | } |
4887 | ||
8da11c19 PB |
4888 | static int io_file_get(struct io_submit_state *state, struct io_kiocb *req, |
4889 | int fd, struct file **out_file, bool fixed) | |
09bb8394 | 4890 | { |
a197f664 | 4891 | struct io_ring_ctx *ctx = req->ctx; |
8da11c19 | 4892 | struct file *file; |
09bb8394 | 4893 | |
8da11c19 | 4894 | if (fixed) { |
05f3fb3c | 4895 | if (unlikely(!ctx->file_data || |
09bb8394 JA |
4896 | (unsigned) fd >= ctx->nr_user_files)) |
4897 | return -EBADF; | |
b7620121 | 4898 | fd = array_index_nospec(fd, ctx->nr_user_files); |
8da11c19 PB |
4899 | file = io_file_from_index(ctx, fd); |
4900 | if (!file) | |
08a45173 | 4901 | return -EBADF; |
05f3fb3c | 4902 | percpu_ref_get(&ctx->file_data->refs); |
09bb8394 | 4903 | } else { |
c826bd7a | 4904 | trace_io_uring_file_get(ctx, fd); |
8da11c19 PB |
4905 | file = __io_file_get(state, fd); |
4906 | if (unlikely(!file)) | |
09bb8394 JA |
4907 | return -EBADF; |
4908 | } | |
4909 | ||
8da11c19 | 4910 | *out_file = file; |
09bb8394 JA |
4911 | return 0; |
4912 | } | |
4913 | ||
8da11c19 PB |
4914 | static int io_req_set_file(struct io_submit_state *state, struct io_kiocb *req, |
4915 | const struct io_uring_sqe *sqe) | |
4916 | { | |
4917 | unsigned flags; | |
4918 | int fd; | |
4919 | bool fixed; | |
4920 | ||
4921 | flags = READ_ONCE(sqe->flags); | |
4922 | fd = READ_ONCE(sqe->fd); | |
4923 | ||
4924 | if (!io_req_needs_file(req, fd)) | |
4925 | return 0; | |
4926 | ||
4927 | fixed = (flags & IOSQE_FIXED_FILE); | |
4928 | if (unlikely(!fixed && req->needs_fixed_file)) | |
4929 | return -EBADF; | |
4930 | ||
4931 | return io_file_get(state, req, fd, &req->file, fixed); | |
4932 | } | |
4933 | ||
a197f664 | 4934 | static int io_grab_files(struct io_kiocb *req) |
fcb323cc JA |
4935 | { |
4936 | int ret = -EBADF; | |
a197f664 | 4937 | struct io_ring_ctx *ctx = req->ctx; |
fcb323cc | 4938 | |
f86cd20c JA |
4939 | if (req->work.files) |
4940 | return 0; | |
b14cca0c | 4941 | if (!ctx->ring_file) |
b5dba59e JA |
4942 | return -EBADF; |
4943 | ||
fcb323cc JA |
4944 | rcu_read_lock(); |
4945 | spin_lock_irq(&ctx->inflight_lock); | |
4946 | /* | |
4947 | * We use the f_ops->flush() handler to ensure that we can flush | |
4948 | * out work accessing these files if the fd is closed. Check if | |
4949 | * the fd has changed since we started down this path, and disallow | |
4950 | * this operation if it has. | |
4951 | */ | |
b14cca0c | 4952 | if (fcheck(ctx->ring_fd) == ctx->ring_file) { |
fcb323cc JA |
4953 | list_add(&req->inflight_entry, &ctx->inflight_list); |
4954 | req->flags |= REQ_F_INFLIGHT; | |
4955 | req->work.files = current->files; | |
4956 | ret = 0; | |
4957 | } | |
4958 | spin_unlock_irq(&ctx->inflight_lock); | |
4959 | rcu_read_unlock(); | |
4960 | ||
4961 | return ret; | |
4962 | } | |
4963 | ||
2665abfd | 4964 | static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer) |
2b188cc1 | 4965 | { |
ad8a48ac JA |
4966 | struct io_timeout_data *data = container_of(timer, |
4967 | struct io_timeout_data, timer); | |
4968 | struct io_kiocb *req = data->req; | |
2665abfd JA |
4969 | struct io_ring_ctx *ctx = req->ctx; |
4970 | struct io_kiocb *prev = NULL; | |
4971 | unsigned long flags; | |
2665abfd JA |
4972 | |
4973 | spin_lock_irqsave(&ctx->completion_lock, flags); | |
4974 | ||
4975 | /* | |
4976 | * We don't expect the list to be empty, that will only happen if we | |
4977 | * race with the completion of the linked work. | |
4978 | */ | |
4493233e PB |
4979 | if (!list_empty(&req->link_list)) { |
4980 | prev = list_entry(req->link_list.prev, struct io_kiocb, | |
4981 | link_list); | |
5d960724 | 4982 | if (refcount_inc_not_zero(&prev->refs)) { |
4493233e | 4983 | list_del_init(&req->link_list); |
5d960724 JA |
4984 | prev->flags &= ~REQ_F_LINK_TIMEOUT; |
4985 | } else | |
76a46e06 | 4986 | prev = NULL; |
2665abfd JA |
4987 | } |
4988 | ||
4989 | spin_unlock_irqrestore(&ctx->completion_lock, flags); | |
4990 | ||
4991 | if (prev) { | |
4e88d6e7 | 4992 | req_set_fail_links(prev); |
014db007 | 4993 | io_async_find_and_cancel(ctx, req, prev->user_data, -ETIME); |
76a46e06 | 4994 | io_put_req(prev); |
47f46768 JA |
4995 | } else { |
4996 | io_cqring_add_event(req, -ETIME); | |
4997 | io_put_req(req); | |
2665abfd | 4998 | } |
2665abfd JA |
4999 | return HRTIMER_NORESTART; |
5000 | } | |
5001 | ||
ad8a48ac | 5002 | static void io_queue_linked_timeout(struct io_kiocb *req) |
2665abfd | 5003 | { |
76a46e06 | 5004 | struct io_ring_ctx *ctx = req->ctx; |
2665abfd | 5005 | |
76a46e06 JA |
5006 | /* |
5007 | * If the list is now empty, then our linked request finished before | |
5008 | * we got a chance to setup the timer | |
5009 | */ | |
5010 | spin_lock_irq(&ctx->completion_lock); | |
4493233e | 5011 | if (!list_empty(&req->link_list)) { |
2d28390a | 5012 | struct io_timeout_data *data = &req->io->timeout; |
94ae5e77 | 5013 | |
ad8a48ac JA |
5014 | data->timer.function = io_link_timeout_fn; |
5015 | hrtimer_start(&data->timer, timespec64_to_ktime(data->ts), | |
5016 | data->mode); | |
2665abfd | 5017 | } |
76a46e06 | 5018 | spin_unlock_irq(&ctx->completion_lock); |
2665abfd | 5019 | |
2665abfd | 5020 | /* drop submission reference */ |
76a46e06 JA |
5021 | io_put_req(req); |
5022 | } | |
2665abfd | 5023 | |
ad8a48ac | 5024 | static struct io_kiocb *io_prep_linked_timeout(struct io_kiocb *req) |
2665abfd JA |
5025 | { |
5026 | struct io_kiocb *nxt; | |
5027 | ||
5028 | if (!(req->flags & REQ_F_LINK)) | |
5029 | return NULL; | |
d7718a9d JA |
5030 | /* for polled retry, if flag is set, we already went through here */ |
5031 | if (req->flags & REQ_F_POLLED) | |
5032 | return NULL; | |
2665abfd | 5033 | |
4493233e PB |
5034 | nxt = list_first_entry_or_null(&req->link_list, struct io_kiocb, |
5035 | link_list); | |
d625c6ee | 5036 | if (!nxt || nxt->opcode != IORING_OP_LINK_TIMEOUT) |
76a46e06 | 5037 | return NULL; |
2665abfd | 5038 | |
76a46e06 | 5039 | req->flags |= REQ_F_LINK_TIMEOUT; |
76a46e06 | 5040 | return nxt; |
2665abfd JA |
5041 | } |
5042 | ||
3529d8c2 | 5043 | static void __io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
2b188cc1 | 5044 | { |
4a0a7a18 | 5045 | struct io_kiocb *linked_timeout; |
4bc4494e | 5046 | struct io_kiocb *nxt; |
193155c8 | 5047 | const struct cred *old_creds = NULL; |
e0c5c576 | 5048 | int ret; |
2b188cc1 | 5049 | |
4a0a7a18 JA |
5050 | again: |
5051 | linked_timeout = io_prep_linked_timeout(req); | |
5052 | ||
193155c8 JA |
5053 | if (req->work.creds && req->work.creds != current_cred()) { |
5054 | if (old_creds) | |
5055 | revert_creds(old_creds); | |
5056 | if (old_creds == req->work.creds) | |
5057 | old_creds = NULL; /* restored original creds */ | |
5058 | else | |
5059 | old_creds = override_creds(req->work.creds); | |
5060 | } | |
5061 | ||
014db007 | 5062 | ret = io_issue_sqe(req, sqe, true); |
491381ce JA |
5063 | |
5064 | /* | |
5065 | * We async punt it if the file wasn't marked NOWAIT, or if the file | |
5066 | * doesn't support non-blocking read/write attempts | |
5067 | */ | |
5068 | if (ret == -EAGAIN && (!(req->flags & REQ_F_NOWAIT) || | |
5069 | (req->flags & REQ_F_MUST_PUNT))) { | |
d7718a9d JA |
5070 | if (io_arm_poll_handler(req)) { |
5071 | if (linked_timeout) | |
5072 | io_queue_linked_timeout(linked_timeout); | |
4bc4494e | 5073 | goto exit; |
d7718a9d | 5074 | } |
86a761f8 | 5075 | punt: |
f86cd20c | 5076 | if (io_op_defs[req->opcode].file_table) { |
bbad27b2 PB |
5077 | ret = io_grab_files(req); |
5078 | if (ret) | |
5079 | goto err; | |
2b188cc1 | 5080 | } |
bbad27b2 PB |
5081 | |
5082 | /* | |
5083 | * Queued up for async execution, worker will release | |
5084 | * submit reference when the iocb is actually submitted. | |
5085 | */ | |
5086 | io_queue_async_work(req); | |
4bc4494e | 5087 | goto exit; |
2b188cc1 | 5088 | } |
e65ef56d | 5089 | |
fcb323cc | 5090 | err: |
4bc4494e | 5091 | nxt = NULL; |
76a46e06 | 5092 | /* drop submission reference */ |
2a44f467 | 5093 | io_put_req_find_next(req, &nxt); |
e65ef56d | 5094 | |
f9bd67f6 | 5095 | if (linked_timeout) { |
76a46e06 | 5096 | if (!ret) |
f9bd67f6 | 5097 | io_queue_linked_timeout(linked_timeout); |
76a46e06 | 5098 | else |
f9bd67f6 | 5099 | io_put_req(linked_timeout); |
76a46e06 JA |
5100 | } |
5101 | ||
e65ef56d | 5102 | /* and drop final reference, if we failed */ |
9e645e11 | 5103 | if (ret) { |
78e19bbe | 5104 | io_cqring_add_event(req, ret); |
4e88d6e7 | 5105 | req_set_fail_links(req); |
e65ef56d | 5106 | io_put_req(req); |
9e645e11 | 5107 | } |
4a0a7a18 JA |
5108 | if (nxt) { |
5109 | req = nxt; | |
86a761f8 PB |
5110 | |
5111 | if (req->flags & REQ_F_FORCE_ASYNC) | |
5112 | goto punt; | |
4a0a7a18 JA |
5113 | goto again; |
5114 | } | |
4bc4494e | 5115 | exit: |
193155c8 JA |
5116 | if (old_creds) |
5117 | revert_creds(old_creds); | |
2b188cc1 JA |
5118 | } |
5119 | ||
3529d8c2 | 5120 | static void io_queue_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
4fe2c963 JL |
5121 | { |
5122 | int ret; | |
5123 | ||
3529d8c2 | 5124 | ret = io_req_defer(req, sqe); |
4fe2c963 JL |
5125 | if (ret) { |
5126 | if (ret != -EIOCBQUEUED) { | |
1118591a | 5127 | fail_req: |
78e19bbe | 5128 | io_cqring_add_event(req, ret); |
4e88d6e7 | 5129 | req_set_fail_links(req); |
78e19bbe | 5130 | io_double_put_req(req); |
4fe2c963 | 5131 | } |
2550878f | 5132 | } else if (req->flags & REQ_F_FORCE_ASYNC) { |
1118591a PB |
5133 | ret = io_req_defer_prep(req, sqe); |
5134 | if (unlikely(ret < 0)) | |
5135 | goto fail_req; | |
ce35a47a JA |
5136 | /* |
5137 | * Never try inline submit of IOSQE_ASYNC is set, go straight | |
5138 | * to async execution. | |
5139 | */ | |
5140 | req->work.flags |= IO_WQ_WORK_CONCURRENT; | |
5141 | io_queue_async_work(req); | |
5142 | } else { | |
3529d8c2 | 5143 | __io_queue_sqe(req, sqe); |
ce35a47a | 5144 | } |
4fe2c963 JL |
5145 | } |
5146 | ||
1b4a51b6 | 5147 | static inline void io_queue_link_head(struct io_kiocb *req) |
4fe2c963 | 5148 | { |
94ae5e77 | 5149 | if (unlikely(req->flags & REQ_F_FAIL_LINK)) { |
1b4a51b6 PB |
5150 | io_cqring_add_event(req, -ECANCELED); |
5151 | io_double_put_req(req); | |
5152 | } else | |
3529d8c2 | 5153 | io_queue_sqe(req, NULL); |
4fe2c963 JL |
5154 | } |
5155 | ||
4e88d6e7 | 5156 | #define SQE_VALID_FLAGS (IOSQE_FIXED_FILE|IOSQE_IO_DRAIN|IOSQE_IO_LINK| \ |
ce35a47a | 5157 | IOSQE_IO_HARDLINK | IOSQE_ASYNC) |
9e645e11 | 5158 | |
3529d8c2 JA |
5159 | static bool io_submit_sqe(struct io_kiocb *req, const struct io_uring_sqe *sqe, |
5160 | struct io_submit_state *state, struct io_kiocb **link) | |
9e645e11 | 5161 | { |
a197f664 | 5162 | struct io_ring_ctx *ctx = req->ctx; |
32fe525b | 5163 | unsigned int sqe_flags; |
75c6a039 | 5164 | int ret, id; |
9e645e11 | 5165 | |
32fe525b | 5166 | sqe_flags = READ_ONCE(sqe->flags); |
9e645e11 JA |
5167 | |
5168 | /* enforce forwards compatibility on users */ | |
32fe525b | 5169 | if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) { |
9e645e11 | 5170 | ret = -EINVAL; |
196be95c | 5171 | goto err_req; |
9e645e11 JA |
5172 | } |
5173 | ||
75c6a039 JA |
5174 | id = READ_ONCE(sqe->personality); |
5175 | if (id) { | |
193155c8 JA |
5176 | req->work.creds = idr_find(&ctx->personality_idr, id); |
5177 | if (unlikely(!req->work.creds)) { | |
75c6a039 JA |
5178 | ret = -EINVAL; |
5179 | goto err_req; | |
5180 | } | |
193155c8 | 5181 | get_cred(req->work.creds); |
75c6a039 JA |
5182 | } |
5183 | ||
6b47ee6e | 5184 | /* same numerical values with corresponding REQ_F_*, safe to copy */ |
8da11c19 PB |
5185 | req->flags |= sqe_flags & (IOSQE_IO_DRAIN | IOSQE_IO_HARDLINK | |
5186 | IOSQE_ASYNC | IOSQE_FIXED_FILE); | |
9e645e11 | 5187 | |
3529d8c2 | 5188 | ret = io_req_set_file(state, req, sqe); |
9e645e11 JA |
5189 | if (unlikely(ret)) { |
5190 | err_req: | |
78e19bbe JA |
5191 | io_cqring_add_event(req, ret); |
5192 | io_double_put_req(req); | |
2e6e1fde | 5193 | return false; |
9e645e11 JA |
5194 | } |
5195 | ||
9e645e11 JA |
5196 | /* |
5197 | * If we already have a head request, queue this one for async | |
5198 | * submittal once the head completes. If we don't have a head but | |
5199 | * IOSQE_IO_LINK is set in the sqe, start a new head. This one will be | |
5200 | * submitted sync once the chain is complete. If none of those | |
5201 | * conditions are true (normal request), then just queue it. | |
5202 | */ | |
5203 | if (*link) { | |
9d76377f | 5204 | struct io_kiocb *head = *link; |
4e88d6e7 | 5205 | |
8cdf2193 PB |
5206 | /* |
5207 | * Taking sequential execution of a link, draining both sides | |
5208 | * of the link also fullfils IOSQE_IO_DRAIN semantics for all | |
5209 | * requests in the link. So, it drains the head and the | |
5210 | * next after the link request. The last one is done via | |
5211 | * drain_next flag to persist the effect across calls. | |
5212 | */ | |
711be031 PB |
5213 | if (sqe_flags & IOSQE_IO_DRAIN) { |
5214 | head->flags |= REQ_F_IO_DRAIN; | |
5215 | ctx->drain_next = 1; | |
5216 | } | |
b7bb4f7d | 5217 | if (io_alloc_async_ctx(req)) { |
9e645e11 JA |
5218 | ret = -EAGAIN; |
5219 | goto err_req; | |
5220 | } | |
5221 | ||
3529d8c2 | 5222 | ret = io_req_defer_prep(req, sqe); |
2d28390a | 5223 | if (ret) { |
4e88d6e7 | 5224 | /* fail even hard links since we don't submit */ |
9d76377f | 5225 | head->flags |= REQ_F_FAIL_LINK; |
f67676d1 | 5226 | goto err_req; |
2d28390a | 5227 | } |
9d76377f PB |
5228 | trace_io_uring_link(ctx, req, head); |
5229 | list_add_tail(&req->link_list, &head->link_list); | |
32fe525b PB |
5230 | |
5231 | /* last request of a link, enqueue the link */ | |
5232 | if (!(sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK))) { | |
5233 | io_queue_link_head(head); | |
5234 | *link = NULL; | |
5235 | } | |
9e645e11 | 5236 | } else { |
711be031 PB |
5237 | if (unlikely(ctx->drain_next)) { |
5238 | req->flags |= REQ_F_IO_DRAIN; | |
5239 | req->ctx->drain_next = 0; | |
5240 | } | |
5241 | if (sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) { | |
5242 | req->flags |= REQ_F_LINK; | |
711be031 PB |
5243 | INIT_LIST_HEAD(&req->link_list); |
5244 | ret = io_req_defer_prep(req, sqe); | |
5245 | if (ret) | |
5246 | req->flags |= REQ_F_FAIL_LINK; | |
5247 | *link = req; | |
5248 | } else { | |
5249 | io_queue_sqe(req, sqe); | |
5250 | } | |
9e645e11 | 5251 | } |
2e6e1fde PB |
5252 | |
5253 | return true; | |
9e645e11 JA |
5254 | } |
5255 | ||
9a56a232 JA |
5256 | /* |
5257 | * Batched submission is done, ensure local IO is flushed out. | |
5258 | */ | |
5259 | static void io_submit_state_end(struct io_submit_state *state) | |
5260 | { | |
5261 | blk_finish_plug(&state->plug); | |
3d6770fb | 5262 | io_file_put(state); |
2579f913 | 5263 | if (state->free_reqs) |
6c8a3134 | 5264 | kmem_cache_free_bulk(req_cachep, state->free_reqs, state->reqs); |
9a56a232 JA |
5265 | } |
5266 | ||
5267 | /* | |
5268 | * Start submission side cache. | |
5269 | */ | |
5270 | static void io_submit_state_start(struct io_submit_state *state, | |
22efde59 | 5271 | unsigned int max_ios) |
9a56a232 JA |
5272 | { |
5273 | blk_start_plug(&state->plug); | |
2579f913 | 5274 | state->free_reqs = 0; |
9a56a232 JA |
5275 | state->file = NULL; |
5276 | state->ios_left = max_ios; | |
5277 | } | |
5278 | ||
2b188cc1 JA |
5279 | static void io_commit_sqring(struct io_ring_ctx *ctx) |
5280 | { | |
75b28aff | 5281 | struct io_rings *rings = ctx->rings; |
2b188cc1 | 5282 | |
caf582c6 PB |
5283 | /* |
5284 | * Ensure any loads from the SQEs are done at this point, | |
5285 | * since once we write the new head, the application could | |
5286 | * write new data to them. | |
5287 | */ | |
5288 | smp_store_release(&rings->sq.head, ctx->cached_sq_head); | |
2b188cc1 JA |
5289 | } |
5290 | ||
2b188cc1 | 5291 | /* |
3529d8c2 | 5292 | * Fetch an sqe, if one is available. Note that sqe_ptr will point to memory |
2b188cc1 JA |
5293 | * that is mapped by userspace. This means that care needs to be taken to |
5294 | * ensure that reads are stable, as we cannot rely on userspace always | |
5295 | * being a good citizen. If members of the sqe are validated and then later | |
5296 | * used, it's important that those reads are done through READ_ONCE() to | |
5297 | * prevent a re-load down the line. | |
5298 | */ | |
3529d8c2 JA |
5299 | static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req, |
5300 | const struct io_uring_sqe **sqe_ptr) | |
2b188cc1 | 5301 | { |
75b28aff | 5302 | u32 *sq_array = ctx->sq_array; |
2b188cc1 JA |
5303 | unsigned head; |
5304 | ||
5305 | /* | |
5306 | * The cached sq head (or cq tail) serves two purposes: | |
5307 | * | |
5308 | * 1) allows us to batch the cost of updating the user visible | |
5309 | * head updates. | |
5310 | * 2) allows the kernel side to track the head on its own, even | |
5311 | * though the application is the one updating it. | |
5312 | */ | |
ee7d46d9 | 5313 | head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]); |
9835d6fa | 5314 | if (likely(head < ctx->sq_entries)) { |
cf6fd4bd PB |
5315 | /* |
5316 | * All io need record the previous position, if LINK vs DARIN, | |
5317 | * it can be used to mark the position of the first IO in the | |
5318 | * link list. | |
5319 | */ | |
5320 | req->sequence = ctx->cached_sq_head; | |
3529d8c2 JA |
5321 | *sqe_ptr = &ctx->sq_sqes[head]; |
5322 | req->opcode = READ_ONCE((*sqe_ptr)->opcode); | |
5323 | req->user_data = READ_ONCE((*sqe_ptr)->user_data); | |
2b188cc1 JA |
5324 | ctx->cached_sq_head++; |
5325 | return true; | |
5326 | } | |
5327 | ||
5328 | /* drop invalid entries */ | |
5329 | ctx->cached_sq_head++; | |
498ccd9e | 5330 | ctx->cached_sq_dropped++; |
ee7d46d9 | 5331 | WRITE_ONCE(ctx->rings->sq_dropped, ctx->cached_sq_dropped); |
2b188cc1 JA |
5332 | return false; |
5333 | } | |
5334 | ||
fb5ccc98 | 5335 | static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr, |
ae9428ca PB |
5336 | struct file *ring_file, int ring_fd, |
5337 | struct mm_struct **mm, bool async) | |
6c271ce2 JA |
5338 | { |
5339 | struct io_submit_state state, *statep = NULL; | |
9e645e11 | 5340 | struct io_kiocb *link = NULL; |
9e645e11 | 5341 | int i, submitted = 0; |
95a1b3ff | 5342 | bool mm_fault = false; |
6c271ce2 | 5343 | |
c4a2ed72 | 5344 | /* if we have a backlog and couldn't flush it all, return BUSY */ |
ad3eb2c8 JA |
5345 | if (test_bit(0, &ctx->sq_check_overflow)) { |
5346 | if (!list_empty(&ctx->cq_overflow_list) && | |
5347 | !io_cqring_overflow_flush(ctx, false)) | |
5348 | return -EBUSY; | |
5349 | } | |
6c271ce2 | 5350 | |
ee7d46d9 PB |
5351 | /* make sure SQ entry isn't read before tail */ |
5352 | nr = min3(nr, ctx->sq_entries, io_sqring_entries(ctx)); | |
9ef4f124 | 5353 | |
2b85edfc PB |
5354 | if (!percpu_ref_tryget_many(&ctx->refs, nr)) |
5355 | return -EAGAIN; | |
6c271ce2 JA |
5356 | |
5357 | if (nr > IO_PLUG_THRESHOLD) { | |
22efde59 | 5358 | io_submit_state_start(&state, nr); |
6c271ce2 JA |
5359 | statep = &state; |
5360 | } | |
5361 | ||
b14cca0c PB |
5362 | ctx->ring_fd = ring_fd; |
5363 | ctx->ring_file = ring_file; | |
5364 | ||
6c271ce2 | 5365 | for (i = 0; i < nr; i++) { |
3529d8c2 | 5366 | const struct io_uring_sqe *sqe; |
196be95c | 5367 | struct io_kiocb *req; |
1cb1edb2 | 5368 | int err; |
fb5ccc98 | 5369 | |
196be95c PB |
5370 | req = io_get_req(ctx, statep); |
5371 | if (unlikely(!req)) { | |
5372 | if (!submitted) | |
5373 | submitted = -EAGAIN; | |
fb5ccc98 | 5374 | break; |
196be95c | 5375 | } |
3529d8c2 | 5376 | if (!io_get_sqring(ctx, req, &sqe)) { |
2b85edfc | 5377 | __io_req_do_free(req); |
196be95c PB |
5378 | break; |
5379 | } | |
fb5ccc98 | 5380 | |
d3656344 JA |
5381 | /* will complete beyond this point, count as submitted */ |
5382 | submitted++; | |
5383 | ||
5384 | if (unlikely(req->opcode >= IORING_OP_LAST)) { | |
1cb1edb2 PB |
5385 | err = -EINVAL; |
5386 | fail_req: | |
5387 | io_cqring_add_event(req, err); | |
d3656344 | 5388 | io_double_put_req(req); |
196be95c PB |
5389 | break; |
5390 | } | |
fb5ccc98 | 5391 | |
d3656344 | 5392 | if (io_op_defs[req->opcode].needs_mm && !*mm) { |
95a1b3ff | 5393 | mm_fault = mm_fault || !mmget_not_zero(ctx->sqo_mm); |
1cb1edb2 PB |
5394 | if (unlikely(mm_fault)) { |
5395 | err = -EFAULT; | |
5396 | goto fail_req; | |
95a1b3ff | 5397 | } |
1cb1edb2 PB |
5398 | use_mm(ctx->sqo_mm); |
5399 | *mm = ctx->sqo_mm; | |
9e645e11 | 5400 | } |
9e645e11 | 5401 | |
cf6fd4bd | 5402 | req->needs_fixed_file = async; |
354420f7 JA |
5403 | trace_io_uring_submit_sqe(ctx, req->opcode, req->user_data, |
5404 | true, async); | |
3529d8c2 | 5405 | if (!io_submit_sqe(req, sqe, statep, &link)) |
2e6e1fde | 5406 | break; |
6c271ce2 JA |
5407 | } |
5408 | ||
9466f437 PB |
5409 | if (unlikely(submitted != nr)) { |
5410 | int ref_used = (submitted == -EAGAIN) ? 0 : submitted; | |
5411 | ||
5412 | percpu_ref_put_many(&ctx->refs, nr - ref_used); | |
5413 | } | |
9e645e11 | 5414 | if (link) |
1b4a51b6 | 5415 | io_queue_link_head(link); |
6c271ce2 JA |
5416 | if (statep) |
5417 | io_submit_state_end(&state); | |
5418 | ||
ae9428ca PB |
5419 | /* Commit SQ ring head once we've consumed and submitted all SQEs */ |
5420 | io_commit_sqring(ctx); | |
5421 | ||
6c271ce2 JA |
5422 | return submitted; |
5423 | } | |
5424 | ||
5425 | static int io_sq_thread(void *data) | |
5426 | { | |
6c271ce2 JA |
5427 | struct io_ring_ctx *ctx = data; |
5428 | struct mm_struct *cur_mm = NULL; | |
181e448d | 5429 | const struct cred *old_cred; |
6c271ce2 JA |
5430 | mm_segment_t old_fs; |
5431 | DEFINE_WAIT(wait); | |
6c271ce2 | 5432 | unsigned long timeout; |
bdcd3eab | 5433 | int ret = 0; |
6c271ce2 | 5434 | |
206aefde | 5435 | complete(&ctx->completions[1]); |
a4c0b3de | 5436 | |
6c271ce2 JA |
5437 | old_fs = get_fs(); |
5438 | set_fs(USER_DS); | |
181e448d | 5439 | old_cred = override_creds(ctx->creds); |
6c271ce2 | 5440 | |
bdcd3eab | 5441 | timeout = jiffies + ctx->sq_thread_idle; |
2bbcd6d3 | 5442 | while (!kthread_should_park()) { |
fb5ccc98 | 5443 | unsigned int to_submit; |
6c271ce2 | 5444 | |
bdcd3eab | 5445 | if (!list_empty(&ctx->poll_list)) { |
6c271ce2 JA |
5446 | unsigned nr_events = 0; |
5447 | ||
bdcd3eab XW |
5448 | mutex_lock(&ctx->uring_lock); |
5449 | if (!list_empty(&ctx->poll_list)) | |
5450 | io_iopoll_getevents(ctx, &nr_events, 0); | |
5451 | else | |
6c271ce2 | 5452 | timeout = jiffies + ctx->sq_thread_idle; |
bdcd3eab | 5453 | mutex_unlock(&ctx->uring_lock); |
6c271ce2 JA |
5454 | } |
5455 | ||
fb5ccc98 | 5456 | to_submit = io_sqring_entries(ctx); |
c1edbf5f JA |
5457 | |
5458 | /* | |
5459 | * If submit got -EBUSY, flag us as needing the application | |
5460 | * to enter the kernel to reap and flush events. | |
5461 | */ | |
5462 | if (!to_submit || ret == -EBUSY) { | |
7143b5ac SG |
5463 | /* |
5464 | * Drop cur_mm before scheduling, we can't hold it for | |
5465 | * long periods (or over schedule()). Do this before | |
5466 | * adding ourselves to the waitqueue, as the unuse/drop | |
5467 | * may sleep. | |
5468 | */ | |
5469 | if (cur_mm) { | |
5470 | unuse_mm(cur_mm); | |
5471 | mmput(cur_mm); | |
5472 | cur_mm = NULL; | |
5473 | } | |
5474 | ||
6c271ce2 JA |
5475 | /* |
5476 | * We're polling. If we're within the defined idle | |
5477 | * period, then let us spin without work before going | |
c1edbf5f JA |
5478 | * to sleep. The exception is if we got EBUSY doing |
5479 | * more IO, we should wait for the application to | |
5480 | * reap events and wake us up. | |
6c271ce2 | 5481 | */ |
bdcd3eab | 5482 | if (!list_empty(&ctx->poll_list) || |
df069d80 JA |
5483 | (!time_after(jiffies, timeout) && ret != -EBUSY && |
5484 | !percpu_ref_is_dying(&ctx->refs))) { | |
b41e9852 JA |
5485 | if (current->task_works) |
5486 | task_work_run(); | |
9831a90c | 5487 | cond_resched(); |
6c271ce2 JA |
5488 | continue; |
5489 | } | |
5490 | ||
6c271ce2 JA |
5491 | prepare_to_wait(&ctx->sqo_wait, &wait, |
5492 | TASK_INTERRUPTIBLE); | |
5493 | ||
bdcd3eab XW |
5494 | /* |
5495 | * While doing polled IO, before going to sleep, we need | |
5496 | * to check if there are new reqs added to poll_list, it | |
5497 | * is because reqs may have been punted to io worker and | |
5498 | * will be added to poll_list later, hence check the | |
5499 | * poll_list again. | |
5500 | */ | |
5501 | if ((ctx->flags & IORING_SETUP_IOPOLL) && | |
5502 | !list_empty_careful(&ctx->poll_list)) { | |
5503 | finish_wait(&ctx->sqo_wait, &wait); | |
5504 | continue; | |
5505 | } | |
5506 | ||
6c271ce2 | 5507 | /* Tell userspace we may need a wakeup call */ |
75b28aff | 5508 | ctx->rings->sq_flags |= IORING_SQ_NEED_WAKEUP; |
0d7bae69 SB |
5509 | /* make sure to read SQ tail after writing flags */ |
5510 | smp_mb(); | |
6c271ce2 | 5511 | |
fb5ccc98 | 5512 | to_submit = io_sqring_entries(ctx); |
c1edbf5f | 5513 | if (!to_submit || ret == -EBUSY) { |
2bbcd6d3 | 5514 | if (kthread_should_park()) { |
6c271ce2 JA |
5515 | finish_wait(&ctx->sqo_wait, &wait); |
5516 | break; | |
5517 | } | |
b41e9852 JA |
5518 | if (current->task_works) { |
5519 | task_work_run(); | |
5520 | continue; | |
5521 | } | |
6c271ce2 JA |
5522 | if (signal_pending(current)) |
5523 | flush_signals(current); | |
5524 | schedule(); | |
5525 | finish_wait(&ctx->sqo_wait, &wait); | |
5526 | ||
75b28aff | 5527 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
5528 | continue; |
5529 | } | |
5530 | finish_wait(&ctx->sqo_wait, &wait); | |
5531 | ||
75b28aff | 5532 | ctx->rings->sq_flags &= ~IORING_SQ_NEED_WAKEUP; |
6c271ce2 JA |
5533 | } |
5534 | ||
8a4955ff | 5535 | mutex_lock(&ctx->uring_lock); |
1d7bb1d5 | 5536 | ret = io_submit_sqes(ctx, to_submit, NULL, -1, &cur_mm, true); |
8a4955ff | 5537 | mutex_unlock(&ctx->uring_lock); |
bdcd3eab | 5538 | timeout = jiffies + ctx->sq_thread_idle; |
6c271ce2 JA |
5539 | } |
5540 | ||
b41e9852 JA |
5541 | if (current->task_works) |
5542 | task_work_run(); | |
5543 | ||
6c271ce2 JA |
5544 | set_fs(old_fs); |
5545 | if (cur_mm) { | |
5546 | unuse_mm(cur_mm); | |
5547 | mmput(cur_mm); | |
5548 | } | |
181e448d | 5549 | revert_creds(old_cred); |
06058632 | 5550 | |
2bbcd6d3 | 5551 | kthread_parkme(); |
06058632 | 5552 | |
6c271ce2 JA |
5553 | return 0; |
5554 | } | |
5555 | ||
bda52162 JA |
5556 | struct io_wait_queue { |
5557 | struct wait_queue_entry wq; | |
5558 | struct io_ring_ctx *ctx; | |
5559 | unsigned to_wait; | |
5560 | unsigned nr_timeouts; | |
5561 | }; | |
5562 | ||
1d7bb1d5 | 5563 | static inline bool io_should_wake(struct io_wait_queue *iowq, bool noflush) |
bda52162 JA |
5564 | { |
5565 | struct io_ring_ctx *ctx = iowq->ctx; | |
5566 | ||
5567 | /* | |
d195a66e | 5568 | * Wake up if we have enough events, or if a timeout occurred since we |
bda52162 JA |
5569 | * started waiting. For timeouts, we always want to return to userspace, |
5570 | * regardless of event count. | |
5571 | */ | |
1d7bb1d5 | 5572 | return io_cqring_events(ctx, noflush) >= iowq->to_wait || |
bda52162 JA |
5573 | atomic_read(&ctx->cq_timeouts) != iowq->nr_timeouts; |
5574 | } | |
5575 | ||
5576 | static int io_wake_function(struct wait_queue_entry *curr, unsigned int mode, | |
5577 | int wake_flags, void *key) | |
5578 | { | |
5579 | struct io_wait_queue *iowq = container_of(curr, struct io_wait_queue, | |
5580 | wq); | |
5581 | ||
1d7bb1d5 JA |
5582 | /* use noflush == true, as we can't safely rely on locking context */ |
5583 | if (!io_should_wake(iowq, true)) | |
bda52162 JA |
5584 | return -1; |
5585 | ||
5586 | return autoremove_wake_function(curr, mode, wake_flags, key); | |
5587 | } | |
5588 | ||
2b188cc1 JA |
5589 | /* |
5590 | * Wait until events become available, if we don't already have some. The | |
5591 | * application must reap them itself, as they reside on the shared cq ring. | |
5592 | */ | |
5593 | static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events, | |
5594 | const sigset_t __user *sig, size_t sigsz) | |
5595 | { | |
bda52162 JA |
5596 | struct io_wait_queue iowq = { |
5597 | .wq = { | |
5598 | .private = current, | |
5599 | .func = io_wake_function, | |
5600 | .entry = LIST_HEAD_INIT(iowq.wq.entry), | |
5601 | }, | |
5602 | .ctx = ctx, | |
5603 | .to_wait = min_events, | |
5604 | }; | |
75b28aff | 5605 | struct io_rings *rings = ctx->rings; |
e9ffa5c2 | 5606 | int ret = 0; |
2b188cc1 | 5607 | |
b41e9852 JA |
5608 | do { |
5609 | if (io_cqring_events(ctx, false) >= min_events) | |
5610 | return 0; | |
5611 | if (!current->task_works) | |
5612 | break; | |
5613 | task_work_run(); | |
5614 | } while (1); | |
2b188cc1 JA |
5615 | |
5616 | if (sig) { | |
9e75ad5d AB |
5617 | #ifdef CONFIG_COMPAT |
5618 | if (in_compat_syscall()) | |
5619 | ret = set_compat_user_sigmask((const compat_sigset_t __user *)sig, | |
b772434b | 5620 | sigsz); |
9e75ad5d AB |
5621 | else |
5622 | #endif | |
b772434b | 5623 | ret = set_user_sigmask(sig, sigsz); |
9e75ad5d | 5624 | |
2b188cc1 JA |
5625 | if (ret) |
5626 | return ret; | |
5627 | } | |
5628 | ||
bda52162 | 5629 | iowq.nr_timeouts = atomic_read(&ctx->cq_timeouts); |
c826bd7a | 5630 | trace_io_uring_cqring_wait(ctx, min_events); |
bda52162 JA |
5631 | do { |
5632 | prepare_to_wait_exclusive(&ctx->wait, &iowq.wq, | |
5633 | TASK_INTERRUPTIBLE); | |
b41e9852 JA |
5634 | if (current->task_works) |
5635 | task_work_run(); | |
1d7bb1d5 | 5636 | if (io_should_wake(&iowq, false)) |
bda52162 JA |
5637 | break; |
5638 | schedule(); | |
5639 | if (signal_pending(current)) { | |
e9ffa5c2 | 5640 | ret = -EINTR; |
bda52162 JA |
5641 | break; |
5642 | } | |
5643 | } while (1); | |
5644 | finish_wait(&ctx->wait, &iowq.wq); | |
5645 | ||
e9ffa5c2 | 5646 | restore_saved_sigmask_unless(ret == -EINTR); |
2b188cc1 | 5647 | |
75b28aff | 5648 | return READ_ONCE(rings->cq.head) == READ_ONCE(rings->cq.tail) ? ret : 0; |
2b188cc1 JA |
5649 | } |
5650 | ||
6b06314c JA |
5651 | static void __io_sqe_files_unregister(struct io_ring_ctx *ctx) |
5652 | { | |
5653 | #if defined(CONFIG_UNIX) | |
5654 | if (ctx->ring_sock) { | |
5655 | struct sock *sock = ctx->ring_sock->sk; | |
5656 | struct sk_buff *skb; | |
5657 | ||
5658 | while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) | |
5659 | kfree_skb(skb); | |
5660 | } | |
5661 | #else | |
5662 | int i; | |
5663 | ||
65e19f54 JA |
5664 | for (i = 0; i < ctx->nr_user_files; i++) { |
5665 | struct file *file; | |
5666 | ||
5667 | file = io_file_from_index(ctx, i); | |
5668 | if (file) | |
5669 | fput(file); | |
5670 | } | |
6b06314c JA |
5671 | #endif |
5672 | } | |
5673 | ||
05f3fb3c JA |
5674 | static void io_file_ref_kill(struct percpu_ref *ref) |
5675 | { | |
5676 | struct fixed_file_data *data; | |
5677 | ||
5678 | data = container_of(ref, struct fixed_file_data, refs); | |
5679 | complete(&data->done); | |
5680 | } | |
5681 | ||
6b06314c JA |
5682 | static int io_sqe_files_unregister(struct io_ring_ctx *ctx) |
5683 | { | |
05f3fb3c | 5684 | struct fixed_file_data *data = ctx->file_data; |
65e19f54 JA |
5685 | unsigned nr_tables, i; |
5686 | ||
05f3fb3c | 5687 | if (!data) |
6b06314c JA |
5688 | return -ENXIO; |
5689 | ||
05f3fb3c | 5690 | percpu_ref_kill_and_confirm(&data->refs, io_file_ref_kill); |
e46a7950 | 5691 | flush_work(&data->ref_work); |
2faf852d JA |
5692 | wait_for_completion(&data->done); |
5693 | io_ring_file_ref_flush(data); | |
05f3fb3c JA |
5694 | percpu_ref_exit(&data->refs); |
5695 | ||
6b06314c | 5696 | __io_sqe_files_unregister(ctx); |
65e19f54 JA |
5697 | nr_tables = DIV_ROUND_UP(ctx->nr_user_files, IORING_MAX_FILES_TABLE); |
5698 | for (i = 0; i < nr_tables; i++) | |
05f3fb3c JA |
5699 | kfree(data->table[i].files); |
5700 | kfree(data->table); | |
5701 | kfree(data); | |
5702 | ctx->file_data = NULL; | |
6b06314c JA |
5703 | ctx->nr_user_files = 0; |
5704 | return 0; | |
5705 | } | |
5706 | ||
6c271ce2 JA |
5707 | static void io_sq_thread_stop(struct io_ring_ctx *ctx) |
5708 | { | |
5709 | if (ctx->sqo_thread) { | |
206aefde | 5710 | wait_for_completion(&ctx->completions[1]); |
2bbcd6d3 RP |
5711 | /* |
5712 | * The park is a bit of a work-around, without it we get | |
5713 | * warning spews on shutdown with SQPOLL set and affinity | |
5714 | * set to a single CPU. | |
5715 | */ | |
06058632 | 5716 | kthread_park(ctx->sqo_thread); |
6c271ce2 JA |
5717 | kthread_stop(ctx->sqo_thread); |
5718 | ctx->sqo_thread = NULL; | |
5719 | } | |
5720 | } | |
5721 | ||
6b06314c JA |
5722 | static void io_finish_async(struct io_ring_ctx *ctx) |
5723 | { | |
6c271ce2 JA |
5724 | io_sq_thread_stop(ctx); |
5725 | ||
561fb04a JA |
5726 | if (ctx->io_wq) { |
5727 | io_wq_destroy(ctx->io_wq); | |
5728 | ctx->io_wq = NULL; | |
6b06314c JA |
5729 | } |
5730 | } | |
5731 | ||
5732 | #if defined(CONFIG_UNIX) | |
6b06314c JA |
5733 | /* |
5734 | * Ensure the UNIX gc is aware of our file set, so we are certain that | |
5735 | * the io_uring can be safely unregistered on process exit, even if we have | |
5736 | * loops in the file referencing. | |
5737 | */ | |
5738 | static int __io_sqe_files_scm(struct io_ring_ctx *ctx, int nr, int offset) | |
5739 | { | |
5740 | struct sock *sk = ctx->ring_sock->sk; | |
5741 | struct scm_fp_list *fpl; | |
5742 | struct sk_buff *skb; | |
08a45173 | 5743 | int i, nr_files; |
6b06314c JA |
5744 | |
5745 | if (!capable(CAP_SYS_RESOURCE) && !capable(CAP_SYS_ADMIN)) { | |
5746 | unsigned long inflight = ctx->user->unix_inflight + nr; | |
5747 | ||
5748 | if (inflight > task_rlimit(current, RLIMIT_NOFILE)) | |
5749 | return -EMFILE; | |
5750 | } | |
5751 | ||
5752 | fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); | |
5753 | if (!fpl) | |
5754 | return -ENOMEM; | |
5755 | ||
5756 | skb = alloc_skb(0, GFP_KERNEL); | |
5757 | if (!skb) { | |
5758 | kfree(fpl); | |
5759 | return -ENOMEM; | |
5760 | } | |
5761 | ||
5762 | skb->sk = sk; | |
6b06314c | 5763 | |
08a45173 | 5764 | nr_files = 0; |
6b06314c JA |
5765 | fpl->user = get_uid(ctx->user); |
5766 | for (i = 0; i < nr; i++) { | |
65e19f54 JA |
5767 | struct file *file = io_file_from_index(ctx, i + offset); |
5768 | ||
5769 | if (!file) | |
08a45173 | 5770 | continue; |
65e19f54 | 5771 | fpl->fp[nr_files] = get_file(file); |
08a45173 JA |
5772 | unix_inflight(fpl->user, fpl->fp[nr_files]); |
5773 | nr_files++; | |
6b06314c JA |
5774 | } |
5775 | ||
08a45173 JA |
5776 | if (nr_files) { |
5777 | fpl->max = SCM_MAX_FD; | |
5778 | fpl->count = nr_files; | |
5779 | UNIXCB(skb).fp = fpl; | |
05f3fb3c | 5780 | skb->destructor = unix_destruct_scm; |
08a45173 JA |
5781 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); |
5782 | skb_queue_head(&sk->sk_receive_queue, skb); | |
6b06314c | 5783 | |
08a45173 JA |
5784 | for (i = 0; i < nr_files; i++) |
5785 | fput(fpl->fp[i]); | |
5786 | } else { | |
5787 | kfree_skb(skb); | |
5788 | kfree(fpl); | |
5789 | } | |
6b06314c JA |
5790 | |
5791 | return 0; | |
5792 | } | |
5793 | ||
5794 | /* | |
5795 | * If UNIX sockets are enabled, fd passing can cause a reference cycle which | |
5796 | * causes regular reference counting to break down. We rely on the UNIX | |
5797 | * garbage collection to take care of this problem for us. | |
5798 | */ | |
5799 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
5800 | { | |
5801 | unsigned left, total; | |
5802 | int ret = 0; | |
5803 | ||
5804 | total = 0; | |
5805 | left = ctx->nr_user_files; | |
5806 | while (left) { | |
5807 | unsigned this_files = min_t(unsigned, left, SCM_MAX_FD); | |
6b06314c JA |
5808 | |
5809 | ret = __io_sqe_files_scm(ctx, this_files, total); | |
5810 | if (ret) | |
5811 | break; | |
5812 | left -= this_files; | |
5813 | total += this_files; | |
5814 | } | |
5815 | ||
5816 | if (!ret) | |
5817 | return 0; | |
5818 | ||
5819 | while (total < ctx->nr_user_files) { | |
65e19f54 JA |
5820 | struct file *file = io_file_from_index(ctx, total); |
5821 | ||
5822 | if (file) | |
5823 | fput(file); | |
6b06314c JA |
5824 | total++; |
5825 | } | |
5826 | ||
5827 | return ret; | |
5828 | } | |
5829 | #else | |
5830 | static int io_sqe_files_scm(struct io_ring_ctx *ctx) | |
5831 | { | |
5832 | return 0; | |
5833 | } | |
5834 | #endif | |
5835 | ||
65e19f54 JA |
5836 | static int io_sqe_alloc_file_tables(struct io_ring_ctx *ctx, unsigned nr_tables, |
5837 | unsigned nr_files) | |
5838 | { | |
5839 | int i; | |
5840 | ||
5841 | for (i = 0; i < nr_tables; i++) { | |
05f3fb3c | 5842 | struct fixed_file_table *table = &ctx->file_data->table[i]; |
65e19f54 JA |
5843 | unsigned this_files; |
5844 | ||
5845 | this_files = min(nr_files, IORING_MAX_FILES_TABLE); | |
5846 | table->files = kcalloc(this_files, sizeof(struct file *), | |
5847 | GFP_KERNEL); | |
5848 | if (!table->files) | |
5849 | break; | |
5850 | nr_files -= this_files; | |
5851 | } | |
5852 | ||
5853 | if (i == nr_tables) | |
5854 | return 0; | |
5855 | ||
5856 | for (i = 0; i < nr_tables; i++) { | |
05f3fb3c | 5857 | struct fixed_file_table *table = &ctx->file_data->table[i]; |
65e19f54 JA |
5858 | kfree(table->files); |
5859 | } | |
5860 | return 1; | |
5861 | } | |
5862 | ||
05f3fb3c JA |
5863 | static void io_ring_file_put(struct io_ring_ctx *ctx, struct file *file) |
5864 | { | |
5865 | #if defined(CONFIG_UNIX) | |
5866 | struct sock *sock = ctx->ring_sock->sk; | |
5867 | struct sk_buff_head list, *head = &sock->sk_receive_queue; | |
5868 | struct sk_buff *skb; | |
5869 | int i; | |
5870 | ||
5871 | __skb_queue_head_init(&list); | |
5872 | ||
5873 | /* | |
5874 | * Find the skb that holds this file in its SCM_RIGHTS. When found, | |
5875 | * remove this entry and rearrange the file array. | |
5876 | */ | |
5877 | skb = skb_dequeue(head); | |
5878 | while (skb) { | |
5879 | struct scm_fp_list *fp; | |
5880 | ||
5881 | fp = UNIXCB(skb).fp; | |
5882 | for (i = 0; i < fp->count; i++) { | |
5883 | int left; | |
5884 | ||
5885 | if (fp->fp[i] != file) | |
5886 | continue; | |
5887 | ||
5888 | unix_notinflight(fp->user, fp->fp[i]); | |
5889 | left = fp->count - 1 - i; | |
5890 | if (left) { | |
5891 | memmove(&fp->fp[i], &fp->fp[i + 1], | |
5892 | left * sizeof(struct file *)); | |
5893 | } | |
5894 | fp->count--; | |
5895 | if (!fp->count) { | |
5896 | kfree_skb(skb); | |
5897 | skb = NULL; | |
5898 | } else { | |
5899 | __skb_queue_tail(&list, skb); | |
5900 | } | |
5901 | fput(file); | |
5902 | file = NULL; | |
5903 | break; | |
5904 | } | |
5905 | ||
5906 | if (!file) | |
5907 | break; | |
5908 | ||
5909 | __skb_queue_tail(&list, skb); | |
5910 | ||
5911 | skb = skb_dequeue(head); | |
5912 | } | |
5913 | ||
5914 | if (skb_peek(&list)) { | |
5915 | spin_lock_irq(&head->lock); | |
5916 | while ((skb = __skb_dequeue(&list)) != NULL) | |
5917 | __skb_queue_tail(head, skb); | |
5918 | spin_unlock_irq(&head->lock); | |
5919 | } | |
5920 | #else | |
5921 | fput(file); | |
5922 | #endif | |
5923 | } | |
5924 | ||
5925 | struct io_file_put { | |
5926 | struct llist_node llist; | |
5927 | struct file *file; | |
5928 | struct completion *done; | |
5929 | }; | |
5930 | ||
2faf852d | 5931 | static void io_ring_file_ref_flush(struct fixed_file_data *data) |
65e19f54 | 5932 | { |
05f3fb3c | 5933 | struct io_file_put *pfile, *tmp; |
05f3fb3c | 5934 | struct llist_node *node; |
65e19f54 | 5935 | |
05f3fb3c JA |
5936 | while ((node = llist_del_all(&data->put_llist)) != NULL) { |
5937 | llist_for_each_entry_safe(pfile, tmp, node, llist) { | |
5938 | io_ring_file_put(data->ctx, pfile->file); | |
5939 | if (pfile->done) | |
5940 | complete(pfile->done); | |
5941 | else | |
5942 | kfree(pfile); | |
5943 | } | |
65e19f54 | 5944 | } |
2faf852d | 5945 | } |
65e19f54 | 5946 | |
2faf852d JA |
5947 | static void io_ring_file_ref_switch(struct work_struct *work) |
5948 | { | |
5949 | struct fixed_file_data *data; | |
65e19f54 | 5950 | |
2faf852d JA |
5951 | data = container_of(work, struct fixed_file_data, ref_work); |
5952 | io_ring_file_ref_flush(data); | |
05f3fb3c JA |
5953 | percpu_ref_switch_to_percpu(&data->refs); |
5954 | } | |
65e19f54 | 5955 | |
05f3fb3c JA |
5956 | static void io_file_data_ref_zero(struct percpu_ref *ref) |
5957 | { | |
5958 | struct fixed_file_data *data; | |
5959 | ||
5960 | data = container_of(ref, struct fixed_file_data, refs); | |
5961 | ||
2faf852d JA |
5962 | /* |
5963 | * We can't safely switch from inside this context, punt to wq. If | |
5964 | * the table ref is going away, the table is being unregistered. | |
5965 | * Don't queue up the async work for that case, the caller will | |
5966 | * handle it. | |
5967 | */ | |
5968 | if (!percpu_ref_is_dying(&data->refs)) | |
5969 | queue_work(system_wq, &data->ref_work); | |
65e19f54 JA |
5970 | } |
5971 | ||
6b06314c JA |
5972 | static int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, |
5973 | unsigned nr_args) | |
5974 | { | |
5975 | __s32 __user *fds = (__s32 __user *) arg; | |
65e19f54 | 5976 | unsigned nr_tables; |
05f3fb3c | 5977 | struct file *file; |
6b06314c JA |
5978 | int fd, ret = 0; |
5979 | unsigned i; | |
5980 | ||
05f3fb3c | 5981 | if (ctx->file_data) |
6b06314c JA |
5982 | return -EBUSY; |
5983 | if (!nr_args) | |
5984 | return -EINVAL; | |
5985 | if (nr_args > IORING_MAX_FIXED_FILES) | |
5986 | return -EMFILE; | |
5987 | ||
05f3fb3c JA |
5988 | ctx->file_data = kzalloc(sizeof(*ctx->file_data), GFP_KERNEL); |
5989 | if (!ctx->file_data) | |
5990 | return -ENOMEM; | |
5991 | ctx->file_data->ctx = ctx; | |
5992 | init_completion(&ctx->file_data->done); | |
5993 | ||
65e19f54 | 5994 | nr_tables = DIV_ROUND_UP(nr_args, IORING_MAX_FILES_TABLE); |
05f3fb3c JA |
5995 | ctx->file_data->table = kcalloc(nr_tables, |
5996 | sizeof(struct fixed_file_table), | |
65e19f54 | 5997 | GFP_KERNEL); |
05f3fb3c JA |
5998 | if (!ctx->file_data->table) { |
5999 | kfree(ctx->file_data); | |
6000 | ctx->file_data = NULL; | |
6b06314c | 6001 | return -ENOMEM; |
05f3fb3c JA |
6002 | } |
6003 | ||
6004 | if (percpu_ref_init(&ctx->file_data->refs, io_file_data_ref_zero, | |
6005 | PERCPU_REF_ALLOW_REINIT, GFP_KERNEL)) { | |
6006 | kfree(ctx->file_data->table); | |
6007 | kfree(ctx->file_data); | |
6008 | ctx->file_data = NULL; | |
6b06314c | 6009 | return -ENOMEM; |
05f3fb3c JA |
6010 | } |
6011 | ctx->file_data->put_llist.first = NULL; | |
6012 | INIT_WORK(&ctx->file_data->ref_work, io_ring_file_ref_switch); | |
6b06314c | 6013 | |
65e19f54 | 6014 | if (io_sqe_alloc_file_tables(ctx, nr_tables, nr_args)) { |
05f3fb3c JA |
6015 | percpu_ref_exit(&ctx->file_data->refs); |
6016 | kfree(ctx->file_data->table); | |
6017 | kfree(ctx->file_data); | |
6018 | ctx->file_data = NULL; | |
65e19f54 JA |
6019 | return -ENOMEM; |
6020 | } | |
6021 | ||
08a45173 | 6022 | for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { |
65e19f54 JA |
6023 | struct fixed_file_table *table; |
6024 | unsigned index; | |
6025 | ||
6b06314c JA |
6026 | ret = -EFAULT; |
6027 | if (copy_from_user(&fd, &fds[i], sizeof(fd))) | |
6028 | break; | |
08a45173 JA |
6029 | /* allow sparse sets */ |
6030 | if (fd == -1) { | |
6031 | ret = 0; | |
6032 | continue; | |
6033 | } | |
6b06314c | 6034 | |
05f3fb3c | 6035 | table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT]; |
65e19f54 | 6036 | index = i & IORING_FILE_TABLE_MASK; |
05f3fb3c | 6037 | file = fget(fd); |
6b06314c JA |
6038 | |
6039 | ret = -EBADF; | |
05f3fb3c | 6040 | if (!file) |
6b06314c | 6041 | break; |
05f3fb3c | 6042 | |
6b06314c JA |
6043 | /* |
6044 | * Don't allow io_uring instances to be registered. If UNIX | |
6045 | * isn't enabled, then this causes a reference cycle and this | |
6046 | * instance can never get freed. If UNIX is enabled we'll | |
6047 | * handle it just fine, but there's still no point in allowing | |
6048 | * a ring fd as it doesn't support regular read/write anyway. | |
6049 | */ | |
05f3fb3c JA |
6050 | if (file->f_op == &io_uring_fops) { |
6051 | fput(file); | |
6b06314c JA |
6052 | break; |
6053 | } | |
6b06314c | 6054 | ret = 0; |
05f3fb3c | 6055 | table->files[index] = file; |
6b06314c JA |
6056 | } |
6057 | ||
6058 | if (ret) { | |
65e19f54 | 6059 | for (i = 0; i < ctx->nr_user_files; i++) { |
65e19f54 JA |
6060 | file = io_file_from_index(ctx, i); |
6061 | if (file) | |
6062 | fput(file); | |
6063 | } | |
6064 | for (i = 0; i < nr_tables; i++) | |
05f3fb3c | 6065 | kfree(ctx->file_data->table[i].files); |
6b06314c | 6066 | |
05f3fb3c JA |
6067 | kfree(ctx->file_data->table); |
6068 | kfree(ctx->file_data); | |
6069 | ctx->file_data = NULL; | |
6b06314c JA |
6070 | ctx->nr_user_files = 0; |
6071 | return ret; | |
6072 | } | |
6073 | ||
6074 | ret = io_sqe_files_scm(ctx); | |
6075 | if (ret) | |
6076 | io_sqe_files_unregister(ctx); | |
6077 | ||
6078 | return ret; | |
6079 | } | |
6080 | ||
c3a31e60 JA |
6081 | static int io_sqe_file_register(struct io_ring_ctx *ctx, struct file *file, |
6082 | int index) | |
6083 | { | |
6084 | #if defined(CONFIG_UNIX) | |
6085 | struct sock *sock = ctx->ring_sock->sk; | |
6086 | struct sk_buff_head *head = &sock->sk_receive_queue; | |
6087 | struct sk_buff *skb; | |
6088 | ||
6089 | /* | |
6090 | * See if we can merge this file into an existing skb SCM_RIGHTS | |
6091 | * file set. If there's no room, fall back to allocating a new skb | |
6092 | * and filling it in. | |
6093 | */ | |
6094 | spin_lock_irq(&head->lock); | |
6095 | skb = skb_peek(head); | |
6096 | if (skb) { | |
6097 | struct scm_fp_list *fpl = UNIXCB(skb).fp; | |
6098 | ||
6099 | if (fpl->count < SCM_MAX_FD) { | |
6100 | __skb_unlink(skb, head); | |
6101 | spin_unlock_irq(&head->lock); | |
6102 | fpl->fp[fpl->count] = get_file(file); | |
6103 | unix_inflight(fpl->user, fpl->fp[fpl->count]); | |
6104 | fpl->count++; | |
6105 | spin_lock_irq(&head->lock); | |
6106 | __skb_queue_head(head, skb); | |
6107 | } else { | |
6108 | skb = NULL; | |
6109 | } | |
6110 | } | |
6111 | spin_unlock_irq(&head->lock); | |
6112 | ||
6113 | if (skb) { | |
6114 | fput(file); | |
6115 | return 0; | |
6116 | } | |
6117 | ||
6118 | return __io_sqe_files_scm(ctx, 1, index); | |
6119 | #else | |
6120 | return 0; | |
6121 | #endif | |
6122 | } | |
6123 | ||
05f3fb3c | 6124 | static void io_atomic_switch(struct percpu_ref *ref) |
c3a31e60 | 6125 | { |
05f3fb3c JA |
6126 | struct fixed_file_data *data; |
6127 | ||
dd3db2a3 JA |
6128 | /* |
6129 | * Juggle reference to ensure we hit zero, if needed, so we can | |
6130 | * switch back to percpu mode | |
6131 | */ | |
05f3fb3c | 6132 | data = container_of(ref, struct fixed_file_data, refs); |
dd3db2a3 JA |
6133 | percpu_ref_put(&data->refs); |
6134 | percpu_ref_get(&data->refs); | |
05f3fb3c JA |
6135 | } |
6136 | ||
6137 | static bool io_queue_file_removal(struct fixed_file_data *data, | |
6138 | struct file *file) | |
6139 | { | |
6140 | struct io_file_put *pfile, pfile_stack; | |
6141 | DECLARE_COMPLETION_ONSTACK(done); | |
6142 | ||
6143 | /* | |
6144 | * If we fail allocating the struct we need for doing async reomval | |
6145 | * of this file, just punt to sync and wait for it. | |
6146 | */ | |
6147 | pfile = kzalloc(sizeof(*pfile), GFP_KERNEL); | |
6148 | if (!pfile) { | |
6149 | pfile = &pfile_stack; | |
6150 | pfile->done = &done; | |
6151 | } | |
6152 | ||
6153 | pfile->file = file; | |
6154 | llist_add(&pfile->llist, &data->put_llist); | |
6155 | ||
6156 | if (pfile == &pfile_stack) { | |
dd3db2a3 | 6157 | percpu_ref_switch_to_atomic(&data->refs, io_atomic_switch); |
05f3fb3c JA |
6158 | wait_for_completion(&done); |
6159 | flush_work(&data->ref_work); | |
6160 | return false; | |
6161 | } | |
6162 | ||
6163 | return true; | |
6164 | } | |
6165 | ||
6166 | static int __io_sqe_files_update(struct io_ring_ctx *ctx, | |
6167 | struct io_uring_files_update *up, | |
6168 | unsigned nr_args) | |
6169 | { | |
6170 | struct fixed_file_data *data = ctx->file_data; | |
6171 | bool ref_switch = false; | |
6172 | struct file *file; | |
c3a31e60 JA |
6173 | __s32 __user *fds; |
6174 | int fd, i, err; | |
6175 | __u32 done; | |
6176 | ||
05f3fb3c | 6177 | if (check_add_overflow(up->offset, nr_args, &done)) |
c3a31e60 JA |
6178 | return -EOVERFLOW; |
6179 | if (done > ctx->nr_user_files) | |
6180 | return -EINVAL; | |
6181 | ||
6182 | done = 0; | |
05f3fb3c | 6183 | fds = u64_to_user_ptr(up->fds); |
c3a31e60 | 6184 | while (nr_args) { |
65e19f54 JA |
6185 | struct fixed_file_table *table; |
6186 | unsigned index; | |
6187 | ||
c3a31e60 JA |
6188 | err = 0; |
6189 | if (copy_from_user(&fd, &fds[done], sizeof(fd))) { | |
6190 | err = -EFAULT; | |
6191 | break; | |
6192 | } | |
05f3fb3c JA |
6193 | i = array_index_nospec(up->offset, ctx->nr_user_files); |
6194 | table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT]; | |
65e19f54 JA |
6195 | index = i & IORING_FILE_TABLE_MASK; |
6196 | if (table->files[index]) { | |
05f3fb3c | 6197 | file = io_file_from_index(ctx, index); |
65e19f54 | 6198 | table->files[index] = NULL; |
05f3fb3c JA |
6199 | if (io_queue_file_removal(data, file)) |
6200 | ref_switch = true; | |
c3a31e60 JA |
6201 | } |
6202 | if (fd != -1) { | |
c3a31e60 JA |
6203 | file = fget(fd); |
6204 | if (!file) { | |
6205 | err = -EBADF; | |
6206 | break; | |
6207 | } | |
6208 | /* | |
6209 | * Don't allow io_uring instances to be registered. If | |
6210 | * UNIX isn't enabled, then this causes a reference | |
6211 | * cycle and this instance can never get freed. If UNIX | |
6212 | * is enabled we'll handle it just fine, but there's | |
6213 | * still no point in allowing a ring fd as it doesn't | |
6214 | * support regular read/write anyway. | |
6215 | */ | |
6216 | if (file->f_op == &io_uring_fops) { | |
6217 | fput(file); | |
6218 | err = -EBADF; | |
6219 | break; | |
6220 | } | |
65e19f54 | 6221 | table->files[index] = file; |
c3a31e60 JA |
6222 | err = io_sqe_file_register(ctx, file, i); |
6223 | if (err) | |
6224 | break; | |
6225 | } | |
6226 | nr_args--; | |
6227 | done++; | |
05f3fb3c JA |
6228 | up->offset++; |
6229 | } | |
6230 | ||
dd3db2a3 | 6231 | if (ref_switch) |
05f3fb3c | 6232 | percpu_ref_switch_to_atomic(&data->refs, io_atomic_switch); |
c3a31e60 JA |
6233 | |
6234 | return done ? done : err; | |
6235 | } | |
05f3fb3c JA |
6236 | static int io_sqe_files_update(struct io_ring_ctx *ctx, void __user *arg, |
6237 | unsigned nr_args) | |
6238 | { | |
6239 | struct io_uring_files_update up; | |
6240 | ||
6241 | if (!ctx->file_data) | |
6242 | return -ENXIO; | |
6243 | if (!nr_args) | |
6244 | return -EINVAL; | |
6245 | if (copy_from_user(&up, arg, sizeof(up))) | |
6246 | return -EFAULT; | |
6247 | if (up.resv) | |
6248 | return -EINVAL; | |
6249 | ||
6250 | return __io_sqe_files_update(ctx, &up, nr_args); | |
6251 | } | |
c3a31e60 | 6252 | |
e9fd9396 | 6253 | static void io_free_work(struct io_wq_work *work) |
7d723065 JA |
6254 | { |
6255 | struct io_kiocb *req = container_of(work, struct io_kiocb, work); | |
6256 | ||
e9fd9396 | 6257 | /* Consider that io_steal_work() relies on this ref */ |
7d723065 JA |
6258 | io_put_req(req); |
6259 | } | |
6260 | ||
24369c2e PB |
6261 | static int io_init_wq_offload(struct io_ring_ctx *ctx, |
6262 | struct io_uring_params *p) | |
6263 | { | |
6264 | struct io_wq_data data; | |
6265 | struct fd f; | |
6266 | struct io_ring_ctx *ctx_attach; | |
6267 | unsigned int concurrency; | |
6268 | int ret = 0; | |
6269 | ||
6270 | data.user = ctx->user; | |
e9fd9396 | 6271 | data.free_work = io_free_work; |
24369c2e PB |
6272 | |
6273 | if (!(p->flags & IORING_SETUP_ATTACH_WQ)) { | |
6274 | /* Do QD, or 4 * CPUS, whatever is smallest */ | |
6275 | concurrency = min(ctx->sq_entries, 4 * num_online_cpus()); | |
6276 | ||
6277 | ctx->io_wq = io_wq_create(concurrency, &data); | |
6278 | if (IS_ERR(ctx->io_wq)) { | |
6279 | ret = PTR_ERR(ctx->io_wq); | |
6280 | ctx->io_wq = NULL; | |
6281 | } | |
6282 | return ret; | |
6283 | } | |
6284 | ||
6285 | f = fdget(p->wq_fd); | |
6286 | if (!f.file) | |
6287 | return -EBADF; | |
6288 | ||
6289 | if (f.file->f_op != &io_uring_fops) { | |
6290 | ret = -EINVAL; | |
6291 | goto out_fput; | |
6292 | } | |
6293 | ||
6294 | ctx_attach = f.file->private_data; | |
6295 | /* @io_wq is protected by holding the fd */ | |
6296 | if (!io_wq_get(ctx_attach->io_wq, &data)) { | |
6297 | ret = -EINVAL; | |
6298 | goto out_fput; | |
6299 | } | |
6300 | ||
6301 | ctx->io_wq = ctx_attach->io_wq; | |
6302 | out_fput: | |
6303 | fdput(f); | |
6304 | return ret; | |
6305 | } | |
6306 | ||
6c271ce2 JA |
6307 | static int io_sq_offload_start(struct io_ring_ctx *ctx, |
6308 | struct io_uring_params *p) | |
2b188cc1 JA |
6309 | { |
6310 | int ret; | |
6311 | ||
6c271ce2 | 6312 | init_waitqueue_head(&ctx->sqo_wait); |
2b188cc1 JA |
6313 | mmgrab(current->mm); |
6314 | ctx->sqo_mm = current->mm; | |
6315 | ||
6c271ce2 | 6316 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
3ec482d1 JA |
6317 | ret = -EPERM; |
6318 | if (!capable(CAP_SYS_ADMIN)) | |
6319 | goto err; | |
6320 | ||
917257da JA |
6321 | ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle); |
6322 | if (!ctx->sq_thread_idle) | |
6323 | ctx->sq_thread_idle = HZ; | |
6324 | ||
6c271ce2 | 6325 | if (p->flags & IORING_SETUP_SQ_AFF) { |
44a9bd18 | 6326 | int cpu = p->sq_thread_cpu; |
6c271ce2 | 6327 | |
917257da | 6328 | ret = -EINVAL; |
44a9bd18 JA |
6329 | if (cpu >= nr_cpu_ids) |
6330 | goto err; | |
7889f44d | 6331 | if (!cpu_online(cpu)) |
917257da JA |
6332 | goto err; |
6333 | ||
6c271ce2 JA |
6334 | ctx->sqo_thread = kthread_create_on_cpu(io_sq_thread, |
6335 | ctx, cpu, | |
6336 | "io_uring-sq"); | |
6337 | } else { | |
6338 | ctx->sqo_thread = kthread_create(io_sq_thread, ctx, | |
6339 | "io_uring-sq"); | |
6340 | } | |
6341 | if (IS_ERR(ctx->sqo_thread)) { | |
6342 | ret = PTR_ERR(ctx->sqo_thread); | |
6343 | ctx->sqo_thread = NULL; | |
6344 | goto err; | |
6345 | } | |
6346 | wake_up_process(ctx->sqo_thread); | |
6347 | } else if (p->flags & IORING_SETUP_SQ_AFF) { | |
6348 | /* Can't have SQ_AFF without SQPOLL */ | |
6349 | ret = -EINVAL; | |
6350 | goto err; | |
6351 | } | |
6352 | ||
24369c2e PB |
6353 | ret = io_init_wq_offload(ctx, p); |
6354 | if (ret) | |
2b188cc1 | 6355 | goto err; |
2b188cc1 JA |
6356 | |
6357 | return 0; | |
6358 | err: | |
54a91f3b | 6359 | io_finish_async(ctx); |
2b188cc1 JA |
6360 | mmdrop(ctx->sqo_mm); |
6361 | ctx->sqo_mm = NULL; | |
6362 | return ret; | |
6363 | } | |
6364 | ||
6365 | static void io_unaccount_mem(struct user_struct *user, unsigned long nr_pages) | |
6366 | { | |
6367 | atomic_long_sub(nr_pages, &user->locked_vm); | |
6368 | } | |
6369 | ||
6370 | static int io_account_mem(struct user_struct *user, unsigned long nr_pages) | |
6371 | { | |
6372 | unsigned long page_limit, cur_pages, new_pages; | |
6373 | ||
6374 | /* Don't allow more pages than we can safely lock */ | |
6375 | page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
6376 | ||
6377 | do { | |
6378 | cur_pages = atomic_long_read(&user->locked_vm); | |
6379 | new_pages = cur_pages + nr_pages; | |
6380 | if (new_pages > page_limit) | |
6381 | return -ENOMEM; | |
6382 | } while (atomic_long_cmpxchg(&user->locked_vm, cur_pages, | |
6383 | new_pages) != cur_pages); | |
6384 | ||
6385 | return 0; | |
6386 | } | |
6387 | ||
6388 | static void io_mem_free(void *ptr) | |
6389 | { | |
52e04ef4 MR |
6390 | struct page *page; |
6391 | ||
6392 | if (!ptr) | |
6393 | return; | |
2b188cc1 | 6394 | |
52e04ef4 | 6395 | page = virt_to_head_page(ptr); |
2b188cc1 JA |
6396 | if (put_page_testzero(page)) |
6397 | free_compound_page(page); | |
6398 | } | |
6399 | ||
6400 | static void *io_mem_alloc(size_t size) | |
6401 | { | |
6402 | gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP | | |
6403 | __GFP_NORETRY; | |
6404 | ||
6405 | return (void *) __get_free_pages(gfp_flags, get_order(size)); | |
6406 | } | |
6407 | ||
75b28aff HV |
6408 | static unsigned long rings_size(unsigned sq_entries, unsigned cq_entries, |
6409 | size_t *sq_offset) | |
6410 | { | |
6411 | struct io_rings *rings; | |
6412 | size_t off, sq_array_size; | |
6413 | ||
6414 | off = struct_size(rings, cqes, cq_entries); | |
6415 | if (off == SIZE_MAX) | |
6416 | return SIZE_MAX; | |
6417 | ||
6418 | #ifdef CONFIG_SMP | |
6419 | off = ALIGN(off, SMP_CACHE_BYTES); | |
6420 | if (off == 0) | |
6421 | return SIZE_MAX; | |
6422 | #endif | |
6423 | ||
6424 | sq_array_size = array_size(sizeof(u32), sq_entries); | |
6425 | if (sq_array_size == SIZE_MAX) | |
6426 | return SIZE_MAX; | |
6427 | ||
6428 | if (check_add_overflow(off, sq_array_size, &off)) | |
6429 | return SIZE_MAX; | |
6430 | ||
6431 | if (sq_offset) | |
6432 | *sq_offset = off; | |
6433 | ||
6434 | return off; | |
6435 | } | |
6436 | ||
2b188cc1 JA |
6437 | static unsigned long ring_pages(unsigned sq_entries, unsigned cq_entries) |
6438 | { | |
75b28aff | 6439 | size_t pages; |
2b188cc1 | 6440 | |
75b28aff HV |
6441 | pages = (size_t)1 << get_order( |
6442 | rings_size(sq_entries, cq_entries, NULL)); | |
6443 | pages += (size_t)1 << get_order( | |
6444 | array_size(sizeof(struct io_uring_sqe), sq_entries)); | |
2b188cc1 | 6445 | |
75b28aff | 6446 | return pages; |
2b188cc1 JA |
6447 | } |
6448 | ||
edafccee JA |
6449 | static int io_sqe_buffer_unregister(struct io_ring_ctx *ctx) |
6450 | { | |
6451 | int i, j; | |
6452 | ||
6453 | if (!ctx->user_bufs) | |
6454 | return -ENXIO; | |
6455 | ||
6456 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
6457 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
6458 | ||
6459 | for (j = 0; j < imu->nr_bvecs; j++) | |
f1f6a7dd | 6460 | unpin_user_page(imu->bvec[j].bv_page); |
edafccee JA |
6461 | |
6462 | if (ctx->account_mem) | |
6463 | io_unaccount_mem(ctx->user, imu->nr_bvecs); | |
d4ef6475 | 6464 | kvfree(imu->bvec); |
edafccee JA |
6465 | imu->nr_bvecs = 0; |
6466 | } | |
6467 | ||
6468 | kfree(ctx->user_bufs); | |
6469 | ctx->user_bufs = NULL; | |
6470 | ctx->nr_user_bufs = 0; | |
6471 | return 0; | |
6472 | } | |
6473 | ||
6474 | static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, | |
6475 | void __user *arg, unsigned index) | |
6476 | { | |
6477 | struct iovec __user *src; | |
6478 | ||
6479 | #ifdef CONFIG_COMPAT | |
6480 | if (ctx->compat) { | |
6481 | struct compat_iovec __user *ciovs; | |
6482 | struct compat_iovec ciov; | |
6483 | ||
6484 | ciovs = (struct compat_iovec __user *) arg; | |
6485 | if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) | |
6486 | return -EFAULT; | |
6487 | ||
d55e5f5b | 6488 | dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base); |
edafccee JA |
6489 | dst->iov_len = ciov.iov_len; |
6490 | return 0; | |
6491 | } | |
6492 | #endif | |
6493 | src = (struct iovec __user *) arg; | |
6494 | if (copy_from_user(dst, &src[index], sizeof(*dst))) | |
6495 | return -EFAULT; | |
6496 | return 0; | |
6497 | } | |
6498 | ||
6499 | static int io_sqe_buffer_register(struct io_ring_ctx *ctx, void __user *arg, | |
6500 | unsigned nr_args) | |
6501 | { | |
6502 | struct vm_area_struct **vmas = NULL; | |
6503 | struct page **pages = NULL; | |
6504 | int i, j, got_pages = 0; | |
6505 | int ret = -EINVAL; | |
6506 | ||
6507 | if (ctx->user_bufs) | |
6508 | return -EBUSY; | |
6509 | if (!nr_args || nr_args > UIO_MAXIOV) | |
6510 | return -EINVAL; | |
6511 | ||
6512 | ctx->user_bufs = kcalloc(nr_args, sizeof(struct io_mapped_ubuf), | |
6513 | GFP_KERNEL); | |
6514 | if (!ctx->user_bufs) | |
6515 | return -ENOMEM; | |
6516 | ||
6517 | for (i = 0; i < nr_args; i++) { | |
6518 | struct io_mapped_ubuf *imu = &ctx->user_bufs[i]; | |
6519 | unsigned long off, start, end, ubuf; | |
6520 | int pret, nr_pages; | |
6521 | struct iovec iov; | |
6522 | size_t size; | |
6523 | ||
6524 | ret = io_copy_iov(ctx, &iov, arg, i); | |
6525 | if (ret) | |
a278682d | 6526 | goto err; |
edafccee JA |
6527 | |
6528 | /* | |
6529 | * Don't impose further limits on the size and buffer | |
6530 | * constraints here, we'll -EINVAL later when IO is | |
6531 | * submitted if they are wrong. | |
6532 | */ | |
6533 | ret = -EFAULT; | |
6534 | if (!iov.iov_base || !iov.iov_len) | |
6535 | goto err; | |
6536 | ||
6537 | /* arbitrary limit, but we need something */ | |
6538 | if (iov.iov_len > SZ_1G) | |
6539 | goto err; | |
6540 | ||
6541 | ubuf = (unsigned long) iov.iov_base; | |
6542 | end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
6543 | start = ubuf >> PAGE_SHIFT; | |
6544 | nr_pages = end - start; | |
6545 | ||
6546 | if (ctx->account_mem) { | |
6547 | ret = io_account_mem(ctx->user, nr_pages); | |
6548 | if (ret) | |
6549 | goto err; | |
6550 | } | |
6551 | ||
6552 | ret = 0; | |
6553 | if (!pages || nr_pages > got_pages) { | |
6554 | kfree(vmas); | |
6555 | kfree(pages); | |
d4ef6475 | 6556 | pages = kvmalloc_array(nr_pages, sizeof(struct page *), |
edafccee | 6557 | GFP_KERNEL); |
d4ef6475 | 6558 | vmas = kvmalloc_array(nr_pages, |
edafccee JA |
6559 | sizeof(struct vm_area_struct *), |
6560 | GFP_KERNEL); | |
6561 | if (!pages || !vmas) { | |
6562 | ret = -ENOMEM; | |
6563 | if (ctx->account_mem) | |
6564 | io_unaccount_mem(ctx->user, nr_pages); | |
6565 | goto err; | |
6566 | } | |
6567 | got_pages = nr_pages; | |
6568 | } | |
6569 | ||
d4ef6475 | 6570 | imu->bvec = kvmalloc_array(nr_pages, sizeof(struct bio_vec), |
edafccee JA |
6571 | GFP_KERNEL); |
6572 | ret = -ENOMEM; | |
6573 | if (!imu->bvec) { | |
6574 | if (ctx->account_mem) | |
6575 | io_unaccount_mem(ctx->user, nr_pages); | |
6576 | goto err; | |
6577 | } | |
6578 | ||
6579 | ret = 0; | |
6580 | down_read(¤t->mm->mmap_sem); | |
2113b05d | 6581 | pret = pin_user_pages(ubuf, nr_pages, |
932f4a63 IW |
6582 | FOLL_WRITE | FOLL_LONGTERM, |
6583 | pages, vmas); | |
edafccee JA |
6584 | if (pret == nr_pages) { |
6585 | /* don't support file backed memory */ | |
6586 | for (j = 0; j < nr_pages; j++) { | |
6587 | struct vm_area_struct *vma = vmas[j]; | |
6588 | ||
6589 | if (vma->vm_file && | |
6590 | !is_file_hugepages(vma->vm_file)) { | |
6591 | ret = -EOPNOTSUPP; | |
6592 | break; | |
6593 | } | |
6594 | } | |
6595 | } else { | |
6596 | ret = pret < 0 ? pret : -EFAULT; | |
6597 | } | |
6598 | up_read(¤t->mm->mmap_sem); | |
6599 | if (ret) { | |
6600 | /* | |
6601 | * if we did partial map, or found file backed vmas, | |
6602 | * release any pages we did get | |
6603 | */ | |
27c4d3a3 | 6604 | if (pret > 0) |
f1f6a7dd | 6605 | unpin_user_pages(pages, pret); |
edafccee JA |
6606 | if (ctx->account_mem) |
6607 | io_unaccount_mem(ctx->user, nr_pages); | |
d4ef6475 | 6608 | kvfree(imu->bvec); |
edafccee JA |
6609 | goto err; |
6610 | } | |
6611 | ||
6612 | off = ubuf & ~PAGE_MASK; | |
6613 | size = iov.iov_len; | |
6614 | for (j = 0; j < nr_pages; j++) { | |
6615 | size_t vec_len; | |
6616 | ||
6617 | vec_len = min_t(size_t, size, PAGE_SIZE - off); | |
6618 | imu->bvec[j].bv_page = pages[j]; | |
6619 | imu->bvec[j].bv_len = vec_len; | |
6620 | imu->bvec[j].bv_offset = off; | |
6621 | off = 0; | |
6622 | size -= vec_len; | |
6623 | } | |
6624 | /* store original address for later verification */ | |
6625 | imu->ubuf = ubuf; | |
6626 | imu->len = iov.iov_len; | |
6627 | imu->nr_bvecs = nr_pages; | |
6628 | ||
6629 | ctx->nr_user_bufs++; | |
6630 | } | |
d4ef6475 MR |
6631 | kvfree(pages); |
6632 | kvfree(vmas); | |
edafccee JA |
6633 | return 0; |
6634 | err: | |
d4ef6475 MR |
6635 | kvfree(pages); |
6636 | kvfree(vmas); | |
edafccee JA |
6637 | io_sqe_buffer_unregister(ctx); |
6638 | return ret; | |
6639 | } | |
6640 | ||
9b402849 JA |
6641 | static int io_eventfd_register(struct io_ring_ctx *ctx, void __user *arg) |
6642 | { | |
6643 | __s32 __user *fds = arg; | |
6644 | int fd; | |
6645 | ||
6646 | if (ctx->cq_ev_fd) | |
6647 | return -EBUSY; | |
6648 | ||
6649 | if (copy_from_user(&fd, fds, sizeof(*fds))) | |
6650 | return -EFAULT; | |
6651 | ||
6652 | ctx->cq_ev_fd = eventfd_ctx_fdget(fd); | |
6653 | if (IS_ERR(ctx->cq_ev_fd)) { | |
6654 | int ret = PTR_ERR(ctx->cq_ev_fd); | |
6655 | ctx->cq_ev_fd = NULL; | |
6656 | return ret; | |
6657 | } | |
6658 | ||
6659 | return 0; | |
6660 | } | |
6661 | ||
6662 | static int io_eventfd_unregister(struct io_ring_ctx *ctx) | |
6663 | { | |
6664 | if (ctx->cq_ev_fd) { | |
6665 | eventfd_ctx_put(ctx->cq_ev_fd); | |
6666 | ctx->cq_ev_fd = NULL; | |
6667 | return 0; | |
6668 | } | |
6669 | ||
6670 | return -ENXIO; | |
6671 | } | |
6672 | ||
5a2e745d JA |
6673 | static int __io_destroy_buffers(int id, void *p, void *data) |
6674 | { | |
6675 | struct io_ring_ctx *ctx = data; | |
6676 | struct io_buffer *buf = p; | |
6677 | ||
6678 | /* the head kbuf is the list itself */ | |
6679 | while (!list_empty(&buf->list)) { | |
6680 | struct io_buffer *nxt; | |
6681 | ||
6682 | nxt = list_first_entry(&buf->list, struct io_buffer, list); | |
6683 | list_del(&nxt->list); | |
6684 | kfree(nxt); | |
6685 | } | |
6686 | kfree(buf); | |
6687 | idr_remove(&ctx->io_buffer_idr, id); | |
6688 | return 0; | |
6689 | } | |
6690 | ||
6691 | static void io_destroy_buffers(struct io_ring_ctx *ctx) | |
6692 | { | |
6693 | idr_for_each(&ctx->io_buffer_idr, __io_destroy_buffers, ctx); | |
6694 | idr_destroy(&ctx->io_buffer_idr); | |
6695 | } | |
6696 | ||
2b188cc1 JA |
6697 | static void io_ring_ctx_free(struct io_ring_ctx *ctx) |
6698 | { | |
6b06314c | 6699 | io_finish_async(ctx); |
2b188cc1 JA |
6700 | if (ctx->sqo_mm) |
6701 | mmdrop(ctx->sqo_mm); | |
def596e9 JA |
6702 | |
6703 | io_iopoll_reap_events(ctx); | |
edafccee | 6704 | io_sqe_buffer_unregister(ctx); |
6b06314c | 6705 | io_sqe_files_unregister(ctx); |
9b402849 | 6706 | io_eventfd_unregister(ctx); |
5a2e745d | 6707 | io_destroy_buffers(ctx); |
41726c9a | 6708 | idr_destroy(&ctx->personality_idr); |
def596e9 | 6709 | |
2b188cc1 | 6710 | #if defined(CONFIG_UNIX) |
355e8d26 EB |
6711 | if (ctx->ring_sock) { |
6712 | ctx->ring_sock->file = NULL; /* so that iput() is called */ | |
2b188cc1 | 6713 | sock_release(ctx->ring_sock); |
355e8d26 | 6714 | } |
2b188cc1 JA |
6715 | #endif |
6716 | ||
75b28aff | 6717 | io_mem_free(ctx->rings); |
2b188cc1 | 6718 | io_mem_free(ctx->sq_sqes); |
2b188cc1 JA |
6719 | |
6720 | percpu_ref_exit(&ctx->refs); | |
6721 | if (ctx->account_mem) | |
6722 | io_unaccount_mem(ctx->user, | |
6723 | ring_pages(ctx->sq_entries, ctx->cq_entries)); | |
6724 | free_uid(ctx->user); | |
181e448d | 6725 | put_cred(ctx->creds); |
206aefde | 6726 | kfree(ctx->completions); |
78076bb6 | 6727 | kfree(ctx->cancel_hash); |
0ddf92e8 | 6728 | kmem_cache_free(req_cachep, ctx->fallback_req); |
2b188cc1 JA |
6729 | kfree(ctx); |
6730 | } | |
6731 | ||
6732 | static __poll_t io_uring_poll(struct file *file, poll_table *wait) | |
6733 | { | |
6734 | struct io_ring_ctx *ctx = file->private_data; | |
6735 | __poll_t mask = 0; | |
6736 | ||
6737 | poll_wait(file, &ctx->cq_wait, wait); | |
4f7067c3 SB |
6738 | /* |
6739 | * synchronizes with barrier from wq_has_sleeper call in | |
6740 | * io_commit_cqring | |
6741 | */ | |
2b188cc1 | 6742 | smp_rmb(); |
75b28aff HV |
6743 | if (READ_ONCE(ctx->rings->sq.tail) - ctx->cached_sq_head != |
6744 | ctx->rings->sq_ring_entries) | |
2b188cc1 | 6745 | mask |= EPOLLOUT | EPOLLWRNORM; |
63e5d81f | 6746 | if (io_cqring_events(ctx, false)) |
2b188cc1 JA |
6747 | mask |= EPOLLIN | EPOLLRDNORM; |
6748 | ||
6749 | return mask; | |
6750 | } | |
6751 | ||
6752 | static int io_uring_fasync(int fd, struct file *file, int on) | |
6753 | { | |
6754 | struct io_ring_ctx *ctx = file->private_data; | |
6755 | ||
6756 | return fasync_helper(fd, file, on, &ctx->cq_fasync); | |
6757 | } | |
6758 | ||
071698e1 JA |
6759 | static int io_remove_personalities(int id, void *p, void *data) |
6760 | { | |
6761 | struct io_ring_ctx *ctx = data; | |
6762 | const struct cred *cred; | |
6763 | ||
6764 | cred = idr_remove(&ctx->personality_idr, id); | |
6765 | if (cred) | |
6766 | put_cred(cred); | |
6767 | return 0; | |
6768 | } | |
6769 | ||
2b188cc1 JA |
6770 | static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx) |
6771 | { | |
6772 | mutex_lock(&ctx->uring_lock); | |
6773 | percpu_ref_kill(&ctx->refs); | |
6774 | mutex_unlock(&ctx->uring_lock); | |
6775 | ||
df069d80 JA |
6776 | /* |
6777 | * Wait for sq thread to idle, if we have one. It won't spin on new | |
6778 | * work after we've killed the ctx ref above. This is important to do | |
6779 | * before we cancel existing commands, as the thread could otherwise | |
6780 | * be queueing new work post that. If that's work we need to cancel, | |
6781 | * it could cause shutdown to hang. | |
6782 | */ | |
6783 | while (ctx->sqo_thread && !wq_has_sleeper(&ctx->sqo_wait)) | |
6784 | cpu_relax(); | |
6785 | ||
5262f567 | 6786 | io_kill_timeouts(ctx); |
221c5eb2 | 6787 | io_poll_remove_all(ctx); |
561fb04a JA |
6788 | |
6789 | if (ctx->io_wq) | |
6790 | io_wq_cancel_all(ctx->io_wq); | |
6791 | ||
def596e9 | 6792 | io_iopoll_reap_events(ctx); |
15dff286 JA |
6793 | /* if we failed setting up the ctx, we might not have any rings */ |
6794 | if (ctx->rings) | |
6795 | io_cqring_overflow_flush(ctx, true); | |
071698e1 | 6796 | idr_for_each(&ctx->personality_idr, io_remove_personalities, ctx); |
206aefde | 6797 | wait_for_completion(&ctx->completions[0]); |
2b188cc1 JA |
6798 | io_ring_ctx_free(ctx); |
6799 | } | |
6800 | ||
6801 | static int io_uring_release(struct inode *inode, struct file *file) | |
6802 | { | |
6803 | struct io_ring_ctx *ctx = file->private_data; | |
6804 | ||
6805 | file->private_data = NULL; | |
6806 | io_ring_ctx_wait_and_kill(ctx); | |
6807 | return 0; | |
6808 | } | |
6809 | ||
fcb323cc JA |
6810 | static void io_uring_cancel_files(struct io_ring_ctx *ctx, |
6811 | struct files_struct *files) | |
6812 | { | |
6813 | struct io_kiocb *req; | |
6814 | DEFINE_WAIT(wait); | |
6815 | ||
6816 | while (!list_empty_careful(&ctx->inflight_list)) { | |
768134d4 | 6817 | struct io_kiocb *cancel_req = NULL; |
fcb323cc JA |
6818 | |
6819 | spin_lock_irq(&ctx->inflight_lock); | |
6820 | list_for_each_entry(req, &ctx->inflight_list, inflight_entry) { | |
768134d4 JA |
6821 | if (req->work.files != files) |
6822 | continue; | |
6823 | /* req is being completed, ignore */ | |
6824 | if (!refcount_inc_not_zero(&req->refs)) | |
6825 | continue; | |
6826 | cancel_req = req; | |
6827 | break; | |
fcb323cc | 6828 | } |
768134d4 | 6829 | if (cancel_req) |
fcb323cc | 6830 | prepare_to_wait(&ctx->inflight_wait, &wait, |
768134d4 | 6831 | TASK_UNINTERRUPTIBLE); |
fcb323cc JA |
6832 | spin_unlock_irq(&ctx->inflight_lock); |
6833 | ||
768134d4 JA |
6834 | /* We need to keep going until we don't find a matching req */ |
6835 | if (!cancel_req) | |
fcb323cc | 6836 | break; |
2f6d9b9d | 6837 | |
2ca10259 JA |
6838 | if (cancel_req->flags & REQ_F_OVERFLOW) { |
6839 | spin_lock_irq(&ctx->completion_lock); | |
6840 | list_del(&cancel_req->list); | |
6841 | cancel_req->flags &= ~REQ_F_OVERFLOW; | |
6842 | if (list_empty(&ctx->cq_overflow_list)) { | |
6843 | clear_bit(0, &ctx->sq_check_overflow); | |
6844 | clear_bit(0, &ctx->cq_check_overflow); | |
6845 | } | |
6846 | spin_unlock_irq(&ctx->completion_lock); | |
6847 | ||
6848 | WRITE_ONCE(ctx->rings->cq_overflow, | |
6849 | atomic_inc_return(&ctx->cached_cq_overflow)); | |
6850 | ||
6851 | /* | |
6852 | * Put inflight ref and overflow ref. If that's | |
6853 | * all we had, then we're done with this request. | |
6854 | */ | |
6855 | if (refcount_sub_and_test(2, &cancel_req->refs)) { | |
6856 | io_put_req(cancel_req); | |
6857 | continue; | |
6858 | } | |
6859 | } | |
6860 | ||
2f6d9b9d BL |
6861 | io_wq_cancel_work(ctx->io_wq, &cancel_req->work); |
6862 | io_put_req(cancel_req); | |
fcb323cc JA |
6863 | schedule(); |
6864 | } | |
768134d4 | 6865 | finish_wait(&ctx->inflight_wait, &wait); |
fcb323cc JA |
6866 | } |
6867 | ||
6868 | static int io_uring_flush(struct file *file, void *data) | |
6869 | { | |
6870 | struct io_ring_ctx *ctx = file->private_data; | |
6871 | ||
6872 | io_uring_cancel_files(ctx, data); | |
6ab23144 JA |
6873 | |
6874 | /* | |
6875 | * If the task is going away, cancel work it may have pending | |
6876 | */ | |
6877 | if (fatal_signal_pending(current) || (current->flags & PF_EXITING)) | |
6878 | io_wq_cancel_pid(ctx->io_wq, task_pid_vnr(current)); | |
6879 | ||
fcb323cc JA |
6880 | return 0; |
6881 | } | |
6882 | ||
6c5c240e RP |
6883 | static void *io_uring_validate_mmap_request(struct file *file, |
6884 | loff_t pgoff, size_t sz) | |
2b188cc1 | 6885 | { |
2b188cc1 | 6886 | struct io_ring_ctx *ctx = file->private_data; |
6c5c240e | 6887 | loff_t offset = pgoff << PAGE_SHIFT; |
2b188cc1 JA |
6888 | struct page *page; |
6889 | void *ptr; | |
6890 | ||
6891 | switch (offset) { | |
6892 | case IORING_OFF_SQ_RING: | |
75b28aff HV |
6893 | case IORING_OFF_CQ_RING: |
6894 | ptr = ctx->rings; | |
2b188cc1 JA |
6895 | break; |
6896 | case IORING_OFF_SQES: | |
6897 | ptr = ctx->sq_sqes; | |
6898 | break; | |
2b188cc1 | 6899 | default: |
6c5c240e | 6900 | return ERR_PTR(-EINVAL); |
2b188cc1 JA |
6901 | } |
6902 | ||
6903 | page = virt_to_head_page(ptr); | |
a50b854e | 6904 | if (sz > page_size(page)) |
6c5c240e RP |
6905 | return ERR_PTR(-EINVAL); |
6906 | ||
6907 | return ptr; | |
6908 | } | |
6909 | ||
6910 | #ifdef CONFIG_MMU | |
6911 | ||
6912 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
6913 | { | |
6914 | size_t sz = vma->vm_end - vma->vm_start; | |
6915 | unsigned long pfn; | |
6916 | void *ptr; | |
6917 | ||
6918 | ptr = io_uring_validate_mmap_request(file, vma->vm_pgoff, sz); | |
6919 | if (IS_ERR(ptr)) | |
6920 | return PTR_ERR(ptr); | |
2b188cc1 JA |
6921 | |
6922 | pfn = virt_to_phys(ptr) >> PAGE_SHIFT; | |
6923 | return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot); | |
6924 | } | |
6925 | ||
6c5c240e RP |
6926 | #else /* !CONFIG_MMU */ |
6927 | ||
6928 | static int io_uring_mmap(struct file *file, struct vm_area_struct *vma) | |
6929 | { | |
6930 | return vma->vm_flags & (VM_SHARED | VM_MAYSHARE) ? 0 : -EINVAL; | |
6931 | } | |
6932 | ||
6933 | static unsigned int io_uring_nommu_mmap_capabilities(struct file *file) | |
6934 | { | |
6935 | return NOMMU_MAP_DIRECT | NOMMU_MAP_READ | NOMMU_MAP_WRITE; | |
6936 | } | |
6937 | ||
6938 | static unsigned long io_uring_nommu_get_unmapped_area(struct file *file, | |
6939 | unsigned long addr, unsigned long len, | |
6940 | unsigned long pgoff, unsigned long flags) | |
6941 | { | |
6942 | void *ptr; | |
6943 | ||
6944 | ptr = io_uring_validate_mmap_request(file, pgoff, len); | |
6945 | if (IS_ERR(ptr)) | |
6946 | return PTR_ERR(ptr); | |
6947 | ||
6948 | return (unsigned long) ptr; | |
6949 | } | |
6950 | ||
6951 | #endif /* !CONFIG_MMU */ | |
6952 | ||
2b188cc1 JA |
6953 | SYSCALL_DEFINE6(io_uring_enter, unsigned int, fd, u32, to_submit, |
6954 | u32, min_complete, u32, flags, const sigset_t __user *, sig, | |
6955 | size_t, sigsz) | |
6956 | { | |
6957 | struct io_ring_ctx *ctx; | |
6958 | long ret = -EBADF; | |
6959 | int submitted = 0; | |
6960 | struct fd f; | |
6961 | ||
b41e9852 JA |
6962 | if (current->task_works) |
6963 | task_work_run(); | |
6964 | ||
6c271ce2 | 6965 | if (flags & ~(IORING_ENTER_GETEVENTS | IORING_ENTER_SQ_WAKEUP)) |
2b188cc1 JA |
6966 | return -EINVAL; |
6967 | ||
6968 | f = fdget(fd); | |
6969 | if (!f.file) | |
6970 | return -EBADF; | |
6971 | ||
6972 | ret = -EOPNOTSUPP; | |
6973 | if (f.file->f_op != &io_uring_fops) | |
6974 | goto out_fput; | |
6975 | ||
6976 | ret = -ENXIO; | |
6977 | ctx = f.file->private_data; | |
6978 | if (!percpu_ref_tryget(&ctx->refs)) | |
6979 | goto out_fput; | |
6980 | ||
6c271ce2 JA |
6981 | /* |
6982 | * For SQ polling, the thread will do all submissions and completions. | |
6983 | * Just return the requested submit count, and wake the thread if | |
6984 | * we were asked to. | |
6985 | */ | |
b2a9eada | 6986 | ret = 0; |
6c271ce2 | 6987 | if (ctx->flags & IORING_SETUP_SQPOLL) { |
c1edbf5f JA |
6988 | if (!list_empty_careful(&ctx->cq_overflow_list)) |
6989 | io_cqring_overflow_flush(ctx, false); | |
6c271ce2 JA |
6990 | if (flags & IORING_ENTER_SQ_WAKEUP) |
6991 | wake_up(&ctx->sqo_wait); | |
6992 | submitted = to_submit; | |
b2a9eada | 6993 | } else if (to_submit) { |
ae9428ca | 6994 | struct mm_struct *cur_mm; |
2b188cc1 JA |
6995 | |
6996 | mutex_lock(&ctx->uring_lock); | |
ae9428ca PB |
6997 | /* already have mm, so io_submit_sqes() won't try to grab it */ |
6998 | cur_mm = ctx->sqo_mm; | |
6999 | submitted = io_submit_sqes(ctx, to_submit, f.file, fd, | |
7000 | &cur_mm, false); | |
2b188cc1 | 7001 | mutex_unlock(&ctx->uring_lock); |
7c504e65 PB |
7002 | |
7003 | if (submitted != to_submit) | |
7004 | goto out; | |
2b188cc1 JA |
7005 | } |
7006 | if (flags & IORING_ENTER_GETEVENTS) { | |
def596e9 JA |
7007 | unsigned nr_events = 0; |
7008 | ||
2b188cc1 JA |
7009 | min_complete = min(min_complete, ctx->cq_entries); |
7010 | ||
def596e9 | 7011 | if (ctx->flags & IORING_SETUP_IOPOLL) { |
def596e9 | 7012 | ret = io_iopoll_check(ctx, &nr_events, min_complete); |
def596e9 JA |
7013 | } else { |
7014 | ret = io_cqring_wait(ctx, min_complete, sig, sigsz); | |
7015 | } | |
2b188cc1 JA |
7016 | } |
7017 | ||
7c504e65 | 7018 | out: |
6805b32e | 7019 | percpu_ref_put(&ctx->refs); |
2b188cc1 JA |
7020 | out_fput: |
7021 | fdput(f); | |
7022 | return submitted ? submitted : ret; | |
7023 | } | |
7024 | ||
bebdb65e | 7025 | #ifdef CONFIG_PROC_FS |
87ce955b JA |
7026 | static int io_uring_show_cred(int id, void *p, void *data) |
7027 | { | |
7028 | const struct cred *cred = p; | |
7029 | struct seq_file *m = data; | |
7030 | struct user_namespace *uns = seq_user_ns(m); | |
7031 | struct group_info *gi; | |
7032 | kernel_cap_t cap; | |
7033 | unsigned __capi; | |
7034 | int g; | |
7035 | ||
7036 | seq_printf(m, "%5d\n", id); | |
7037 | seq_put_decimal_ull(m, "\tUid:\t", from_kuid_munged(uns, cred->uid)); | |
7038 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->euid)); | |
7039 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->suid)); | |
7040 | seq_put_decimal_ull(m, "\t\t", from_kuid_munged(uns, cred->fsuid)); | |
7041 | seq_put_decimal_ull(m, "\n\tGid:\t", from_kgid_munged(uns, cred->gid)); | |
7042 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->egid)); | |
7043 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->sgid)); | |
7044 | seq_put_decimal_ull(m, "\t\t", from_kgid_munged(uns, cred->fsgid)); | |
7045 | seq_puts(m, "\n\tGroups:\t"); | |
7046 | gi = cred->group_info; | |
7047 | for (g = 0; g < gi->ngroups; g++) { | |
7048 | seq_put_decimal_ull(m, g ? " " : "", | |
7049 | from_kgid_munged(uns, gi->gid[g])); | |
7050 | } | |
7051 | seq_puts(m, "\n\tCapEff:\t"); | |
7052 | cap = cred->cap_effective; | |
7053 | CAP_FOR_EACH_U32(__capi) | |
7054 | seq_put_hex_ll(m, NULL, cap.cap[CAP_LAST_U32 - __capi], 8); | |
7055 | seq_putc(m, '\n'); | |
7056 | return 0; | |
7057 | } | |
7058 | ||
7059 | static void __io_uring_show_fdinfo(struct io_ring_ctx *ctx, struct seq_file *m) | |
7060 | { | |
7061 | int i; | |
7062 | ||
7063 | mutex_lock(&ctx->uring_lock); | |
7064 | seq_printf(m, "UserFiles:\t%u\n", ctx->nr_user_files); | |
7065 | for (i = 0; i < ctx->nr_user_files; i++) { | |
7066 | struct fixed_file_table *table; | |
7067 | struct file *f; | |
7068 | ||
7069 | table = &ctx->file_data->table[i >> IORING_FILE_TABLE_SHIFT]; | |
7070 | f = table->files[i & IORING_FILE_TABLE_MASK]; | |
7071 | if (f) | |
7072 | seq_printf(m, "%5u: %s\n", i, file_dentry(f)->d_iname); | |
7073 | else | |
7074 | seq_printf(m, "%5u: <none>\n", i); | |
7075 | } | |
7076 | seq_printf(m, "UserBufs:\t%u\n", ctx->nr_user_bufs); | |
7077 | for (i = 0; i < ctx->nr_user_bufs; i++) { | |
7078 | struct io_mapped_ubuf *buf = &ctx->user_bufs[i]; | |
7079 | ||
7080 | seq_printf(m, "%5u: 0x%llx/%u\n", i, buf->ubuf, | |
7081 | (unsigned int) buf->len); | |
7082 | } | |
7083 | if (!idr_is_empty(&ctx->personality_idr)) { | |
7084 | seq_printf(m, "Personalities:\n"); | |
7085 | idr_for_each(&ctx->personality_idr, io_uring_show_cred, m); | |
7086 | } | |
d7718a9d JA |
7087 | seq_printf(m, "PollList:\n"); |
7088 | spin_lock_irq(&ctx->completion_lock); | |
7089 | for (i = 0; i < (1U << ctx->cancel_hash_bits); i++) { | |
7090 | struct hlist_head *list = &ctx->cancel_hash[i]; | |
7091 | struct io_kiocb *req; | |
7092 | ||
7093 | hlist_for_each_entry(req, list, hash_node) | |
7094 | seq_printf(m, " op=%d, task_works=%d\n", req->opcode, | |
7095 | req->task->task_works != NULL); | |
7096 | } | |
7097 | spin_unlock_irq(&ctx->completion_lock); | |
87ce955b JA |
7098 | mutex_unlock(&ctx->uring_lock); |
7099 | } | |
7100 | ||
7101 | static void io_uring_show_fdinfo(struct seq_file *m, struct file *f) | |
7102 | { | |
7103 | struct io_ring_ctx *ctx = f->private_data; | |
7104 | ||
7105 | if (percpu_ref_tryget(&ctx->refs)) { | |
7106 | __io_uring_show_fdinfo(ctx, m); | |
7107 | percpu_ref_put(&ctx->refs); | |
7108 | } | |
7109 | } | |
bebdb65e | 7110 | #endif |
87ce955b | 7111 | |
2b188cc1 JA |
7112 | static const struct file_operations io_uring_fops = { |
7113 | .release = io_uring_release, | |
fcb323cc | 7114 | .flush = io_uring_flush, |
2b188cc1 | 7115 | .mmap = io_uring_mmap, |
6c5c240e RP |
7116 | #ifndef CONFIG_MMU |
7117 | .get_unmapped_area = io_uring_nommu_get_unmapped_area, | |
7118 | .mmap_capabilities = io_uring_nommu_mmap_capabilities, | |
7119 | #endif | |
2b188cc1 JA |
7120 | .poll = io_uring_poll, |
7121 | .fasync = io_uring_fasync, | |
bebdb65e | 7122 | #ifdef CONFIG_PROC_FS |
87ce955b | 7123 | .show_fdinfo = io_uring_show_fdinfo, |
bebdb65e | 7124 | #endif |
2b188cc1 JA |
7125 | }; |
7126 | ||
7127 | static int io_allocate_scq_urings(struct io_ring_ctx *ctx, | |
7128 | struct io_uring_params *p) | |
7129 | { | |
75b28aff HV |
7130 | struct io_rings *rings; |
7131 | size_t size, sq_array_offset; | |
2b188cc1 | 7132 | |
75b28aff HV |
7133 | size = rings_size(p->sq_entries, p->cq_entries, &sq_array_offset); |
7134 | if (size == SIZE_MAX) | |
7135 | return -EOVERFLOW; | |
7136 | ||
7137 | rings = io_mem_alloc(size); | |
7138 | if (!rings) | |
2b188cc1 JA |
7139 | return -ENOMEM; |
7140 | ||
75b28aff HV |
7141 | ctx->rings = rings; |
7142 | ctx->sq_array = (u32 *)((char *)rings + sq_array_offset); | |
7143 | rings->sq_ring_mask = p->sq_entries - 1; | |
7144 | rings->cq_ring_mask = p->cq_entries - 1; | |
7145 | rings->sq_ring_entries = p->sq_entries; | |
7146 | rings->cq_ring_entries = p->cq_entries; | |
7147 | ctx->sq_mask = rings->sq_ring_mask; | |
7148 | ctx->cq_mask = rings->cq_ring_mask; | |
7149 | ctx->sq_entries = rings->sq_ring_entries; | |
7150 | ctx->cq_entries = rings->cq_ring_entries; | |
2b188cc1 JA |
7151 | |
7152 | size = array_size(sizeof(struct io_uring_sqe), p->sq_entries); | |
eb065d30 JA |
7153 | if (size == SIZE_MAX) { |
7154 | io_mem_free(ctx->rings); | |
7155 | ctx->rings = NULL; | |
2b188cc1 | 7156 | return -EOVERFLOW; |
eb065d30 | 7157 | } |
2b188cc1 JA |
7158 | |
7159 | ctx->sq_sqes = io_mem_alloc(size); | |
eb065d30 JA |
7160 | if (!ctx->sq_sqes) { |
7161 | io_mem_free(ctx->rings); | |
7162 | ctx->rings = NULL; | |
2b188cc1 | 7163 | return -ENOMEM; |
eb065d30 | 7164 | } |
2b188cc1 | 7165 | |
2b188cc1 JA |
7166 | return 0; |
7167 | } | |
7168 | ||
7169 | /* | |
7170 | * Allocate an anonymous fd, this is what constitutes the application | |
7171 | * visible backing of an io_uring instance. The application mmaps this | |
7172 | * fd to gain access to the SQ/CQ ring details. If UNIX sockets are enabled, | |
7173 | * we have to tie this fd to a socket for file garbage collection purposes. | |
7174 | */ | |
7175 | static int io_uring_get_fd(struct io_ring_ctx *ctx) | |
7176 | { | |
7177 | struct file *file; | |
7178 | int ret; | |
7179 | ||
7180 | #if defined(CONFIG_UNIX) | |
7181 | ret = sock_create_kern(&init_net, PF_UNIX, SOCK_RAW, IPPROTO_IP, | |
7182 | &ctx->ring_sock); | |
7183 | if (ret) | |
7184 | return ret; | |
7185 | #endif | |
7186 | ||
7187 | ret = get_unused_fd_flags(O_RDWR | O_CLOEXEC); | |
7188 | if (ret < 0) | |
7189 | goto err; | |
7190 | ||
7191 | file = anon_inode_getfile("[io_uring]", &io_uring_fops, ctx, | |
7192 | O_RDWR | O_CLOEXEC); | |
7193 | if (IS_ERR(file)) { | |
7194 | put_unused_fd(ret); | |
7195 | ret = PTR_ERR(file); | |
7196 | goto err; | |
7197 | } | |
7198 | ||
7199 | #if defined(CONFIG_UNIX) | |
7200 | ctx->ring_sock->file = file; | |
7201 | #endif | |
7202 | fd_install(ret, file); | |
7203 | return ret; | |
7204 | err: | |
7205 | #if defined(CONFIG_UNIX) | |
7206 | sock_release(ctx->ring_sock); | |
7207 | ctx->ring_sock = NULL; | |
7208 | #endif | |
7209 | return ret; | |
7210 | } | |
7211 | ||
7212 | static int io_uring_create(unsigned entries, struct io_uring_params *p) | |
7213 | { | |
7214 | struct user_struct *user = NULL; | |
7215 | struct io_ring_ctx *ctx; | |
7216 | bool account_mem; | |
7217 | int ret; | |
7218 | ||
8110c1a6 | 7219 | if (!entries) |
2b188cc1 | 7220 | return -EINVAL; |
8110c1a6 JA |
7221 | if (entries > IORING_MAX_ENTRIES) { |
7222 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
7223 | return -EINVAL; | |
7224 | entries = IORING_MAX_ENTRIES; | |
7225 | } | |
2b188cc1 JA |
7226 | |
7227 | /* | |
7228 | * Use twice as many entries for the CQ ring. It's possible for the | |
7229 | * application to drive a higher depth than the size of the SQ ring, | |
7230 | * since the sqes are only used at submission time. This allows for | |
33a107f0 JA |
7231 | * some flexibility in overcommitting a bit. If the application has |
7232 | * set IORING_SETUP_CQSIZE, it will have passed in the desired number | |
7233 | * of CQ ring entries manually. | |
2b188cc1 JA |
7234 | */ |
7235 | p->sq_entries = roundup_pow_of_two(entries); | |
33a107f0 JA |
7236 | if (p->flags & IORING_SETUP_CQSIZE) { |
7237 | /* | |
7238 | * If IORING_SETUP_CQSIZE is set, we do the same roundup | |
7239 | * to a power-of-two, if it isn't already. We do NOT impose | |
7240 | * any cq vs sq ring sizing. | |
7241 | */ | |
8110c1a6 | 7242 | if (p->cq_entries < p->sq_entries) |
33a107f0 | 7243 | return -EINVAL; |
8110c1a6 JA |
7244 | if (p->cq_entries > IORING_MAX_CQ_ENTRIES) { |
7245 | if (!(p->flags & IORING_SETUP_CLAMP)) | |
7246 | return -EINVAL; | |
7247 | p->cq_entries = IORING_MAX_CQ_ENTRIES; | |
7248 | } | |
33a107f0 JA |
7249 | p->cq_entries = roundup_pow_of_two(p->cq_entries); |
7250 | } else { | |
7251 | p->cq_entries = 2 * p->sq_entries; | |
7252 | } | |
2b188cc1 JA |
7253 | |
7254 | user = get_uid(current_user()); | |
7255 | account_mem = !capable(CAP_IPC_LOCK); | |
7256 | ||
7257 | if (account_mem) { | |
7258 | ret = io_account_mem(user, | |
7259 | ring_pages(p->sq_entries, p->cq_entries)); | |
7260 | if (ret) { | |
7261 | free_uid(user); | |
7262 | return ret; | |
7263 | } | |
7264 | } | |
7265 | ||
7266 | ctx = io_ring_ctx_alloc(p); | |
7267 | if (!ctx) { | |
7268 | if (account_mem) | |
7269 | io_unaccount_mem(user, ring_pages(p->sq_entries, | |
7270 | p->cq_entries)); | |
7271 | free_uid(user); | |
7272 | return -ENOMEM; | |
7273 | } | |
7274 | ctx->compat = in_compat_syscall(); | |
7275 | ctx->account_mem = account_mem; | |
7276 | ctx->user = user; | |
0b8c0ec7 | 7277 | ctx->creds = get_current_cred(); |
2b188cc1 JA |
7278 | |
7279 | ret = io_allocate_scq_urings(ctx, p); | |
7280 | if (ret) | |
7281 | goto err; | |
7282 | ||
6c271ce2 | 7283 | ret = io_sq_offload_start(ctx, p); |
2b188cc1 JA |
7284 | if (ret) |
7285 | goto err; | |
7286 | ||
2b188cc1 | 7287 | memset(&p->sq_off, 0, sizeof(p->sq_off)); |
75b28aff HV |
7288 | p->sq_off.head = offsetof(struct io_rings, sq.head); |
7289 | p->sq_off.tail = offsetof(struct io_rings, sq.tail); | |
7290 | p->sq_off.ring_mask = offsetof(struct io_rings, sq_ring_mask); | |
7291 | p->sq_off.ring_entries = offsetof(struct io_rings, sq_ring_entries); | |
7292 | p->sq_off.flags = offsetof(struct io_rings, sq_flags); | |
7293 | p->sq_off.dropped = offsetof(struct io_rings, sq_dropped); | |
7294 | p->sq_off.array = (char *)ctx->sq_array - (char *)ctx->rings; | |
2b188cc1 JA |
7295 | |
7296 | memset(&p->cq_off, 0, sizeof(p->cq_off)); | |
75b28aff HV |
7297 | p->cq_off.head = offsetof(struct io_rings, cq.head); |
7298 | p->cq_off.tail = offsetof(struct io_rings, cq.tail); | |
7299 | p->cq_off.ring_mask = offsetof(struct io_rings, cq_ring_mask); | |
7300 | p->cq_off.ring_entries = offsetof(struct io_rings, cq_ring_entries); | |
7301 | p->cq_off.overflow = offsetof(struct io_rings, cq_overflow); | |
7302 | p->cq_off.cqes = offsetof(struct io_rings, cqes); | |
ac90f249 | 7303 | |
044c1ab3 JA |
7304 | /* |
7305 | * Install ring fd as the very last thing, so we don't risk someone | |
7306 | * having closed it before we finish setup | |
7307 | */ | |
7308 | ret = io_uring_get_fd(ctx); | |
7309 | if (ret < 0) | |
7310 | goto err; | |
7311 | ||
da8c9690 | 7312 | p->features = IORING_FEAT_SINGLE_MMAP | IORING_FEAT_NODROP | |
cccf0ee8 | 7313 | IORING_FEAT_SUBMIT_STABLE | IORING_FEAT_RW_CUR_POS | |
d7718a9d | 7314 | IORING_FEAT_CUR_PERSONALITY | IORING_FEAT_FAST_POLL; |
c826bd7a | 7315 | trace_io_uring_create(ret, ctx, p->sq_entries, p->cq_entries, p->flags); |
2b188cc1 JA |
7316 | return ret; |
7317 | err: | |
7318 | io_ring_ctx_wait_and_kill(ctx); | |
7319 | return ret; | |
7320 | } | |
7321 | ||
7322 | /* | |
7323 | * Sets up an aio uring context, and returns the fd. Applications asks for a | |
7324 | * ring size, we return the actual sq/cq ring sizes (among other things) in the | |
7325 | * params structure passed in. | |
7326 | */ | |
7327 | static long io_uring_setup(u32 entries, struct io_uring_params __user *params) | |
7328 | { | |
7329 | struct io_uring_params p; | |
7330 | long ret; | |
7331 | int i; | |
7332 | ||
7333 | if (copy_from_user(&p, params, sizeof(p))) | |
7334 | return -EFAULT; | |
7335 | for (i = 0; i < ARRAY_SIZE(p.resv); i++) { | |
7336 | if (p.resv[i]) | |
7337 | return -EINVAL; | |
7338 | } | |
7339 | ||
6c271ce2 | 7340 | if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQPOLL | |
8110c1a6 | 7341 | IORING_SETUP_SQ_AFF | IORING_SETUP_CQSIZE | |
24369c2e | 7342 | IORING_SETUP_CLAMP | IORING_SETUP_ATTACH_WQ)) |
2b188cc1 JA |
7343 | return -EINVAL; |
7344 | ||
7345 | ret = io_uring_create(entries, &p); | |
7346 | if (ret < 0) | |
7347 | return ret; | |
7348 | ||
7349 | if (copy_to_user(params, &p, sizeof(p))) | |
7350 | return -EFAULT; | |
7351 | ||
7352 | return ret; | |
7353 | } | |
7354 | ||
7355 | SYSCALL_DEFINE2(io_uring_setup, u32, entries, | |
7356 | struct io_uring_params __user *, params) | |
7357 | { | |
7358 | return io_uring_setup(entries, params); | |
7359 | } | |
7360 | ||
66f4af93 JA |
7361 | static int io_probe(struct io_ring_ctx *ctx, void __user *arg, unsigned nr_args) |
7362 | { | |
7363 | struct io_uring_probe *p; | |
7364 | size_t size; | |
7365 | int i, ret; | |
7366 | ||
7367 | size = struct_size(p, ops, nr_args); | |
7368 | if (size == SIZE_MAX) | |
7369 | return -EOVERFLOW; | |
7370 | p = kzalloc(size, GFP_KERNEL); | |
7371 | if (!p) | |
7372 | return -ENOMEM; | |
7373 | ||
7374 | ret = -EFAULT; | |
7375 | if (copy_from_user(p, arg, size)) | |
7376 | goto out; | |
7377 | ret = -EINVAL; | |
7378 | if (memchr_inv(p, 0, size)) | |
7379 | goto out; | |
7380 | ||
7381 | p->last_op = IORING_OP_LAST - 1; | |
7382 | if (nr_args > IORING_OP_LAST) | |
7383 | nr_args = IORING_OP_LAST; | |
7384 | ||
7385 | for (i = 0; i < nr_args; i++) { | |
7386 | p->ops[i].op = i; | |
7387 | if (!io_op_defs[i].not_supported) | |
7388 | p->ops[i].flags = IO_URING_OP_SUPPORTED; | |
7389 | } | |
7390 | p->ops_len = i; | |
7391 | ||
7392 | ret = 0; | |
7393 | if (copy_to_user(arg, p, size)) | |
7394 | ret = -EFAULT; | |
7395 | out: | |
7396 | kfree(p); | |
7397 | return ret; | |
7398 | } | |
7399 | ||
071698e1 JA |
7400 | static int io_register_personality(struct io_ring_ctx *ctx) |
7401 | { | |
7402 | const struct cred *creds = get_current_cred(); | |
7403 | int id; | |
7404 | ||
7405 | id = idr_alloc_cyclic(&ctx->personality_idr, (void *) creds, 1, | |
7406 | USHRT_MAX, GFP_KERNEL); | |
7407 | if (id < 0) | |
7408 | put_cred(creds); | |
7409 | return id; | |
7410 | } | |
7411 | ||
7412 | static int io_unregister_personality(struct io_ring_ctx *ctx, unsigned id) | |
7413 | { | |
7414 | const struct cred *old_creds; | |
7415 | ||
7416 | old_creds = idr_remove(&ctx->personality_idr, id); | |
7417 | if (old_creds) { | |
7418 | put_cred(old_creds); | |
7419 | return 0; | |
7420 | } | |
7421 | ||
7422 | return -EINVAL; | |
7423 | } | |
7424 | ||
7425 | static bool io_register_op_must_quiesce(int op) | |
7426 | { | |
7427 | switch (op) { | |
7428 | case IORING_UNREGISTER_FILES: | |
7429 | case IORING_REGISTER_FILES_UPDATE: | |
7430 | case IORING_REGISTER_PROBE: | |
7431 | case IORING_REGISTER_PERSONALITY: | |
7432 | case IORING_UNREGISTER_PERSONALITY: | |
7433 | return false; | |
7434 | default: | |
7435 | return true; | |
7436 | } | |
7437 | } | |
7438 | ||
edafccee JA |
7439 | static int __io_uring_register(struct io_ring_ctx *ctx, unsigned opcode, |
7440 | void __user *arg, unsigned nr_args) | |
b19062a5 JA |
7441 | __releases(ctx->uring_lock) |
7442 | __acquires(ctx->uring_lock) | |
edafccee JA |
7443 | { |
7444 | int ret; | |
7445 | ||
35fa71a0 JA |
7446 | /* |
7447 | * We're inside the ring mutex, if the ref is already dying, then | |
7448 | * someone else killed the ctx or is already going through | |
7449 | * io_uring_register(). | |
7450 | */ | |
7451 | if (percpu_ref_is_dying(&ctx->refs)) | |
7452 | return -ENXIO; | |
7453 | ||
071698e1 | 7454 | if (io_register_op_must_quiesce(opcode)) { |
05f3fb3c | 7455 | percpu_ref_kill(&ctx->refs); |
b19062a5 | 7456 | |
05f3fb3c JA |
7457 | /* |
7458 | * Drop uring mutex before waiting for references to exit. If | |
7459 | * another thread is currently inside io_uring_enter() it might | |
7460 | * need to grab the uring_lock to make progress. If we hold it | |
7461 | * here across the drain wait, then we can deadlock. It's safe | |
7462 | * to drop the mutex here, since no new references will come in | |
7463 | * after we've killed the percpu ref. | |
7464 | */ | |
7465 | mutex_unlock(&ctx->uring_lock); | |
c150368b | 7466 | ret = wait_for_completion_interruptible(&ctx->completions[0]); |
05f3fb3c | 7467 | mutex_lock(&ctx->uring_lock); |
c150368b JA |
7468 | if (ret) { |
7469 | percpu_ref_resurrect(&ctx->refs); | |
7470 | ret = -EINTR; | |
7471 | goto out; | |
7472 | } | |
05f3fb3c | 7473 | } |
edafccee JA |
7474 | |
7475 | switch (opcode) { | |
7476 | case IORING_REGISTER_BUFFERS: | |
7477 | ret = io_sqe_buffer_register(ctx, arg, nr_args); | |
7478 | break; | |
7479 | case IORING_UNREGISTER_BUFFERS: | |
7480 | ret = -EINVAL; | |
7481 | if (arg || nr_args) | |
7482 | break; | |
7483 | ret = io_sqe_buffer_unregister(ctx); | |
7484 | break; | |
6b06314c JA |
7485 | case IORING_REGISTER_FILES: |
7486 | ret = io_sqe_files_register(ctx, arg, nr_args); | |
7487 | break; | |
7488 | case IORING_UNREGISTER_FILES: | |
7489 | ret = -EINVAL; | |
7490 | if (arg || nr_args) | |
7491 | break; | |
7492 | ret = io_sqe_files_unregister(ctx); | |
7493 | break; | |
c3a31e60 JA |
7494 | case IORING_REGISTER_FILES_UPDATE: |
7495 | ret = io_sqe_files_update(ctx, arg, nr_args); | |
7496 | break; | |
9b402849 | 7497 | case IORING_REGISTER_EVENTFD: |
f2842ab5 | 7498 | case IORING_REGISTER_EVENTFD_ASYNC: |
9b402849 JA |
7499 | ret = -EINVAL; |
7500 | if (nr_args != 1) | |
7501 | break; | |
7502 | ret = io_eventfd_register(ctx, arg); | |
f2842ab5 JA |
7503 | if (ret) |
7504 | break; | |
7505 | if (opcode == IORING_REGISTER_EVENTFD_ASYNC) | |
7506 | ctx->eventfd_async = 1; | |
7507 | else | |
7508 | ctx->eventfd_async = 0; | |
9b402849 JA |
7509 | break; |
7510 | case IORING_UNREGISTER_EVENTFD: | |
7511 | ret = -EINVAL; | |
7512 | if (arg || nr_args) | |
7513 | break; | |
7514 | ret = io_eventfd_unregister(ctx); | |
7515 | break; | |
66f4af93 JA |
7516 | case IORING_REGISTER_PROBE: |
7517 | ret = -EINVAL; | |
7518 | if (!arg || nr_args > 256) | |
7519 | break; | |
7520 | ret = io_probe(ctx, arg, nr_args); | |
7521 | break; | |
071698e1 JA |
7522 | case IORING_REGISTER_PERSONALITY: |
7523 | ret = -EINVAL; | |
7524 | if (arg || nr_args) | |
7525 | break; | |
7526 | ret = io_register_personality(ctx); | |
7527 | break; | |
7528 | case IORING_UNREGISTER_PERSONALITY: | |
7529 | ret = -EINVAL; | |
7530 | if (arg) | |
7531 | break; | |
7532 | ret = io_unregister_personality(ctx, nr_args); | |
7533 | break; | |
edafccee JA |
7534 | default: |
7535 | ret = -EINVAL; | |
7536 | break; | |
7537 | } | |
7538 | ||
071698e1 | 7539 | if (io_register_op_must_quiesce(opcode)) { |
05f3fb3c | 7540 | /* bring the ctx back to life */ |
05f3fb3c | 7541 | percpu_ref_reinit(&ctx->refs); |
c150368b JA |
7542 | out: |
7543 | reinit_completion(&ctx->completions[0]); | |
05f3fb3c | 7544 | } |
edafccee JA |
7545 | return ret; |
7546 | } | |
7547 | ||
7548 | SYSCALL_DEFINE4(io_uring_register, unsigned int, fd, unsigned int, opcode, | |
7549 | void __user *, arg, unsigned int, nr_args) | |
7550 | { | |
7551 | struct io_ring_ctx *ctx; | |
7552 | long ret = -EBADF; | |
7553 | struct fd f; | |
7554 | ||
7555 | f = fdget(fd); | |
7556 | if (!f.file) | |
7557 | return -EBADF; | |
7558 | ||
7559 | ret = -EOPNOTSUPP; | |
7560 | if (f.file->f_op != &io_uring_fops) | |
7561 | goto out_fput; | |
7562 | ||
7563 | ctx = f.file->private_data; | |
7564 | ||
7565 | mutex_lock(&ctx->uring_lock); | |
7566 | ret = __io_uring_register(ctx, opcode, arg, nr_args); | |
7567 | mutex_unlock(&ctx->uring_lock); | |
c826bd7a DD |
7568 | trace_io_uring_register(ctx, opcode, ctx->nr_user_files, ctx->nr_user_bufs, |
7569 | ctx->cq_ev_fd != NULL, ret); | |
edafccee JA |
7570 | out_fput: |
7571 | fdput(f); | |
7572 | return ret; | |
7573 | } | |
7574 | ||
2b188cc1 JA |
7575 | static int __init io_uring_init(void) |
7576 | { | |
d7f62e82 SM |
7577 | #define __BUILD_BUG_VERIFY_ELEMENT(stype, eoffset, etype, ename) do { \ |
7578 | BUILD_BUG_ON(offsetof(stype, ename) != eoffset); \ | |
7579 | BUILD_BUG_ON(sizeof(etype) != sizeof_field(stype, ename)); \ | |
7580 | } while (0) | |
7581 | ||
7582 | #define BUILD_BUG_SQE_ELEM(eoffset, etype, ename) \ | |
7583 | __BUILD_BUG_VERIFY_ELEMENT(struct io_uring_sqe, eoffset, etype, ename) | |
7584 | BUILD_BUG_ON(sizeof(struct io_uring_sqe) != 64); | |
7585 | BUILD_BUG_SQE_ELEM(0, __u8, opcode); | |
7586 | BUILD_BUG_SQE_ELEM(1, __u8, flags); | |
7587 | BUILD_BUG_SQE_ELEM(2, __u16, ioprio); | |
7588 | BUILD_BUG_SQE_ELEM(4, __s32, fd); | |
7589 | BUILD_BUG_SQE_ELEM(8, __u64, off); | |
7590 | BUILD_BUG_SQE_ELEM(8, __u64, addr2); | |
7591 | BUILD_BUG_SQE_ELEM(16, __u64, addr); | |
7d67af2c | 7592 | BUILD_BUG_SQE_ELEM(16, __u64, splice_off_in); |
d7f62e82 SM |
7593 | BUILD_BUG_SQE_ELEM(24, __u32, len); |
7594 | BUILD_BUG_SQE_ELEM(28, __kernel_rwf_t, rw_flags); | |
7595 | BUILD_BUG_SQE_ELEM(28, /* compat */ int, rw_flags); | |
7596 | BUILD_BUG_SQE_ELEM(28, /* compat */ __u32, rw_flags); | |
7597 | BUILD_BUG_SQE_ELEM(28, __u32, fsync_flags); | |
7598 | BUILD_BUG_SQE_ELEM(28, __u16, poll_events); | |
7599 | BUILD_BUG_SQE_ELEM(28, __u32, sync_range_flags); | |
7600 | BUILD_BUG_SQE_ELEM(28, __u32, msg_flags); | |
7601 | BUILD_BUG_SQE_ELEM(28, __u32, timeout_flags); | |
7602 | BUILD_BUG_SQE_ELEM(28, __u32, accept_flags); | |
7603 | BUILD_BUG_SQE_ELEM(28, __u32, cancel_flags); | |
7604 | BUILD_BUG_SQE_ELEM(28, __u32, open_flags); | |
7605 | BUILD_BUG_SQE_ELEM(28, __u32, statx_flags); | |
7606 | BUILD_BUG_SQE_ELEM(28, __u32, fadvise_advice); | |
7d67af2c | 7607 | BUILD_BUG_SQE_ELEM(28, __u32, splice_flags); |
d7f62e82 SM |
7608 | BUILD_BUG_SQE_ELEM(32, __u64, user_data); |
7609 | BUILD_BUG_SQE_ELEM(40, __u16, buf_index); | |
7610 | BUILD_BUG_SQE_ELEM(42, __u16, personality); | |
7d67af2c | 7611 | BUILD_BUG_SQE_ELEM(44, __s32, splice_fd_in); |
d7f62e82 | 7612 | |
d3656344 | 7613 | BUILD_BUG_ON(ARRAY_SIZE(io_op_defs) != IORING_OP_LAST); |
2b188cc1 JA |
7614 | req_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC); |
7615 | return 0; | |
7616 | }; | |
7617 | __initcall(io_uring_init); |