]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/fs/read_write.c | |
3 | * | |
4 | * Copyright (C) 1991, 1992 Linus Torvalds | |
5 | */ | |
6 | ||
7 | #include <linux/slab.h> | |
8 | #include <linux/stat.h> | |
9 | #include <linux/fcntl.h> | |
10 | #include <linux/file.h> | |
11 | #include <linux/uio.h> | |
0eeca283 | 12 | #include <linux/fsnotify.h> |
1da177e4 | 13 | #include <linux/security.h> |
630d9c47 | 14 | #include <linux/export.h> |
1da177e4 | 15 | #include <linux/syscalls.h> |
e28cc715 | 16 | #include <linux/pagemap.h> |
d6b29d7c | 17 | #include <linux/splice.h> |
561c6731 | 18 | #include <linux/compat.h> |
29732938 | 19 | #include <linux/mount.h> |
2feb55f8 | 20 | #include <linux/fs.h> |
06ae43f3 | 21 | #include "internal.h" |
1da177e4 | 22 | |
7c0f6ba6 | 23 | #include <linux/uaccess.h> |
1da177e4 LT |
24 | #include <asm/unistd.h> |
25 | ||
c0bd14af | 26 | typedef ssize_t (*io_fn_t)(struct file *, char __user *, size_t, loff_t *); |
293bc982 | 27 | typedef ssize_t (*iter_fn_t)(struct kiocb *, struct iov_iter *); |
c0bd14af | 28 | |
4b6f5d20 | 29 | const struct file_operations generic_ro_fops = { |
1da177e4 | 30 | .llseek = generic_file_llseek, |
aad4f8bb | 31 | .read_iter = generic_file_read_iter, |
1da177e4 | 32 | .mmap = generic_file_readonly_mmap, |
534f2aaa | 33 | .splice_read = generic_file_splice_read, |
1da177e4 LT |
34 | }; |
35 | ||
36 | EXPORT_SYMBOL(generic_ro_fops); | |
37 | ||
cccb5a1e | 38 | static inline int unsigned_offsets(struct file *file) |
4a3956c7 | 39 | { |
cccb5a1e | 40 | return file->f_mode & FMODE_UNSIGNED_OFFSET; |
4a3956c7 KH |
41 | } |
42 | ||
46a1c2c7 JL |
43 | /** |
44 | * vfs_setpos - update the file offset for lseek | |
45 | * @file: file structure in question | |
46 | * @offset: file offset to seek to | |
47 | * @maxsize: maximum file size | |
48 | * | |
49 | * This is a low-level filesystem helper for updating the file offset to | |
50 | * the value specified by @offset if the given offset is valid and it is | |
51 | * not equal to the current file offset. | |
52 | * | |
53 | * Return the specified offset on success and -EINVAL on invalid offset. | |
54 | */ | |
55 | loff_t vfs_setpos(struct file *file, loff_t offset, loff_t maxsize) | |
ef3d0fd2 AK |
56 | { |
57 | if (offset < 0 && !unsigned_offsets(file)) | |
58 | return -EINVAL; | |
59 | if (offset > maxsize) | |
60 | return -EINVAL; | |
61 | ||
62 | if (offset != file->f_pos) { | |
63 | file->f_pos = offset; | |
64 | file->f_version = 0; | |
65 | } | |
66 | return offset; | |
67 | } | |
46a1c2c7 | 68 | EXPORT_SYMBOL(vfs_setpos); |
ef3d0fd2 | 69 | |
3a8cff4f | 70 | /** |
5760495a | 71 | * generic_file_llseek_size - generic llseek implementation for regular files |
3a8cff4f CH |
72 | * @file: file structure to seek on |
73 | * @offset: file offset to seek to | |
965c8e59 | 74 | * @whence: type of seek |
e8b96eb5 ES |
75 | * @size: max size of this file in file system |
76 | * @eof: offset used for SEEK_END position | |
3a8cff4f | 77 | * |
5760495a | 78 | * This is a variant of generic_file_llseek that allows passing in a custom |
e8b96eb5 | 79 | * maximum file size and a custom EOF position, for e.g. hashed directories |
ef3d0fd2 AK |
80 | * |
81 | * Synchronization: | |
5760495a | 82 | * SEEK_SET and SEEK_END are unsynchronized (but atomic on 64bit platforms) |
ef3d0fd2 AK |
83 | * SEEK_CUR is synchronized against other SEEK_CURs, but not read/writes. |
84 | * read/writes behave like SEEK_SET against seeks. | |
3a8cff4f | 85 | */ |
9465efc9 | 86 | loff_t |
965c8e59 | 87 | generic_file_llseek_size(struct file *file, loff_t offset, int whence, |
e8b96eb5 | 88 | loff_t maxsize, loff_t eof) |
1da177e4 | 89 | { |
965c8e59 | 90 | switch (whence) { |
3a8cff4f | 91 | case SEEK_END: |
e8b96eb5 | 92 | offset += eof; |
3a8cff4f CH |
93 | break; |
94 | case SEEK_CUR: | |
5b6f1eb9 AK |
95 | /* |
96 | * Here we special-case the lseek(fd, 0, SEEK_CUR) | |
97 | * position-querying operation. Avoid rewriting the "same" | |
98 | * f_pos value back to the file because a concurrent read(), | |
99 | * write() or lseek() might have altered it | |
100 | */ | |
101 | if (offset == 0) | |
102 | return file->f_pos; | |
ef3d0fd2 AK |
103 | /* |
104 | * f_lock protects against read/modify/write race with other | |
105 | * SEEK_CURs. Note that parallel writes and reads behave | |
106 | * like SEEK_SET. | |
107 | */ | |
108 | spin_lock(&file->f_lock); | |
46a1c2c7 | 109 | offset = vfs_setpos(file, file->f_pos + offset, maxsize); |
ef3d0fd2 AK |
110 | spin_unlock(&file->f_lock); |
111 | return offset; | |
982d8165 JB |
112 | case SEEK_DATA: |
113 | /* | |
114 | * In the generic case the entire file is data, so as long as | |
115 | * offset isn't at the end of the file then the offset is data. | |
116 | */ | |
e8b96eb5 | 117 | if (offset >= eof) |
982d8165 JB |
118 | return -ENXIO; |
119 | break; | |
120 | case SEEK_HOLE: | |
121 | /* | |
122 | * There is a virtual hole at the end of the file, so as long as | |
123 | * offset isn't i_size or larger, return i_size. | |
124 | */ | |
e8b96eb5 | 125 | if (offset >= eof) |
982d8165 | 126 | return -ENXIO; |
e8b96eb5 | 127 | offset = eof; |
982d8165 | 128 | break; |
1da177e4 | 129 | } |
3a8cff4f | 130 | |
46a1c2c7 | 131 | return vfs_setpos(file, offset, maxsize); |
5760495a AK |
132 | } |
133 | EXPORT_SYMBOL(generic_file_llseek_size); | |
134 | ||
135 | /** | |
136 | * generic_file_llseek - generic llseek implementation for regular files | |
137 | * @file: file structure to seek on | |
138 | * @offset: file offset to seek to | |
965c8e59 | 139 | * @whence: type of seek |
5760495a AK |
140 | * |
141 | * This is a generic implemenation of ->llseek useable for all normal local | |
142 | * filesystems. It just updates the file offset to the value specified by | |
546ae2d2 | 143 | * @offset and @whence. |
5760495a | 144 | */ |
965c8e59 | 145 | loff_t generic_file_llseek(struct file *file, loff_t offset, int whence) |
5760495a AK |
146 | { |
147 | struct inode *inode = file->f_mapping->host; | |
148 | ||
965c8e59 | 149 | return generic_file_llseek_size(file, offset, whence, |
e8b96eb5 ES |
150 | inode->i_sb->s_maxbytes, |
151 | i_size_read(inode)); | |
1da177e4 | 152 | } |
9465efc9 | 153 | EXPORT_SYMBOL(generic_file_llseek); |
1da177e4 | 154 | |
1bf9d14d AV |
155 | /** |
156 | * fixed_size_llseek - llseek implementation for fixed-sized devices | |
157 | * @file: file structure to seek on | |
158 | * @offset: file offset to seek to | |
159 | * @whence: type of seek | |
160 | * @size: size of the file | |
161 | * | |
162 | */ | |
163 | loff_t fixed_size_llseek(struct file *file, loff_t offset, int whence, loff_t size) | |
164 | { | |
165 | switch (whence) { | |
166 | case SEEK_SET: case SEEK_CUR: case SEEK_END: | |
167 | return generic_file_llseek_size(file, offset, whence, | |
168 | size, size); | |
169 | default: | |
170 | return -EINVAL; | |
171 | } | |
172 | } | |
173 | EXPORT_SYMBOL(fixed_size_llseek); | |
174 | ||
b25472f9 AV |
175 | /** |
176 | * no_seek_end_llseek - llseek implementation for fixed-sized devices | |
177 | * @file: file structure to seek on | |
178 | * @offset: file offset to seek to | |
179 | * @whence: type of seek | |
180 | * | |
181 | */ | |
182 | loff_t no_seek_end_llseek(struct file *file, loff_t offset, int whence) | |
183 | { | |
184 | switch (whence) { | |
185 | case SEEK_SET: case SEEK_CUR: | |
186 | return generic_file_llseek_size(file, offset, whence, | |
2feb55f8 | 187 | OFFSET_MAX, 0); |
b25472f9 AV |
188 | default: |
189 | return -EINVAL; | |
190 | } | |
191 | } | |
192 | EXPORT_SYMBOL(no_seek_end_llseek); | |
193 | ||
194 | /** | |
195 | * no_seek_end_llseek_size - llseek implementation for fixed-sized devices | |
196 | * @file: file structure to seek on | |
197 | * @offset: file offset to seek to | |
198 | * @whence: type of seek | |
199 | * @size: maximal offset allowed | |
200 | * | |
201 | */ | |
202 | loff_t no_seek_end_llseek_size(struct file *file, loff_t offset, int whence, loff_t size) | |
203 | { | |
204 | switch (whence) { | |
205 | case SEEK_SET: case SEEK_CUR: | |
206 | return generic_file_llseek_size(file, offset, whence, | |
207 | size, 0); | |
208 | default: | |
209 | return -EINVAL; | |
210 | } | |
211 | } | |
212 | EXPORT_SYMBOL(no_seek_end_llseek_size); | |
213 | ||
ae6afc3f B |
214 | /** |
215 | * noop_llseek - No Operation Performed llseek implementation | |
216 | * @file: file structure to seek on | |
217 | * @offset: file offset to seek to | |
965c8e59 | 218 | * @whence: type of seek |
ae6afc3f B |
219 | * |
220 | * This is an implementation of ->llseek useable for the rare special case when | |
221 | * userspace expects the seek to succeed but the (device) file is actually not | |
222 | * able to perform the seek. In this case you use noop_llseek() instead of | |
223 | * falling back to the default implementation of ->llseek. | |
224 | */ | |
965c8e59 | 225 | loff_t noop_llseek(struct file *file, loff_t offset, int whence) |
ae6afc3f B |
226 | { |
227 | return file->f_pos; | |
228 | } | |
229 | EXPORT_SYMBOL(noop_llseek); | |
230 | ||
965c8e59 | 231 | loff_t no_llseek(struct file *file, loff_t offset, int whence) |
1da177e4 LT |
232 | { |
233 | return -ESPIPE; | |
234 | } | |
235 | EXPORT_SYMBOL(no_llseek); | |
236 | ||
965c8e59 | 237 | loff_t default_llseek(struct file *file, loff_t offset, int whence) |
1da177e4 | 238 | { |
496ad9aa | 239 | struct inode *inode = file_inode(file); |
16abef0e | 240 | loff_t retval; |
1da177e4 | 241 | |
5955102c | 242 | inode_lock(inode); |
965c8e59 | 243 | switch (whence) { |
7b8e8924 | 244 | case SEEK_END: |
982d8165 | 245 | offset += i_size_read(inode); |
1da177e4 | 246 | break; |
7b8e8924 | 247 | case SEEK_CUR: |
5b6f1eb9 AK |
248 | if (offset == 0) { |
249 | retval = file->f_pos; | |
250 | goto out; | |
251 | } | |
1da177e4 | 252 | offset += file->f_pos; |
982d8165 JB |
253 | break; |
254 | case SEEK_DATA: | |
255 | /* | |
256 | * In the generic case the entire file is data, so as | |
257 | * long as offset isn't at the end of the file then the | |
258 | * offset is data. | |
259 | */ | |
bacb2d81 DC |
260 | if (offset >= inode->i_size) { |
261 | retval = -ENXIO; | |
262 | goto out; | |
263 | } | |
982d8165 JB |
264 | break; |
265 | case SEEK_HOLE: | |
266 | /* | |
267 | * There is a virtual hole at the end of the file, so | |
268 | * as long as offset isn't i_size or larger, return | |
269 | * i_size. | |
270 | */ | |
bacb2d81 DC |
271 | if (offset >= inode->i_size) { |
272 | retval = -ENXIO; | |
273 | goto out; | |
274 | } | |
982d8165 JB |
275 | offset = inode->i_size; |
276 | break; | |
1da177e4 LT |
277 | } |
278 | retval = -EINVAL; | |
cccb5a1e | 279 | if (offset >= 0 || unsigned_offsets(file)) { |
1da177e4 LT |
280 | if (offset != file->f_pos) { |
281 | file->f_pos = offset; | |
282 | file->f_version = 0; | |
283 | } | |
284 | retval = offset; | |
285 | } | |
5b6f1eb9 | 286 | out: |
5955102c | 287 | inode_unlock(inode); |
1da177e4 LT |
288 | return retval; |
289 | } | |
290 | EXPORT_SYMBOL(default_llseek); | |
291 | ||
965c8e59 | 292 | loff_t vfs_llseek(struct file *file, loff_t offset, int whence) |
1da177e4 LT |
293 | { |
294 | loff_t (*fn)(struct file *, loff_t, int); | |
295 | ||
296 | fn = no_llseek; | |
297 | if (file->f_mode & FMODE_LSEEK) { | |
72c2d531 | 298 | if (file->f_op->llseek) |
1da177e4 LT |
299 | fn = file->f_op->llseek; |
300 | } | |
965c8e59 | 301 | return fn(file, offset, whence); |
1da177e4 LT |
302 | } |
303 | EXPORT_SYMBOL(vfs_llseek); | |
304 | ||
965c8e59 | 305 | SYSCALL_DEFINE3(lseek, unsigned int, fd, off_t, offset, unsigned int, whence) |
1da177e4 LT |
306 | { |
307 | off_t retval; | |
9c225f26 | 308 | struct fd f = fdget_pos(fd); |
2903ff01 AV |
309 | if (!f.file) |
310 | return -EBADF; | |
1da177e4 LT |
311 | |
312 | retval = -EINVAL; | |
965c8e59 AM |
313 | if (whence <= SEEK_MAX) { |
314 | loff_t res = vfs_llseek(f.file, offset, whence); | |
1da177e4 LT |
315 | retval = res; |
316 | if (res != (loff_t)retval) | |
317 | retval = -EOVERFLOW; /* LFS: should only happen on 32 bit platforms */ | |
318 | } | |
9c225f26 | 319 | fdput_pos(f); |
1da177e4 LT |
320 | return retval; |
321 | } | |
322 | ||
561c6731 AV |
323 | #ifdef CONFIG_COMPAT |
324 | COMPAT_SYSCALL_DEFINE3(lseek, unsigned int, fd, compat_off_t, offset, unsigned int, whence) | |
325 | { | |
326 | return sys_lseek(fd, offset, whence); | |
327 | } | |
328 | #endif | |
329 | ||
1da177e4 | 330 | #ifdef __ARCH_WANT_SYS_LLSEEK |
003d7ab4 HC |
331 | SYSCALL_DEFINE5(llseek, unsigned int, fd, unsigned long, offset_high, |
332 | unsigned long, offset_low, loff_t __user *, result, | |
965c8e59 | 333 | unsigned int, whence) |
1da177e4 LT |
334 | { |
335 | int retval; | |
d7a15f8d | 336 | struct fd f = fdget_pos(fd); |
1da177e4 | 337 | loff_t offset; |
1da177e4 | 338 | |
2903ff01 AV |
339 | if (!f.file) |
340 | return -EBADF; | |
1da177e4 LT |
341 | |
342 | retval = -EINVAL; | |
965c8e59 | 343 | if (whence > SEEK_MAX) |
1da177e4 LT |
344 | goto out_putf; |
345 | ||
2903ff01 | 346 | offset = vfs_llseek(f.file, ((loff_t) offset_high << 32) | offset_low, |
965c8e59 | 347 | whence); |
1da177e4 LT |
348 | |
349 | retval = (int)offset; | |
350 | if (offset >= 0) { | |
351 | retval = -EFAULT; | |
352 | if (!copy_to_user(result, &offset, sizeof(offset))) | |
353 | retval = 0; | |
354 | } | |
355 | out_putf: | |
d7a15f8d | 356 | fdput_pos(f); |
1da177e4 LT |
357 | return retval; |
358 | } | |
359 | #endif | |
360 | ||
dbe4e192 CH |
361 | ssize_t vfs_iter_read(struct file *file, struct iov_iter *iter, loff_t *ppos) |
362 | { | |
363 | struct kiocb kiocb; | |
364 | ssize_t ret; | |
365 | ||
366 | if (!file->f_op->read_iter) | |
367 | return -EINVAL; | |
368 | ||
369 | init_sync_kiocb(&kiocb, file); | |
370 | kiocb.ki_pos = *ppos; | |
dbe4e192 CH |
371 | |
372 | iter->type |= READ; | |
373 | ret = file->f_op->read_iter(&kiocb, iter); | |
599bd19b | 374 | BUG_ON(ret == -EIOCBQUEUED); |
dbe4e192 CH |
375 | if (ret > 0) |
376 | *ppos = kiocb.ki_pos; | |
377 | return ret; | |
378 | } | |
379 | EXPORT_SYMBOL(vfs_iter_read); | |
380 | ||
381 | ssize_t vfs_iter_write(struct file *file, struct iov_iter *iter, loff_t *ppos) | |
382 | { | |
383 | struct kiocb kiocb; | |
384 | ssize_t ret; | |
385 | ||
386 | if (!file->f_op->write_iter) | |
387 | return -EINVAL; | |
388 | ||
389 | init_sync_kiocb(&kiocb, file); | |
390 | kiocb.ki_pos = *ppos; | |
dbe4e192 CH |
391 | |
392 | iter->type |= WRITE; | |
393 | ret = file->f_op->write_iter(&kiocb, iter); | |
599bd19b | 394 | BUG_ON(ret == -EIOCBQUEUED); |
dbe4e192 CH |
395 | if (ret > 0) |
396 | *ppos = kiocb.ki_pos; | |
397 | return ret; | |
398 | } | |
399 | EXPORT_SYMBOL(vfs_iter_write); | |
400 | ||
68d70d03 | 401 | int rw_verify_area(int read_write, struct file *file, const loff_t *ppos, size_t count) |
1da177e4 LT |
402 | { |
403 | struct inode *inode; | |
404 | loff_t pos; | |
c43e259c | 405 | int retval = -EINVAL; |
1da177e4 | 406 | |
496ad9aa | 407 | inode = file_inode(file); |
e28cc715 | 408 | if (unlikely((ssize_t) count < 0)) |
c43e259c | 409 | return retval; |
1da177e4 | 410 | pos = *ppos; |
cccb5a1e AV |
411 | if (unlikely(pos < 0)) { |
412 | if (!unsigned_offsets(file)) | |
413 | return retval; | |
414 | if (count >= -pos) /* both values are in 0..LLONG_MAX */ | |
415 | return -EOVERFLOW; | |
416 | } else if (unlikely((loff_t) (pos + count) < 0)) { | |
417 | if (!unsigned_offsets(file)) | |
4a3956c7 KH |
418 | return retval; |
419 | } | |
1da177e4 | 420 | |
bd61e0a9 | 421 | if (unlikely(inode->i_flctx && mandatory_lock(inode))) { |
acc15575 CH |
422 | retval = locks_mandatory_area(inode, file, pos, pos + count - 1, |
423 | read_write == READ ? F_RDLCK : F_WRLCK); | |
e28cc715 LT |
424 | if (retval < 0) |
425 | return retval; | |
426 | } | |
bc61384d | 427 | return security_file_permission(file, |
c43e259c | 428 | read_write == READ ? MAY_READ : MAY_WRITE); |
1da177e4 LT |
429 | } |
430 | ||
5d5d5689 | 431 | static ssize_t new_sync_read(struct file *filp, char __user *buf, size_t len, loff_t *ppos) |
293bc982 AV |
432 | { |
433 | struct iovec iov = { .iov_base = buf, .iov_len = len }; | |
434 | struct kiocb kiocb; | |
435 | struct iov_iter iter; | |
436 | ssize_t ret; | |
437 | ||
438 | init_sync_kiocb(&kiocb, filp); | |
439 | kiocb.ki_pos = *ppos; | |
293bc982 AV |
440 | iov_iter_init(&iter, READ, &iov, 1, len); |
441 | ||
442 | ret = filp->f_op->read_iter(&kiocb, &iter); | |
599bd19b | 443 | BUG_ON(ret == -EIOCBQUEUED); |
293bc982 AV |
444 | *ppos = kiocb.ki_pos; |
445 | return ret; | |
446 | } | |
447 | ||
6fb5032e DK |
448 | ssize_t __vfs_read(struct file *file, char __user *buf, size_t count, |
449 | loff_t *pos) | |
450 | { | |
6fb5032e | 451 | if (file->f_op->read) |
3d04c8a1 | 452 | return file->f_op->read(file, buf, count, pos); |
6fb5032e | 453 | else if (file->f_op->read_iter) |
3d04c8a1 | 454 | return new_sync_read(file, buf, count, pos); |
6fb5032e | 455 | else |
3d04c8a1 | 456 | return -EINVAL; |
6fb5032e | 457 | } |
3d04c8a1 | 458 | EXPORT_SYMBOL(__vfs_read); |
6fb5032e | 459 | |
1da177e4 LT |
460 | ssize_t vfs_read(struct file *file, char __user *buf, size_t count, loff_t *pos) |
461 | { | |
462 | ssize_t ret; | |
463 | ||
464 | if (!(file->f_mode & FMODE_READ)) | |
465 | return -EBADF; | |
7f7f25e8 | 466 | if (!(file->f_mode & FMODE_CAN_READ)) |
1da177e4 LT |
467 | return -EINVAL; |
468 | if (unlikely(!access_ok(VERIFY_WRITE, buf, count))) | |
469 | return -EFAULT; | |
470 | ||
471 | ret = rw_verify_area(READ, file, pos, count); | |
bc61384d AV |
472 | if (!ret) { |
473 | if (count > MAX_RW_COUNT) | |
474 | count = MAX_RW_COUNT; | |
6fb5032e | 475 | ret = __vfs_read(file, buf, count, pos); |
c43e259c | 476 | if (ret > 0) { |
2a12a9d7 | 477 | fsnotify_access(file); |
c43e259c | 478 | add_rchar(current, ret); |
1da177e4 | 479 | } |
c43e259c | 480 | inc_syscr(current); |
1da177e4 LT |
481 | } |
482 | ||
483 | return ret; | |
484 | } | |
485 | ||
486 | EXPORT_SYMBOL(vfs_read); | |
487 | ||
5d5d5689 | 488 | static ssize_t new_sync_write(struct file *filp, const char __user *buf, size_t len, loff_t *ppos) |
293bc982 AV |
489 | { |
490 | struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = len }; | |
491 | struct kiocb kiocb; | |
492 | struct iov_iter iter; | |
493 | ssize_t ret; | |
494 | ||
495 | init_sync_kiocb(&kiocb, filp); | |
496 | kiocb.ki_pos = *ppos; | |
293bc982 AV |
497 | iov_iter_init(&iter, WRITE, &iov, 1, len); |
498 | ||
499 | ret = filp->f_op->write_iter(&kiocb, &iter); | |
599bd19b | 500 | BUG_ON(ret == -EIOCBQUEUED); |
f765b134 AV |
501 | if (ret > 0) |
502 | *ppos = kiocb.ki_pos; | |
293bc982 AV |
503 | return ret; |
504 | } | |
505 | ||
493c84c0 AV |
506 | ssize_t __vfs_write(struct file *file, const char __user *p, size_t count, |
507 | loff_t *pos) | |
508 | { | |
509 | if (file->f_op->write) | |
510 | return file->f_op->write(file, p, count, pos); | |
493c84c0 AV |
511 | else if (file->f_op->write_iter) |
512 | return new_sync_write(file, p, count, pos); | |
513 | else | |
514 | return -EINVAL; | |
515 | } | |
516 | EXPORT_SYMBOL(__vfs_write); | |
517 | ||
e14748e8 SF |
518 | vfs_readf_t vfs_readf(struct file *file) |
519 | { | |
520 | const struct file_operations *fop = file->f_op; | |
521 | ||
522 | if (fop->read) | |
523 | return fop->read; | |
524 | if (fop->read_iter) | |
525 | return new_sync_read; | |
526 | return ERR_PTR(-ENOSYS); | |
527 | } | |
528 | EXPORT_SYMBOL_GPL(vfs_readf); | |
529 | ||
530 | vfs_writef_t vfs_writef(struct file *file) | |
531 | { | |
532 | const struct file_operations *fop = file->f_op; | |
533 | ||
534 | if (fop->write) | |
535 | return fop->write; | |
536 | if (fop->write_iter) | |
537 | return new_sync_write; | |
538 | return ERR_PTR(-ENOSYS); | |
539 | } | |
540 | EXPORT_SYMBOL_GPL(vfs_writef); | |
541 | ||
06ae43f3 AV |
542 | ssize_t __kernel_write(struct file *file, const char *buf, size_t count, loff_t *pos) |
543 | { | |
544 | mm_segment_t old_fs; | |
545 | const char __user *p; | |
546 | ssize_t ret; | |
547 | ||
7f7f25e8 | 548 | if (!(file->f_mode & FMODE_CAN_WRITE)) |
3e84f48e AV |
549 | return -EINVAL; |
550 | ||
06ae43f3 AV |
551 | old_fs = get_fs(); |
552 | set_fs(get_ds()); | |
553 | p = (__force const char __user *)buf; | |
554 | if (count > MAX_RW_COUNT) | |
555 | count = MAX_RW_COUNT; | |
493c84c0 | 556 | ret = __vfs_write(file, p, count, pos); |
06ae43f3 AV |
557 | set_fs(old_fs); |
558 | if (ret > 0) { | |
559 | fsnotify_modify(file); | |
560 | add_wchar(current, ret); | |
561 | } | |
562 | inc_syscw(current); | |
563 | return ret; | |
564 | } | |
565 | ||
2ec3a12a AV |
566 | EXPORT_SYMBOL(__kernel_write); |
567 | ||
1da177e4 LT |
568 | ssize_t vfs_write(struct file *file, const char __user *buf, size_t count, loff_t *pos) |
569 | { | |
570 | ssize_t ret; | |
571 | ||
572 | if (!(file->f_mode & FMODE_WRITE)) | |
573 | return -EBADF; | |
7f7f25e8 | 574 | if (!(file->f_mode & FMODE_CAN_WRITE)) |
1da177e4 LT |
575 | return -EINVAL; |
576 | if (unlikely(!access_ok(VERIFY_READ, buf, count))) | |
577 | return -EFAULT; | |
578 | ||
579 | ret = rw_verify_area(WRITE, file, pos, count); | |
bc61384d AV |
580 | if (!ret) { |
581 | if (count > MAX_RW_COUNT) | |
582 | count = MAX_RW_COUNT; | |
03d95eb2 | 583 | file_start_write(file); |
493c84c0 | 584 | ret = __vfs_write(file, buf, count, pos); |
c43e259c | 585 | if (ret > 0) { |
2a12a9d7 | 586 | fsnotify_modify(file); |
c43e259c | 587 | add_wchar(current, ret); |
1da177e4 | 588 | } |
c43e259c | 589 | inc_syscw(current); |
03d95eb2 | 590 | file_end_write(file); |
1da177e4 LT |
591 | } |
592 | ||
593 | return ret; | |
594 | } | |
595 | ||
596 | EXPORT_SYMBOL(vfs_write); | |
597 | ||
598 | static inline loff_t file_pos_read(struct file *file) | |
599 | { | |
600 | return file->f_pos; | |
601 | } | |
602 | ||
603 | static inline void file_pos_write(struct file *file, loff_t pos) | |
604 | { | |
605 | file->f_pos = pos; | |
606 | } | |
607 | ||
3cdad428 | 608 | SYSCALL_DEFINE3(read, unsigned int, fd, char __user *, buf, size_t, count) |
1da177e4 | 609 | { |
9c225f26 | 610 | struct fd f = fdget_pos(fd); |
1da177e4 | 611 | ssize_t ret = -EBADF; |
1da177e4 | 612 | |
2903ff01 AV |
613 | if (f.file) { |
614 | loff_t pos = file_pos_read(f.file); | |
615 | ret = vfs_read(f.file, buf, count, &pos); | |
5faf153e AV |
616 | if (ret >= 0) |
617 | file_pos_write(f.file, pos); | |
9c225f26 | 618 | fdput_pos(f); |
1da177e4 | 619 | } |
1da177e4 LT |
620 | return ret; |
621 | } | |
1da177e4 | 622 | |
3cdad428 HC |
623 | SYSCALL_DEFINE3(write, unsigned int, fd, const char __user *, buf, |
624 | size_t, count) | |
1da177e4 | 625 | { |
9c225f26 | 626 | struct fd f = fdget_pos(fd); |
1da177e4 | 627 | ssize_t ret = -EBADF; |
1da177e4 | 628 | |
2903ff01 AV |
629 | if (f.file) { |
630 | loff_t pos = file_pos_read(f.file); | |
631 | ret = vfs_write(f.file, buf, count, &pos); | |
5faf153e AV |
632 | if (ret >= 0) |
633 | file_pos_write(f.file, pos); | |
9c225f26 | 634 | fdput_pos(f); |
1da177e4 LT |
635 | } |
636 | ||
637 | return ret; | |
638 | } | |
639 | ||
4a0fd5bf AV |
640 | SYSCALL_DEFINE4(pread64, unsigned int, fd, char __user *, buf, |
641 | size_t, count, loff_t, pos) | |
1da177e4 | 642 | { |
2903ff01 | 643 | struct fd f; |
1da177e4 | 644 | ssize_t ret = -EBADF; |
1da177e4 LT |
645 | |
646 | if (pos < 0) | |
647 | return -EINVAL; | |
648 | ||
2903ff01 AV |
649 | f = fdget(fd); |
650 | if (f.file) { | |
1da177e4 | 651 | ret = -ESPIPE; |
2903ff01 AV |
652 | if (f.file->f_mode & FMODE_PREAD) |
653 | ret = vfs_read(f.file, buf, count, &pos); | |
654 | fdput(f); | |
1da177e4 LT |
655 | } |
656 | ||
657 | return ret; | |
658 | } | |
659 | ||
4a0fd5bf AV |
660 | SYSCALL_DEFINE4(pwrite64, unsigned int, fd, const char __user *, buf, |
661 | size_t, count, loff_t, pos) | |
1da177e4 | 662 | { |
2903ff01 | 663 | struct fd f; |
1da177e4 | 664 | ssize_t ret = -EBADF; |
1da177e4 LT |
665 | |
666 | if (pos < 0) | |
667 | return -EINVAL; | |
668 | ||
2903ff01 AV |
669 | f = fdget(fd); |
670 | if (f.file) { | |
1da177e4 | 671 | ret = -ESPIPE; |
2903ff01 AV |
672 | if (f.file->f_mode & FMODE_PWRITE) |
673 | ret = vfs_write(f.file, buf, count, &pos); | |
674 | fdput(f); | |
1da177e4 LT |
675 | } |
676 | ||
677 | return ret; | |
678 | } | |
679 | ||
680 | /* | |
681 | * Reduce an iovec's length in-place. Return the resulting number of segments | |
682 | */ | |
683 | unsigned long iov_shorten(struct iovec *iov, unsigned long nr_segs, size_t to) | |
684 | { | |
685 | unsigned long seg = 0; | |
686 | size_t len = 0; | |
687 | ||
688 | while (seg < nr_segs) { | |
689 | seg++; | |
690 | if (len + iov->iov_len >= to) { | |
691 | iov->iov_len = to - len; | |
692 | break; | |
693 | } | |
694 | len += iov->iov_len; | |
695 | iov++; | |
696 | } | |
697 | return seg; | |
698 | } | |
19295529 | 699 | EXPORT_SYMBOL(iov_shorten); |
1da177e4 | 700 | |
ac15ac06 | 701 | static ssize_t do_iter_readv_writev(struct file *filp, struct iov_iter *iter, |
793b80ef | 702 | loff_t *ppos, iter_fn_t fn, int flags) |
293bc982 AV |
703 | { |
704 | struct kiocb kiocb; | |
293bc982 AV |
705 | ssize_t ret; |
706 | ||
e864f395 | 707 | if (flags & ~(RWF_HIPRI | RWF_DSYNC | RWF_SYNC)) |
793b80ef CH |
708 | return -EOPNOTSUPP; |
709 | ||
293bc982 | 710 | init_sync_kiocb(&kiocb, filp); |
97be7ebe CH |
711 | if (flags & RWF_HIPRI) |
712 | kiocb.ki_flags |= IOCB_HIPRI; | |
e864f395 CH |
713 | if (flags & RWF_DSYNC) |
714 | kiocb.ki_flags |= IOCB_DSYNC; | |
715 | if (flags & RWF_SYNC) | |
716 | kiocb.ki_flags |= (IOCB_DSYNC | IOCB_SYNC); | |
293bc982 | 717 | kiocb.ki_pos = *ppos; |
293bc982 | 718 | |
ac15ac06 | 719 | ret = fn(&kiocb, iter); |
599bd19b | 720 | BUG_ON(ret == -EIOCBQUEUED); |
293bc982 AV |
721 | *ppos = kiocb.ki_pos; |
722 | return ret; | |
723 | } | |
724 | ||
ee0b3e67 | 725 | /* Do it by hand, with file-ops */ |
ac15ac06 | 726 | static ssize_t do_loop_readv_writev(struct file *filp, struct iov_iter *iter, |
793b80ef | 727 | loff_t *ppos, io_fn_t fn, int flags) |
ee0b3e67 | 728 | { |
ee0b3e67 BP |
729 | ssize_t ret = 0; |
730 | ||
97be7ebe | 731 | if (flags & ~RWF_HIPRI) |
793b80ef CH |
732 | return -EOPNOTSUPP; |
733 | ||
ac15ac06 AV |
734 | while (iov_iter_count(iter)) { |
735 | struct iovec iovec = iov_iter_iovec(iter); | |
ee0b3e67 BP |
736 | ssize_t nr; |
737 | ||
ac15ac06 | 738 | nr = fn(filp, iovec.iov_base, iovec.iov_len, ppos); |
ee0b3e67 BP |
739 | |
740 | if (nr < 0) { | |
741 | if (!ret) | |
742 | ret = nr; | |
743 | break; | |
744 | } | |
745 | ret += nr; | |
ac15ac06 | 746 | if (nr != iovec.iov_len) |
ee0b3e67 | 747 | break; |
ac15ac06 | 748 | iov_iter_advance(iter, nr); |
ee0b3e67 BP |
749 | } |
750 | ||
751 | return ret; | |
752 | } | |
753 | ||
1da177e4 LT |
754 | /* A write operation does a read from user space and vice versa */ |
755 | #define vrfy_dir(type) ((type) == READ ? VERIFY_WRITE : VERIFY_READ) | |
756 | ||
ffecee4f VN |
757 | /** |
758 | * rw_copy_check_uvector() - Copy an array of &struct iovec from userspace | |
759 | * into the kernel and check that it is valid. | |
760 | * | |
761 | * @type: One of %CHECK_IOVEC_ONLY, %READ, or %WRITE. | |
762 | * @uvector: Pointer to the userspace array. | |
763 | * @nr_segs: Number of elements in userspace array. | |
764 | * @fast_segs: Number of elements in @fast_pointer. | |
765 | * @fast_pointer: Pointer to (usually small on-stack) kernel array. | |
766 | * @ret_pointer: (output parameter) Pointer to a variable that will point to | |
767 | * either @fast_pointer, a newly allocated kernel array, or NULL, | |
768 | * depending on which array was used. | |
769 | * | |
770 | * This function copies an array of &struct iovec of @nr_segs from | |
771 | * userspace into the kernel and checks that each element is valid (e.g. | |
772 | * it does not point to a kernel address or cause overflow by being too | |
773 | * large, etc.). | |
774 | * | |
775 | * As an optimization, the caller may provide a pointer to a small | |
776 | * on-stack array in @fast_pointer, typically %UIO_FASTIOV elements long | |
777 | * (the size of this array, or 0 if unused, should be given in @fast_segs). | |
778 | * | |
779 | * @ret_pointer will always point to the array that was used, so the | |
780 | * caller must take care not to call kfree() on it e.g. in case the | |
781 | * @fast_pointer array was used and it was allocated on the stack. | |
782 | * | |
783 | * Return: The total number of bytes covered by the iovec array on success | |
784 | * or a negative error code on error. | |
785 | */ | |
eed4e51f BP |
786 | ssize_t rw_copy_check_uvector(int type, const struct iovec __user * uvector, |
787 | unsigned long nr_segs, unsigned long fast_segs, | |
788 | struct iovec *fast_pointer, | |
ac34ebb3 | 789 | struct iovec **ret_pointer) |
435f49a5 | 790 | { |
eed4e51f | 791 | unsigned long seg; |
435f49a5 | 792 | ssize_t ret; |
eed4e51f BP |
793 | struct iovec *iov = fast_pointer; |
794 | ||
435f49a5 LT |
795 | /* |
796 | * SuS says "The readv() function *may* fail if the iovcnt argument | |
797 | * was less than or equal to 0, or greater than {IOV_MAX}. Linux has | |
798 | * traditionally returned zero for zero segments, so... | |
799 | */ | |
eed4e51f BP |
800 | if (nr_segs == 0) { |
801 | ret = 0; | |
435f49a5 | 802 | goto out; |
eed4e51f BP |
803 | } |
804 | ||
435f49a5 LT |
805 | /* |
806 | * First get the "struct iovec" from user memory and | |
807 | * verify all the pointers | |
808 | */ | |
eed4e51f BP |
809 | if (nr_segs > UIO_MAXIOV) { |
810 | ret = -EINVAL; | |
435f49a5 | 811 | goto out; |
eed4e51f BP |
812 | } |
813 | if (nr_segs > fast_segs) { | |
435f49a5 | 814 | iov = kmalloc(nr_segs*sizeof(struct iovec), GFP_KERNEL); |
eed4e51f BP |
815 | if (iov == NULL) { |
816 | ret = -ENOMEM; | |
435f49a5 | 817 | goto out; |
eed4e51f | 818 | } |
435f49a5 | 819 | } |
eed4e51f BP |
820 | if (copy_from_user(iov, uvector, nr_segs*sizeof(*uvector))) { |
821 | ret = -EFAULT; | |
435f49a5 | 822 | goto out; |
eed4e51f BP |
823 | } |
824 | ||
435f49a5 | 825 | /* |
eed4e51f BP |
826 | * According to the Single Unix Specification we should return EINVAL |
827 | * if an element length is < 0 when cast to ssize_t or if the | |
828 | * total length would overflow the ssize_t return value of the | |
829 | * system call. | |
435f49a5 LT |
830 | * |
831 | * Linux caps all read/write calls to MAX_RW_COUNT, and avoids the | |
832 | * overflow case. | |
833 | */ | |
eed4e51f | 834 | ret = 0; |
435f49a5 LT |
835 | for (seg = 0; seg < nr_segs; seg++) { |
836 | void __user *buf = iov[seg].iov_base; | |
837 | ssize_t len = (ssize_t)iov[seg].iov_len; | |
eed4e51f BP |
838 | |
839 | /* see if we we're about to use an invalid len or if | |
840 | * it's about to overflow ssize_t */ | |
435f49a5 | 841 | if (len < 0) { |
eed4e51f | 842 | ret = -EINVAL; |
435f49a5 | 843 | goto out; |
eed4e51f | 844 | } |
ac34ebb3 | 845 | if (type >= 0 |
fcf63409 | 846 | && unlikely(!access_ok(vrfy_dir(type), buf, len))) { |
eed4e51f | 847 | ret = -EFAULT; |
435f49a5 LT |
848 | goto out; |
849 | } | |
850 | if (len > MAX_RW_COUNT - ret) { | |
851 | len = MAX_RW_COUNT - ret; | |
852 | iov[seg].iov_len = len; | |
eed4e51f | 853 | } |
eed4e51f | 854 | ret += len; |
435f49a5 | 855 | } |
eed4e51f BP |
856 | out: |
857 | *ret_pointer = iov; | |
858 | return ret; | |
859 | } | |
860 | ||
1da177e4 LT |
861 | static ssize_t do_readv_writev(int type, struct file *file, |
862 | const struct iovec __user * uvector, | |
793b80ef CH |
863 | unsigned long nr_segs, loff_t *pos, |
864 | int flags) | |
1da177e4 | 865 | { |
1da177e4 LT |
866 | size_t tot_len; |
867 | struct iovec iovstack[UIO_FASTIOV]; | |
ee0b3e67 | 868 | struct iovec *iov = iovstack; |
ac15ac06 | 869 | struct iov_iter iter; |
1da177e4 | 870 | ssize_t ret; |
1da177e4 | 871 | io_fn_t fn; |
293bc982 | 872 | iter_fn_t iter_fn; |
1da177e4 | 873 | |
0504c074 AV |
874 | ret = import_iovec(type, uvector, nr_segs, |
875 | ARRAY_SIZE(iovstack), &iov, &iter); | |
876 | if (ret < 0) | |
877 | return ret; | |
1da177e4 | 878 | |
0504c074 AV |
879 | tot_len = iov_iter_count(&iter); |
880 | if (!tot_len) | |
881 | goto out; | |
1da177e4 | 882 | ret = rw_verify_area(type, file, pos, tot_len); |
e28cc715 | 883 | if (ret < 0) |
411b67b4 | 884 | goto out; |
1da177e4 | 885 | |
1da177e4 LT |
886 | if (type == READ) { |
887 | fn = file->f_op->read; | |
293bc982 | 888 | iter_fn = file->f_op->read_iter; |
1da177e4 LT |
889 | } else { |
890 | fn = (io_fn_t)file->f_op->write; | |
293bc982 | 891 | iter_fn = file->f_op->write_iter; |
03d95eb2 | 892 | file_start_write(file); |
1da177e4 LT |
893 | } |
894 | ||
293bc982 | 895 | if (iter_fn) |
793b80ef | 896 | ret = do_iter_readv_writev(file, &iter, pos, iter_fn, flags); |
ee0b3e67 | 897 | else |
793b80ef | 898 | ret = do_loop_readv_writev(file, &iter, pos, fn, flags); |
1da177e4 | 899 | |
03d95eb2 AV |
900 | if (type != READ) |
901 | file_end_write(file); | |
902 | ||
1da177e4 | 903 | out: |
0504c074 | 904 | kfree(iov); |
0eeca283 RL |
905 | if ((ret + (type == READ)) > 0) { |
906 | if (type == READ) | |
2a12a9d7 | 907 | fsnotify_access(file); |
0eeca283 | 908 | else |
2a12a9d7 | 909 | fsnotify_modify(file); |
0eeca283 | 910 | } |
1da177e4 | 911 | return ret; |
1da177e4 LT |
912 | } |
913 | ||
914 | ssize_t vfs_readv(struct file *file, const struct iovec __user *vec, | |
793b80ef | 915 | unsigned long vlen, loff_t *pos, int flags) |
1da177e4 LT |
916 | { |
917 | if (!(file->f_mode & FMODE_READ)) | |
918 | return -EBADF; | |
7f7f25e8 | 919 | if (!(file->f_mode & FMODE_CAN_READ)) |
1da177e4 LT |
920 | return -EINVAL; |
921 | ||
793b80ef | 922 | return do_readv_writev(READ, file, vec, vlen, pos, flags); |
1da177e4 LT |
923 | } |
924 | ||
925 | EXPORT_SYMBOL(vfs_readv); | |
926 | ||
927 | ssize_t vfs_writev(struct file *file, const struct iovec __user *vec, | |
793b80ef | 928 | unsigned long vlen, loff_t *pos, int flags) |
1da177e4 LT |
929 | { |
930 | if (!(file->f_mode & FMODE_WRITE)) | |
931 | return -EBADF; | |
7f7f25e8 | 932 | if (!(file->f_mode & FMODE_CAN_WRITE)) |
1da177e4 LT |
933 | return -EINVAL; |
934 | ||
793b80ef | 935 | return do_readv_writev(WRITE, file, vec, vlen, pos, flags); |
1da177e4 LT |
936 | } |
937 | ||
938 | EXPORT_SYMBOL(vfs_writev); | |
939 | ||
f17d8b35 MT |
940 | static ssize_t do_readv(unsigned long fd, const struct iovec __user *vec, |
941 | unsigned long vlen, int flags) | |
1da177e4 | 942 | { |
9c225f26 | 943 | struct fd f = fdget_pos(fd); |
1da177e4 | 944 | ssize_t ret = -EBADF; |
1da177e4 | 945 | |
2903ff01 AV |
946 | if (f.file) { |
947 | loff_t pos = file_pos_read(f.file); | |
f17d8b35 | 948 | ret = vfs_readv(f.file, vec, vlen, &pos, flags); |
5faf153e AV |
949 | if (ret >= 0) |
950 | file_pos_write(f.file, pos); | |
9c225f26 | 951 | fdput_pos(f); |
1da177e4 LT |
952 | } |
953 | ||
954 | if (ret > 0) | |
4b98d11b AD |
955 | add_rchar(current, ret); |
956 | inc_syscr(current); | |
1da177e4 LT |
957 | return ret; |
958 | } | |
959 | ||
f17d8b35 MT |
960 | static ssize_t do_writev(unsigned long fd, const struct iovec __user *vec, |
961 | unsigned long vlen, int flags) | |
1da177e4 | 962 | { |
9c225f26 | 963 | struct fd f = fdget_pos(fd); |
1da177e4 | 964 | ssize_t ret = -EBADF; |
1da177e4 | 965 | |
2903ff01 AV |
966 | if (f.file) { |
967 | loff_t pos = file_pos_read(f.file); | |
f17d8b35 | 968 | ret = vfs_writev(f.file, vec, vlen, &pos, flags); |
5faf153e AV |
969 | if (ret >= 0) |
970 | file_pos_write(f.file, pos); | |
9c225f26 | 971 | fdput_pos(f); |
1da177e4 LT |
972 | } |
973 | ||
974 | if (ret > 0) | |
4b98d11b AD |
975 | add_wchar(current, ret); |
976 | inc_syscw(current); | |
1da177e4 LT |
977 | return ret; |
978 | } | |
979 | ||
601cc11d LT |
980 | static inline loff_t pos_from_hilo(unsigned long high, unsigned long low) |
981 | { | |
982 | #define HALF_LONG_BITS (BITS_PER_LONG / 2) | |
983 | return (((loff_t)high << HALF_LONG_BITS) << HALF_LONG_BITS) | low; | |
984 | } | |
985 | ||
f17d8b35 MT |
986 | static ssize_t do_preadv(unsigned long fd, const struct iovec __user *vec, |
987 | unsigned long vlen, loff_t pos, int flags) | |
f3554f4b | 988 | { |
2903ff01 | 989 | struct fd f; |
f3554f4b | 990 | ssize_t ret = -EBADF; |
f3554f4b GH |
991 | |
992 | if (pos < 0) | |
993 | return -EINVAL; | |
994 | ||
2903ff01 AV |
995 | f = fdget(fd); |
996 | if (f.file) { | |
f3554f4b | 997 | ret = -ESPIPE; |
2903ff01 | 998 | if (f.file->f_mode & FMODE_PREAD) |
f17d8b35 | 999 | ret = vfs_readv(f.file, vec, vlen, &pos, flags); |
2903ff01 | 1000 | fdput(f); |
f3554f4b GH |
1001 | } |
1002 | ||
1003 | if (ret > 0) | |
1004 | add_rchar(current, ret); | |
1005 | inc_syscr(current); | |
1006 | return ret; | |
1007 | } | |
1008 | ||
f17d8b35 MT |
1009 | static ssize_t do_pwritev(unsigned long fd, const struct iovec __user *vec, |
1010 | unsigned long vlen, loff_t pos, int flags) | |
f3554f4b | 1011 | { |
2903ff01 | 1012 | struct fd f; |
f3554f4b | 1013 | ssize_t ret = -EBADF; |
f3554f4b GH |
1014 | |
1015 | if (pos < 0) | |
1016 | return -EINVAL; | |
1017 | ||
2903ff01 AV |
1018 | f = fdget(fd); |
1019 | if (f.file) { | |
f3554f4b | 1020 | ret = -ESPIPE; |
2903ff01 | 1021 | if (f.file->f_mode & FMODE_PWRITE) |
f17d8b35 | 1022 | ret = vfs_writev(f.file, vec, vlen, &pos, flags); |
2903ff01 | 1023 | fdput(f); |
f3554f4b GH |
1024 | } |
1025 | ||
1026 | if (ret > 0) | |
1027 | add_wchar(current, ret); | |
1028 | inc_syscw(current); | |
1029 | return ret; | |
1030 | } | |
1031 | ||
f17d8b35 MT |
1032 | SYSCALL_DEFINE3(readv, unsigned long, fd, const struct iovec __user *, vec, |
1033 | unsigned long, vlen) | |
1034 | { | |
1035 | return do_readv(fd, vec, vlen, 0); | |
1036 | } | |
1037 | ||
1038 | SYSCALL_DEFINE3(writev, unsigned long, fd, const struct iovec __user *, vec, | |
1039 | unsigned long, vlen) | |
1040 | { | |
1041 | return do_writev(fd, vec, vlen, 0); | |
1042 | } | |
1043 | ||
1044 | SYSCALL_DEFINE5(preadv, unsigned long, fd, const struct iovec __user *, vec, | |
1045 | unsigned long, vlen, unsigned long, pos_l, unsigned long, pos_h) | |
1046 | { | |
1047 | loff_t pos = pos_from_hilo(pos_h, pos_l); | |
1048 | ||
1049 | return do_preadv(fd, vec, vlen, pos, 0); | |
1050 | } | |
1051 | ||
1052 | SYSCALL_DEFINE6(preadv2, unsigned long, fd, const struct iovec __user *, vec, | |
1053 | unsigned long, vlen, unsigned long, pos_l, unsigned long, pos_h, | |
1054 | int, flags) | |
1055 | { | |
1056 | loff_t pos = pos_from_hilo(pos_h, pos_l); | |
1057 | ||
1058 | if (pos == -1) | |
1059 | return do_readv(fd, vec, vlen, flags); | |
1060 | ||
1061 | return do_preadv(fd, vec, vlen, pos, flags); | |
1062 | } | |
1063 | ||
1064 | SYSCALL_DEFINE5(pwritev, unsigned long, fd, const struct iovec __user *, vec, | |
1065 | unsigned long, vlen, unsigned long, pos_l, unsigned long, pos_h) | |
1066 | { | |
1067 | loff_t pos = pos_from_hilo(pos_h, pos_l); | |
1068 | ||
1069 | return do_pwritev(fd, vec, vlen, pos, 0); | |
1070 | } | |
1071 | ||
1072 | SYSCALL_DEFINE6(pwritev2, unsigned long, fd, const struct iovec __user *, vec, | |
1073 | unsigned long, vlen, unsigned long, pos_l, unsigned long, pos_h, | |
1074 | int, flags) | |
1075 | { | |
1076 | loff_t pos = pos_from_hilo(pos_h, pos_l); | |
1077 | ||
1078 | if (pos == -1) | |
1079 | return do_writev(fd, vec, vlen, flags); | |
1080 | ||
1081 | return do_pwritev(fd, vec, vlen, pos, flags); | |
1082 | } | |
1083 | ||
72ec3516 AV |
1084 | #ifdef CONFIG_COMPAT |
1085 | ||
1086 | static ssize_t compat_do_readv_writev(int type, struct file *file, | |
1087 | const struct compat_iovec __user *uvector, | |
793b80ef CH |
1088 | unsigned long nr_segs, loff_t *pos, |
1089 | int flags) | |
72ec3516 AV |
1090 | { |
1091 | compat_ssize_t tot_len; | |
1092 | struct iovec iovstack[UIO_FASTIOV]; | |
1093 | struct iovec *iov = iovstack; | |
ac15ac06 | 1094 | struct iov_iter iter; |
72ec3516 AV |
1095 | ssize_t ret; |
1096 | io_fn_t fn; | |
293bc982 | 1097 | iter_fn_t iter_fn; |
72ec3516 | 1098 | |
0504c074 AV |
1099 | ret = compat_import_iovec(type, uvector, nr_segs, |
1100 | UIO_FASTIOV, &iov, &iter); | |
1101 | if (ret < 0) | |
1102 | return ret; | |
72ec3516 | 1103 | |
0504c074 AV |
1104 | tot_len = iov_iter_count(&iter); |
1105 | if (!tot_len) | |
1106 | goto out; | |
72ec3516 AV |
1107 | ret = rw_verify_area(type, file, pos, tot_len); |
1108 | if (ret < 0) | |
1109 | goto out; | |
1110 | ||
72ec3516 AV |
1111 | if (type == READ) { |
1112 | fn = file->f_op->read; | |
293bc982 | 1113 | iter_fn = file->f_op->read_iter; |
72ec3516 AV |
1114 | } else { |
1115 | fn = (io_fn_t)file->f_op->write; | |
293bc982 | 1116 | iter_fn = file->f_op->write_iter; |
03d95eb2 | 1117 | file_start_write(file); |
72ec3516 AV |
1118 | } |
1119 | ||
293bc982 | 1120 | if (iter_fn) |
793b80ef | 1121 | ret = do_iter_readv_writev(file, &iter, pos, iter_fn, flags); |
03d95eb2 | 1122 | else |
793b80ef | 1123 | ret = do_loop_readv_writev(file, &iter, pos, fn, flags); |
72ec3516 | 1124 | |
03d95eb2 AV |
1125 | if (type != READ) |
1126 | file_end_write(file); | |
1127 | ||
72ec3516 | 1128 | out: |
0504c074 | 1129 | kfree(iov); |
72ec3516 AV |
1130 | if ((ret + (type == READ)) > 0) { |
1131 | if (type == READ) | |
1132 | fsnotify_access(file); | |
1133 | else | |
1134 | fsnotify_modify(file); | |
1135 | } | |
1136 | return ret; | |
1137 | } | |
1138 | ||
1139 | static size_t compat_readv(struct file *file, | |
1140 | const struct compat_iovec __user *vec, | |
f17d8b35 | 1141 | unsigned long vlen, loff_t *pos, int flags) |
72ec3516 AV |
1142 | { |
1143 | ssize_t ret = -EBADF; | |
1144 | ||
1145 | if (!(file->f_mode & FMODE_READ)) | |
1146 | goto out; | |
1147 | ||
1148 | ret = -EINVAL; | |
7f7f25e8 | 1149 | if (!(file->f_mode & FMODE_CAN_READ)) |
72ec3516 AV |
1150 | goto out; |
1151 | ||
f17d8b35 | 1152 | ret = compat_do_readv_writev(READ, file, vec, vlen, pos, flags); |
72ec3516 AV |
1153 | |
1154 | out: | |
1155 | if (ret > 0) | |
1156 | add_rchar(current, ret); | |
1157 | inc_syscr(current); | |
1158 | return ret; | |
1159 | } | |
1160 | ||
f17d8b35 MT |
1161 | static size_t do_compat_readv(compat_ulong_t fd, |
1162 | const struct compat_iovec __user *vec, | |
1163 | compat_ulong_t vlen, int flags) | |
72ec3516 | 1164 | { |
9c225f26 | 1165 | struct fd f = fdget_pos(fd); |
72ec3516 AV |
1166 | ssize_t ret; |
1167 | loff_t pos; | |
1168 | ||
1169 | if (!f.file) | |
1170 | return -EBADF; | |
1171 | pos = f.file->f_pos; | |
f17d8b35 | 1172 | ret = compat_readv(f.file, vec, vlen, &pos, flags); |
5faf153e AV |
1173 | if (ret >= 0) |
1174 | f.file->f_pos = pos; | |
9c225f26 | 1175 | fdput_pos(f); |
72ec3516 | 1176 | return ret; |
f17d8b35 | 1177 | |
72ec3516 AV |
1178 | } |
1179 | ||
f17d8b35 MT |
1180 | COMPAT_SYSCALL_DEFINE3(readv, compat_ulong_t, fd, |
1181 | const struct compat_iovec __user *,vec, | |
1182 | compat_ulong_t, vlen) | |
1183 | { | |
1184 | return do_compat_readv(fd, vec, vlen, 0); | |
1185 | } | |
1186 | ||
1187 | static long do_compat_preadv64(unsigned long fd, | |
378a10f3 | 1188 | const struct compat_iovec __user *vec, |
f17d8b35 | 1189 | unsigned long vlen, loff_t pos, int flags) |
72ec3516 AV |
1190 | { |
1191 | struct fd f; | |
1192 | ssize_t ret; | |
1193 | ||
1194 | if (pos < 0) | |
1195 | return -EINVAL; | |
1196 | f = fdget(fd); | |
1197 | if (!f.file) | |
1198 | return -EBADF; | |
1199 | ret = -ESPIPE; | |
1200 | if (f.file->f_mode & FMODE_PREAD) | |
f17d8b35 | 1201 | ret = compat_readv(f.file, vec, vlen, &pos, flags); |
72ec3516 AV |
1202 | fdput(f); |
1203 | return ret; | |
1204 | } | |
1205 | ||
378a10f3 HC |
1206 | #ifdef __ARCH_WANT_COMPAT_SYS_PREADV64 |
1207 | COMPAT_SYSCALL_DEFINE4(preadv64, unsigned long, fd, | |
1208 | const struct compat_iovec __user *,vec, | |
1209 | unsigned long, vlen, loff_t, pos) | |
1210 | { | |
f17d8b35 | 1211 | return do_compat_preadv64(fd, vec, vlen, pos, 0); |
378a10f3 HC |
1212 | } |
1213 | #endif | |
1214 | ||
dfd948e3 | 1215 | COMPAT_SYSCALL_DEFINE5(preadv, compat_ulong_t, fd, |
72ec3516 | 1216 | const struct compat_iovec __user *,vec, |
dfd948e3 | 1217 | compat_ulong_t, vlen, u32, pos_low, u32, pos_high) |
72ec3516 AV |
1218 | { |
1219 | loff_t pos = ((loff_t)pos_high << 32) | pos_low; | |
378a10f3 | 1220 | |
f17d8b35 MT |
1221 | return do_compat_preadv64(fd, vec, vlen, pos, 0); |
1222 | } | |
1223 | ||
3ebfd81f L |
1224 | #ifdef __ARCH_WANT_COMPAT_SYS_PREADV64V2 |
1225 | COMPAT_SYSCALL_DEFINE5(preadv64v2, unsigned long, fd, | |
1226 | const struct compat_iovec __user *,vec, | |
1227 | unsigned long, vlen, loff_t, pos, int, flags) | |
1228 | { | |
1229 | return do_compat_preadv64(fd, vec, vlen, pos, flags); | |
1230 | } | |
1231 | #endif | |
1232 | ||
f17d8b35 MT |
1233 | COMPAT_SYSCALL_DEFINE6(preadv2, compat_ulong_t, fd, |
1234 | const struct compat_iovec __user *,vec, | |
1235 | compat_ulong_t, vlen, u32, pos_low, u32, pos_high, | |
1236 | int, flags) | |
1237 | { | |
1238 | loff_t pos = ((loff_t)pos_high << 32) | pos_low; | |
1239 | ||
1240 | if (pos == -1) | |
1241 | return do_compat_readv(fd, vec, vlen, flags); | |
1242 | ||
1243 | return do_compat_preadv64(fd, vec, vlen, pos, flags); | |
72ec3516 AV |
1244 | } |
1245 | ||
1246 | static size_t compat_writev(struct file *file, | |
1247 | const struct compat_iovec __user *vec, | |
f17d8b35 | 1248 | unsigned long vlen, loff_t *pos, int flags) |
72ec3516 AV |
1249 | { |
1250 | ssize_t ret = -EBADF; | |
1251 | ||
1252 | if (!(file->f_mode & FMODE_WRITE)) | |
1253 | goto out; | |
1254 | ||
1255 | ret = -EINVAL; | |
7f7f25e8 | 1256 | if (!(file->f_mode & FMODE_CAN_WRITE)) |
72ec3516 AV |
1257 | goto out; |
1258 | ||
793b80ef | 1259 | ret = compat_do_readv_writev(WRITE, file, vec, vlen, pos, 0); |
72ec3516 AV |
1260 | |
1261 | out: | |
1262 | if (ret > 0) | |
1263 | add_wchar(current, ret); | |
1264 | inc_syscw(current); | |
1265 | return ret; | |
1266 | } | |
1267 | ||
f17d8b35 MT |
1268 | static size_t do_compat_writev(compat_ulong_t fd, |
1269 | const struct compat_iovec __user* vec, | |
1270 | compat_ulong_t vlen, int flags) | |
72ec3516 | 1271 | { |
9c225f26 | 1272 | struct fd f = fdget_pos(fd); |
72ec3516 AV |
1273 | ssize_t ret; |
1274 | loff_t pos; | |
1275 | ||
1276 | if (!f.file) | |
1277 | return -EBADF; | |
1278 | pos = f.file->f_pos; | |
f17d8b35 | 1279 | ret = compat_writev(f.file, vec, vlen, &pos, flags); |
5faf153e AV |
1280 | if (ret >= 0) |
1281 | f.file->f_pos = pos; | |
9c225f26 | 1282 | fdput_pos(f); |
72ec3516 AV |
1283 | return ret; |
1284 | } | |
1285 | ||
f17d8b35 MT |
1286 | COMPAT_SYSCALL_DEFINE3(writev, compat_ulong_t, fd, |
1287 | const struct compat_iovec __user *, vec, | |
1288 | compat_ulong_t, vlen) | |
1289 | { | |
1290 | return do_compat_writev(fd, vec, vlen, 0); | |
1291 | } | |
1292 | ||
1293 | static long do_compat_pwritev64(unsigned long fd, | |
378a10f3 | 1294 | const struct compat_iovec __user *vec, |
f17d8b35 | 1295 | unsigned long vlen, loff_t pos, int flags) |
72ec3516 AV |
1296 | { |
1297 | struct fd f; | |
1298 | ssize_t ret; | |
1299 | ||
1300 | if (pos < 0) | |
1301 | return -EINVAL; | |
1302 | f = fdget(fd); | |
1303 | if (!f.file) | |
1304 | return -EBADF; | |
1305 | ret = -ESPIPE; | |
1306 | if (f.file->f_mode & FMODE_PWRITE) | |
f17d8b35 | 1307 | ret = compat_writev(f.file, vec, vlen, &pos, flags); |
72ec3516 AV |
1308 | fdput(f); |
1309 | return ret; | |
1310 | } | |
1311 | ||
378a10f3 HC |
1312 | #ifdef __ARCH_WANT_COMPAT_SYS_PWRITEV64 |
1313 | COMPAT_SYSCALL_DEFINE4(pwritev64, unsigned long, fd, | |
1314 | const struct compat_iovec __user *,vec, | |
1315 | unsigned long, vlen, loff_t, pos) | |
1316 | { | |
f17d8b35 | 1317 | return do_compat_pwritev64(fd, vec, vlen, pos, 0); |
378a10f3 HC |
1318 | } |
1319 | #endif | |
1320 | ||
dfd948e3 | 1321 | COMPAT_SYSCALL_DEFINE5(pwritev, compat_ulong_t, fd, |
72ec3516 | 1322 | const struct compat_iovec __user *,vec, |
dfd948e3 | 1323 | compat_ulong_t, vlen, u32, pos_low, u32, pos_high) |
72ec3516 AV |
1324 | { |
1325 | loff_t pos = ((loff_t)pos_high << 32) | pos_low; | |
378a10f3 | 1326 | |
f17d8b35 | 1327 | return do_compat_pwritev64(fd, vec, vlen, pos, 0); |
72ec3516 | 1328 | } |
f17d8b35 | 1329 | |
3ebfd81f L |
1330 | #ifdef __ARCH_WANT_COMPAT_SYS_PWRITEV64V2 |
1331 | COMPAT_SYSCALL_DEFINE5(pwritev64v2, unsigned long, fd, | |
1332 | const struct compat_iovec __user *,vec, | |
1333 | unsigned long, vlen, loff_t, pos, int, flags) | |
1334 | { | |
1335 | return do_compat_pwritev64(fd, vec, vlen, pos, flags); | |
1336 | } | |
1337 | #endif | |
1338 | ||
f17d8b35 MT |
1339 | COMPAT_SYSCALL_DEFINE6(pwritev2, compat_ulong_t, fd, |
1340 | const struct compat_iovec __user *,vec, | |
1341 | compat_ulong_t, vlen, u32, pos_low, u32, pos_high, int, flags) | |
1342 | { | |
1343 | loff_t pos = ((loff_t)pos_high << 32) | pos_low; | |
1344 | ||
1345 | if (pos == -1) | |
1346 | return do_compat_writev(fd, vec, vlen, flags); | |
1347 | ||
1348 | return do_compat_pwritev64(fd, vec, vlen, pos, flags); | |
72ec3516 | 1349 | } |
f17d8b35 | 1350 | |
72ec3516 AV |
1351 | #endif |
1352 | ||
19f4fc3a AV |
1353 | static ssize_t do_sendfile(int out_fd, int in_fd, loff_t *ppos, |
1354 | size_t count, loff_t max) | |
1da177e4 | 1355 | { |
2903ff01 AV |
1356 | struct fd in, out; |
1357 | struct inode *in_inode, *out_inode; | |
1da177e4 | 1358 | loff_t pos; |
7995bd28 | 1359 | loff_t out_pos; |
1da177e4 | 1360 | ssize_t retval; |
2903ff01 | 1361 | int fl; |
1da177e4 LT |
1362 | |
1363 | /* | |
1364 | * Get input file, and verify that it is ok.. | |
1365 | */ | |
1366 | retval = -EBADF; | |
2903ff01 AV |
1367 | in = fdget(in_fd); |
1368 | if (!in.file) | |
1da177e4 | 1369 | goto out; |
2903ff01 | 1370 | if (!(in.file->f_mode & FMODE_READ)) |
1da177e4 | 1371 | goto fput_in; |
1da177e4 | 1372 | retval = -ESPIPE; |
7995bd28 AV |
1373 | if (!ppos) { |
1374 | pos = in.file->f_pos; | |
1375 | } else { | |
1376 | pos = *ppos; | |
2903ff01 | 1377 | if (!(in.file->f_mode & FMODE_PREAD)) |
1da177e4 | 1378 | goto fput_in; |
7995bd28 AV |
1379 | } |
1380 | retval = rw_verify_area(READ, in.file, &pos, count); | |
e28cc715 | 1381 | if (retval < 0) |
1da177e4 | 1382 | goto fput_in; |
bc61384d AV |
1383 | if (count > MAX_RW_COUNT) |
1384 | count = MAX_RW_COUNT; | |
1da177e4 | 1385 | |
1da177e4 LT |
1386 | /* |
1387 | * Get output file, and verify that it is ok.. | |
1388 | */ | |
1389 | retval = -EBADF; | |
2903ff01 AV |
1390 | out = fdget(out_fd); |
1391 | if (!out.file) | |
1da177e4 | 1392 | goto fput_in; |
2903ff01 | 1393 | if (!(out.file->f_mode & FMODE_WRITE)) |
1da177e4 LT |
1394 | goto fput_out; |
1395 | retval = -EINVAL; | |
496ad9aa AV |
1396 | in_inode = file_inode(in.file); |
1397 | out_inode = file_inode(out.file); | |
7995bd28 AV |
1398 | out_pos = out.file->f_pos; |
1399 | retval = rw_verify_area(WRITE, out.file, &out_pos, count); | |
e28cc715 | 1400 | if (retval < 0) |
1da177e4 LT |
1401 | goto fput_out; |
1402 | ||
1da177e4 LT |
1403 | if (!max) |
1404 | max = min(in_inode->i_sb->s_maxbytes, out_inode->i_sb->s_maxbytes); | |
1405 | ||
1da177e4 LT |
1406 | if (unlikely(pos + count > max)) { |
1407 | retval = -EOVERFLOW; | |
1408 | if (pos >= max) | |
1409 | goto fput_out; | |
1410 | count = max - pos; | |
1411 | } | |
1412 | ||
d96e6e71 | 1413 | fl = 0; |
534f2aaa | 1414 | #if 0 |
d96e6e71 JA |
1415 | /* |
1416 | * We need to debate whether we can enable this or not. The | |
1417 | * man page documents EAGAIN return for the output at least, | |
1418 | * and the application is arguably buggy if it doesn't expect | |
1419 | * EAGAIN on a non-blocking file descriptor. | |
1420 | */ | |
2903ff01 | 1421 | if (in.file->f_flags & O_NONBLOCK) |
d96e6e71 | 1422 | fl = SPLICE_F_NONBLOCK; |
534f2aaa | 1423 | #endif |
50cd2c57 | 1424 | file_start_write(out.file); |
7995bd28 | 1425 | retval = do_splice_direct(in.file, &pos, out.file, &out_pos, count, fl); |
50cd2c57 | 1426 | file_end_write(out.file); |
1da177e4 LT |
1427 | |
1428 | if (retval > 0) { | |
4b98d11b AD |
1429 | add_rchar(current, retval); |
1430 | add_wchar(current, retval); | |
a68c2f12 SW |
1431 | fsnotify_access(in.file); |
1432 | fsnotify_modify(out.file); | |
7995bd28 AV |
1433 | out.file->f_pos = out_pos; |
1434 | if (ppos) | |
1435 | *ppos = pos; | |
1436 | else | |
1437 | in.file->f_pos = pos; | |
1da177e4 | 1438 | } |
1da177e4 | 1439 | |
4b98d11b AD |
1440 | inc_syscr(current); |
1441 | inc_syscw(current); | |
7995bd28 | 1442 | if (pos > max) |
1da177e4 LT |
1443 | retval = -EOVERFLOW; |
1444 | ||
1445 | fput_out: | |
2903ff01 | 1446 | fdput(out); |
1da177e4 | 1447 | fput_in: |
2903ff01 | 1448 | fdput(in); |
1da177e4 LT |
1449 | out: |
1450 | return retval; | |
1451 | } | |
1452 | ||
002c8976 | 1453 | SYSCALL_DEFINE4(sendfile, int, out_fd, int, in_fd, off_t __user *, offset, size_t, count) |
1da177e4 LT |
1454 | { |
1455 | loff_t pos; | |
1456 | off_t off; | |
1457 | ssize_t ret; | |
1458 | ||
1459 | if (offset) { | |
1460 | if (unlikely(get_user(off, offset))) | |
1461 | return -EFAULT; | |
1462 | pos = off; | |
1463 | ret = do_sendfile(out_fd, in_fd, &pos, count, MAX_NON_LFS); | |
1464 | if (unlikely(put_user(pos, offset))) | |
1465 | return -EFAULT; | |
1466 | return ret; | |
1467 | } | |
1468 | ||
1469 | return do_sendfile(out_fd, in_fd, NULL, count, 0); | |
1470 | } | |
1471 | ||
002c8976 | 1472 | SYSCALL_DEFINE4(sendfile64, int, out_fd, int, in_fd, loff_t __user *, offset, size_t, count) |
1da177e4 LT |
1473 | { |
1474 | loff_t pos; | |
1475 | ssize_t ret; | |
1476 | ||
1477 | if (offset) { | |
1478 | if (unlikely(copy_from_user(&pos, offset, sizeof(loff_t)))) | |
1479 | return -EFAULT; | |
1480 | ret = do_sendfile(out_fd, in_fd, &pos, count, 0); | |
1481 | if (unlikely(put_user(pos, offset))) | |
1482 | return -EFAULT; | |
1483 | return ret; | |
1484 | } | |
1485 | ||
1486 | return do_sendfile(out_fd, in_fd, NULL, count, 0); | |
1487 | } | |
19f4fc3a AV |
1488 | |
1489 | #ifdef CONFIG_COMPAT | |
1490 | COMPAT_SYSCALL_DEFINE4(sendfile, int, out_fd, int, in_fd, | |
1491 | compat_off_t __user *, offset, compat_size_t, count) | |
1492 | { | |
1493 | loff_t pos; | |
1494 | off_t off; | |
1495 | ssize_t ret; | |
1496 | ||
1497 | if (offset) { | |
1498 | if (unlikely(get_user(off, offset))) | |
1499 | return -EFAULT; | |
1500 | pos = off; | |
1501 | ret = do_sendfile(out_fd, in_fd, &pos, count, MAX_NON_LFS); | |
1502 | if (unlikely(put_user(pos, offset))) | |
1503 | return -EFAULT; | |
1504 | return ret; | |
1505 | } | |
1506 | ||
1507 | return do_sendfile(out_fd, in_fd, NULL, count, 0); | |
1508 | } | |
1509 | ||
1510 | COMPAT_SYSCALL_DEFINE4(sendfile64, int, out_fd, int, in_fd, | |
1511 | compat_loff_t __user *, offset, compat_size_t, count) | |
1512 | { | |
1513 | loff_t pos; | |
1514 | ssize_t ret; | |
1515 | ||
1516 | if (offset) { | |
1517 | if (unlikely(copy_from_user(&pos, offset, sizeof(loff_t)))) | |
1518 | return -EFAULT; | |
1519 | ret = do_sendfile(out_fd, in_fd, &pos, count, 0); | |
1520 | if (unlikely(put_user(pos, offset))) | |
1521 | return -EFAULT; | |
1522 | return ret; | |
1523 | } | |
1524 | ||
1525 | return do_sendfile(out_fd, in_fd, NULL, count, 0); | |
1526 | } | |
1527 | #endif | |
29732938 ZB |
1528 | |
1529 | /* | |
1530 | * copy_file_range() differs from regular file read and write in that it | |
1531 | * specifically allows return partial success. When it does so is up to | |
1532 | * the copy_file_range method. | |
1533 | */ | |
1534 | ssize_t vfs_copy_file_range(struct file *file_in, loff_t pos_in, | |
1535 | struct file *file_out, loff_t pos_out, | |
1536 | size_t len, unsigned int flags) | |
1537 | { | |
1538 | struct inode *inode_in = file_inode(file_in); | |
1539 | struct inode *inode_out = file_inode(file_out); | |
1540 | ssize_t ret; | |
1541 | ||
1542 | if (flags != 0) | |
1543 | return -EINVAL; | |
1544 | ||
29732938 | 1545 | ret = rw_verify_area(READ, file_in, &pos_in, len); |
bc61384d AV |
1546 | if (unlikely(ret)) |
1547 | return ret; | |
1548 | ||
1549 | ret = rw_verify_area(WRITE, file_out, &pos_out, len); | |
1550 | if (unlikely(ret)) | |
29732938 ZB |
1551 | return ret; |
1552 | ||
1553 | if (!(file_in->f_mode & FMODE_READ) || | |
1554 | !(file_out->f_mode & FMODE_WRITE) || | |
eac70053 | 1555 | (file_out->f_flags & O_APPEND)) |
29732938 ZB |
1556 | return -EBADF; |
1557 | ||
1558 | /* this could be relaxed once a method supports cross-fs copies */ | |
1559 | if (inode_in->i_sb != inode_out->i_sb) | |
1560 | return -EXDEV; | |
1561 | ||
1562 | if (len == 0) | |
1563 | return 0; | |
1564 | ||
3616119d | 1565 | sb_start_write(inode_out->i_sb); |
29732938 | 1566 | |
a76b5b04 CH |
1567 | /* |
1568 | * Try cloning first, this is supported by more file systems, and | |
1569 | * more efficient if both clone and copy are supported (e.g. NFS). | |
1570 | */ | |
1571 | if (file_in->f_op->clone_file_range) { | |
1572 | ret = file_in->f_op->clone_file_range(file_in, pos_in, | |
1573 | file_out, pos_out, len); | |
1574 | if (ret == 0) { | |
1575 | ret = len; | |
1576 | goto done; | |
1577 | } | |
1578 | } | |
1579 | ||
1580 | if (file_out->f_op->copy_file_range) { | |
eac70053 AS |
1581 | ret = file_out->f_op->copy_file_range(file_in, pos_in, file_out, |
1582 | pos_out, len, flags); | |
a76b5b04 CH |
1583 | if (ret != -EOPNOTSUPP) |
1584 | goto done; | |
1585 | } | |
eac70053 | 1586 | |
a76b5b04 CH |
1587 | ret = do_splice_direct(file_in, &pos_in, file_out, &pos_out, |
1588 | len > MAX_RW_COUNT ? MAX_RW_COUNT : len, 0); | |
eac70053 | 1589 | |
a76b5b04 | 1590 | done: |
29732938 ZB |
1591 | if (ret > 0) { |
1592 | fsnotify_access(file_in); | |
1593 | add_rchar(current, ret); | |
1594 | fsnotify_modify(file_out); | |
1595 | add_wchar(current, ret); | |
1596 | } | |
a76b5b04 | 1597 | |
29732938 ZB |
1598 | inc_syscr(current); |
1599 | inc_syscw(current); | |
1600 | ||
3616119d | 1601 | sb_end_write(inode_out->i_sb); |
29732938 ZB |
1602 | |
1603 | return ret; | |
1604 | } | |
1605 | EXPORT_SYMBOL(vfs_copy_file_range); | |
1606 | ||
1607 | SYSCALL_DEFINE6(copy_file_range, int, fd_in, loff_t __user *, off_in, | |
1608 | int, fd_out, loff_t __user *, off_out, | |
1609 | size_t, len, unsigned int, flags) | |
1610 | { | |
1611 | loff_t pos_in; | |
1612 | loff_t pos_out; | |
1613 | struct fd f_in; | |
1614 | struct fd f_out; | |
1615 | ssize_t ret = -EBADF; | |
1616 | ||
1617 | f_in = fdget(fd_in); | |
1618 | if (!f_in.file) | |
1619 | goto out2; | |
1620 | ||
1621 | f_out = fdget(fd_out); | |
1622 | if (!f_out.file) | |
1623 | goto out1; | |
1624 | ||
1625 | ret = -EFAULT; | |
1626 | if (off_in) { | |
1627 | if (copy_from_user(&pos_in, off_in, sizeof(loff_t))) | |
1628 | goto out; | |
1629 | } else { | |
1630 | pos_in = f_in.file->f_pos; | |
1631 | } | |
1632 | ||
1633 | if (off_out) { | |
1634 | if (copy_from_user(&pos_out, off_out, sizeof(loff_t))) | |
1635 | goto out; | |
1636 | } else { | |
1637 | pos_out = f_out.file->f_pos; | |
1638 | } | |
1639 | ||
1640 | ret = vfs_copy_file_range(f_in.file, pos_in, f_out.file, pos_out, len, | |
1641 | flags); | |
1642 | if (ret > 0) { | |
1643 | pos_in += ret; | |
1644 | pos_out += ret; | |
1645 | ||
1646 | if (off_in) { | |
1647 | if (copy_to_user(off_in, &pos_in, sizeof(loff_t))) | |
1648 | ret = -EFAULT; | |
1649 | } else { | |
1650 | f_in.file->f_pos = pos_in; | |
1651 | } | |
1652 | ||
1653 | if (off_out) { | |
1654 | if (copy_to_user(off_out, &pos_out, sizeof(loff_t))) | |
1655 | ret = -EFAULT; | |
1656 | } else { | |
1657 | f_out.file->f_pos = pos_out; | |
1658 | } | |
1659 | } | |
1660 | ||
1661 | out: | |
1662 | fdput(f_out); | |
1663 | out1: | |
1664 | fdput(f_in); | |
1665 | out2: | |
1666 | return ret; | |
1667 | } | |
04b38d60 CH |
1668 | |
1669 | static int clone_verify_area(struct file *file, loff_t pos, u64 len, bool write) | |
1670 | { | |
1671 | struct inode *inode = file_inode(file); | |
1672 | ||
1673 | if (unlikely(pos < 0)) | |
1674 | return -EINVAL; | |
1675 | ||
1676 | if (unlikely((loff_t) (pos + len) < 0)) | |
1677 | return -EINVAL; | |
1678 | ||
1679 | if (unlikely(inode->i_flctx && mandatory_lock(inode))) { | |
1680 | loff_t end = len ? pos + len - 1 : OFFSET_MAX; | |
1681 | int retval; | |
1682 | ||
1683 | retval = locks_mandatory_area(inode, file, pos, end, | |
1684 | write ? F_WRLCK : F_RDLCK); | |
1685 | if (retval < 0) | |
1686 | return retval; | |
1687 | } | |
1688 | ||
1689 | return security_file_permission(file, write ? MAY_WRITE : MAY_READ); | |
1690 | } | |
1691 | ||
876bec6f DW |
1692 | /* |
1693 | * Check that the two inodes are eligible for cloning, the ranges make | |
1694 | * sense, and then flush all dirty data. Caller must ensure that the | |
1695 | * inodes have been locked against any other modifications. | |
22725ce4 DW |
1696 | * |
1697 | * Returns: 0 for "nothing to clone", 1 for "something to clone", or | |
1698 | * the usual negative error code. | |
876bec6f DW |
1699 | */ |
1700 | int vfs_clone_file_prep_inodes(struct inode *inode_in, loff_t pos_in, | |
1701 | struct inode *inode_out, loff_t pos_out, | |
1702 | u64 *len, bool is_dedupe) | |
1703 | { | |
1704 | loff_t bs = inode_out->i_sb->s_blocksize; | |
1705 | loff_t blen; | |
1706 | loff_t isize; | |
1707 | bool same_inode = (inode_in == inode_out); | |
1708 | int ret; | |
1709 | ||
1710 | /* Don't touch certain kinds of inodes */ | |
1711 | if (IS_IMMUTABLE(inode_out)) | |
1712 | return -EPERM; | |
1713 | ||
1714 | if (IS_SWAPFILE(inode_in) || IS_SWAPFILE(inode_out)) | |
1715 | return -ETXTBSY; | |
1716 | ||
1717 | /* Don't reflink dirs, pipes, sockets... */ | |
1718 | if (S_ISDIR(inode_in->i_mode) || S_ISDIR(inode_out->i_mode)) | |
1719 | return -EISDIR; | |
1720 | if (!S_ISREG(inode_in->i_mode) || !S_ISREG(inode_out->i_mode)) | |
1721 | return -EINVAL; | |
1722 | ||
1723 | /* Are we going all the way to the end? */ | |
1724 | isize = i_size_read(inode_in); | |
22725ce4 | 1725 | if (isize == 0) |
876bec6f | 1726 | return 0; |
876bec6f DW |
1727 | |
1728 | /* Zero length dedupe exits immediately; reflink goes to EOF. */ | |
1729 | if (*len == 0) { | |
22725ce4 | 1730 | if (is_dedupe || pos_in == isize) |
876bec6f | 1731 | return 0; |
22725ce4 DW |
1732 | if (pos_in > isize) |
1733 | return -EINVAL; | |
876bec6f DW |
1734 | *len = isize - pos_in; |
1735 | } | |
1736 | ||
1737 | /* Ensure offsets don't wrap and the input is inside i_size */ | |
1738 | if (pos_in + *len < pos_in || pos_out + *len < pos_out || | |
1739 | pos_in + *len > isize) | |
1740 | return -EINVAL; | |
1741 | ||
1742 | /* Don't allow dedupe past EOF in the dest file */ | |
1743 | if (is_dedupe) { | |
1744 | loff_t disize; | |
1745 | ||
1746 | disize = i_size_read(inode_out); | |
1747 | if (pos_out >= disize || pos_out + *len > disize) | |
1748 | return -EINVAL; | |
1749 | } | |
1750 | ||
1751 | /* If we're linking to EOF, continue to the block boundary. */ | |
1752 | if (pos_in + *len == isize) | |
1753 | blen = ALIGN(isize, bs) - pos_in; | |
1754 | else | |
1755 | blen = *len; | |
1756 | ||
1757 | /* Only reflink if we're aligned to block boundaries */ | |
1758 | if (!IS_ALIGNED(pos_in, bs) || !IS_ALIGNED(pos_in + blen, bs) || | |
1759 | !IS_ALIGNED(pos_out, bs) || !IS_ALIGNED(pos_out + blen, bs)) | |
1760 | return -EINVAL; | |
1761 | ||
1762 | /* Don't allow overlapped reflink within the same file */ | |
1763 | if (same_inode) { | |
1764 | if (pos_out + blen > pos_in && pos_out < pos_in + blen) | |
1765 | return -EINVAL; | |
1766 | } | |
1767 | ||
1768 | /* Wait for the completion of any pending IOs on both files */ | |
1769 | inode_dio_wait(inode_in); | |
1770 | if (!same_inode) | |
1771 | inode_dio_wait(inode_out); | |
1772 | ||
1773 | ret = filemap_write_and_wait_range(inode_in->i_mapping, | |
1774 | pos_in, pos_in + *len - 1); | |
1775 | if (ret) | |
1776 | return ret; | |
1777 | ||
1778 | ret = filemap_write_and_wait_range(inode_out->i_mapping, | |
1779 | pos_out, pos_out + *len - 1); | |
1780 | if (ret) | |
1781 | return ret; | |
1782 | ||
1783 | /* | |
1784 | * Check that the extents are the same. | |
1785 | */ | |
1786 | if (is_dedupe) { | |
1787 | bool is_same = false; | |
1788 | ||
1789 | ret = vfs_dedupe_file_range_compare(inode_in, pos_in, | |
1790 | inode_out, pos_out, *len, &is_same); | |
1791 | if (ret) | |
1792 | return ret; | |
1793 | if (!is_same) | |
1794 | return -EBADE; | |
1795 | } | |
1796 | ||
22725ce4 | 1797 | return 1; |
876bec6f DW |
1798 | } |
1799 | EXPORT_SYMBOL(vfs_clone_file_prep_inodes); | |
1800 | ||
04b38d60 CH |
1801 | int vfs_clone_file_range(struct file *file_in, loff_t pos_in, |
1802 | struct file *file_out, loff_t pos_out, u64 len) | |
1803 | { | |
1804 | struct inode *inode_in = file_inode(file_in); | |
1805 | struct inode *inode_out = file_inode(file_out); | |
1806 | int ret; | |
1807 | ||
b335e9d9 AG |
1808 | if (S_ISDIR(inode_in->i_mode) || S_ISDIR(inode_out->i_mode)) |
1809 | return -EISDIR; | |
1810 | if (!S_ISREG(inode_in->i_mode) || !S_ISREG(inode_out->i_mode)) | |
1811 | return -EINVAL; | |
1812 | ||
913b86e9 AG |
1813 | /* |
1814 | * FICLONE/FICLONERANGE ioctls enforce that src and dest files are on | |
1815 | * the same mount. Practically, they only need to be on the same file | |
1816 | * system. | |
1817 | */ | |
1818 | if (inode_in->i_sb != inode_out->i_sb) | |
04b38d60 CH |
1819 | return -EXDEV; |
1820 | ||
04b38d60 CH |
1821 | if (!(file_in->f_mode & FMODE_READ) || |
1822 | !(file_out->f_mode & FMODE_WRITE) || | |
0fcbf996 | 1823 | (file_out->f_flags & O_APPEND)) |
04b38d60 CH |
1824 | return -EBADF; |
1825 | ||
0fcbf996 CH |
1826 | if (!file_in->f_op->clone_file_range) |
1827 | return -EOPNOTSUPP; | |
1828 | ||
04b38d60 CH |
1829 | ret = clone_verify_area(file_in, pos_in, len, false); |
1830 | if (ret) | |
1831 | return ret; | |
1832 | ||
1833 | ret = clone_verify_area(file_out, pos_out, len, true); | |
1834 | if (ret) | |
1835 | return ret; | |
1836 | ||
1837 | if (pos_in + len > i_size_read(inode_in)) | |
1838 | return -EINVAL; | |
1839 | ||
04b38d60 CH |
1840 | ret = file_in->f_op->clone_file_range(file_in, pos_in, |
1841 | file_out, pos_out, len); | |
1842 | if (!ret) { | |
1843 | fsnotify_access(file_in); | |
1844 | fsnotify_modify(file_out); | |
1845 | } | |
1846 | ||
04b38d60 CH |
1847 | return ret; |
1848 | } | |
1849 | EXPORT_SYMBOL(vfs_clone_file_range); | |
54dbc151 | 1850 | |
876bec6f DW |
1851 | /* |
1852 | * Read a page's worth of file data into the page cache. Return the page | |
1853 | * locked. | |
1854 | */ | |
1855 | static struct page *vfs_dedupe_get_page(struct inode *inode, loff_t offset) | |
1856 | { | |
1857 | struct address_space *mapping; | |
1858 | struct page *page; | |
1859 | pgoff_t n; | |
1860 | ||
1861 | n = offset >> PAGE_SHIFT; | |
1862 | mapping = inode->i_mapping; | |
1863 | page = read_mapping_page(mapping, n, NULL); | |
1864 | if (IS_ERR(page)) | |
1865 | return page; | |
1866 | if (!PageUptodate(page)) { | |
1867 | put_page(page); | |
1868 | return ERR_PTR(-EIO); | |
1869 | } | |
1870 | lock_page(page); | |
1871 | return page; | |
1872 | } | |
1873 | ||
1874 | /* | |
1875 | * Compare extents of two files to see if they are the same. | |
1876 | * Caller must have locked both inodes to prevent write races. | |
1877 | */ | |
1878 | int vfs_dedupe_file_range_compare(struct inode *src, loff_t srcoff, | |
1879 | struct inode *dest, loff_t destoff, | |
1880 | loff_t len, bool *is_same) | |
1881 | { | |
1882 | loff_t src_poff; | |
1883 | loff_t dest_poff; | |
1884 | void *src_addr; | |
1885 | void *dest_addr; | |
1886 | struct page *src_page; | |
1887 | struct page *dest_page; | |
1888 | loff_t cmp_len; | |
1889 | bool same; | |
1890 | int error; | |
1891 | ||
1892 | error = -EINVAL; | |
1893 | same = true; | |
1894 | while (len) { | |
1895 | src_poff = srcoff & (PAGE_SIZE - 1); | |
1896 | dest_poff = destoff & (PAGE_SIZE - 1); | |
1897 | cmp_len = min(PAGE_SIZE - src_poff, | |
1898 | PAGE_SIZE - dest_poff); | |
1899 | cmp_len = min(cmp_len, len); | |
1900 | if (cmp_len <= 0) | |
1901 | goto out_error; | |
1902 | ||
1903 | src_page = vfs_dedupe_get_page(src, srcoff); | |
1904 | if (IS_ERR(src_page)) { | |
1905 | error = PTR_ERR(src_page); | |
1906 | goto out_error; | |
1907 | } | |
1908 | dest_page = vfs_dedupe_get_page(dest, destoff); | |
1909 | if (IS_ERR(dest_page)) { | |
1910 | error = PTR_ERR(dest_page); | |
1911 | unlock_page(src_page); | |
1912 | put_page(src_page); | |
1913 | goto out_error; | |
1914 | } | |
1915 | src_addr = kmap_atomic(src_page); | |
1916 | dest_addr = kmap_atomic(dest_page); | |
1917 | ||
1918 | flush_dcache_page(src_page); | |
1919 | flush_dcache_page(dest_page); | |
1920 | ||
1921 | if (memcmp(src_addr + src_poff, dest_addr + dest_poff, cmp_len)) | |
1922 | same = false; | |
1923 | ||
1924 | kunmap_atomic(dest_addr); | |
1925 | kunmap_atomic(src_addr); | |
1926 | unlock_page(dest_page); | |
1927 | unlock_page(src_page); | |
1928 | put_page(dest_page); | |
1929 | put_page(src_page); | |
1930 | ||
1931 | if (!same) | |
1932 | break; | |
1933 | ||
1934 | srcoff += cmp_len; | |
1935 | destoff += cmp_len; | |
1936 | len -= cmp_len; | |
1937 | } | |
1938 | ||
1939 | *is_same = same; | |
1940 | return 0; | |
1941 | ||
1942 | out_error: | |
1943 | return error; | |
1944 | } | |
1945 | EXPORT_SYMBOL(vfs_dedupe_file_range_compare); | |
1946 | ||
54dbc151 DW |
1947 | int vfs_dedupe_file_range(struct file *file, struct file_dedupe_range *same) |
1948 | { | |
1949 | struct file_dedupe_range_info *info; | |
1950 | struct inode *src = file_inode(file); | |
1951 | u64 off; | |
1952 | u64 len; | |
1953 | int i; | |
1954 | int ret; | |
1955 | bool is_admin = capable(CAP_SYS_ADMIN); | |
1956 | u16 count = same->dest_count; | |
1957 | struct file *dst_file; | |
1958 | loff_t dst_off; | |
1959 | ssize_t deduped; | |
1960 | ||
1961 | if (!(file->f_mode & FMODE_READ)) | |
1962 | return -EINVAL; | |
1963 | ||
1964 | if (same->reserved1 || same->reserved2) | |
1965 | return -EINVAL; | |
1966 | ||
1967 | off = same->src_offset; | |
1968 | len = same->src_length; | |
1969 | ||
1970 | ret = -EISDIR; | |
1971 | if (S_ISDIR(src->i_mode)) | |
1972 | goto out; | |
1973 | ||
1974 | ret = -EINVAL; | |
1975 | if (!S_ISREG(src->i_mode)) | |
1976 | goto out; | |
1977 | ||
1978 | ret = clone_verify_area(file, off, len, false); | |
1979 | if (ret < 0) | |
1980 | goto out; | |
1981 | ret = 0; | |
1982 | ||
22725ce4 DW |
1983 | if (off + len > i_size_read(src)) |
1984 | return -EINVAL; | |
1985 | ||
54dbc151 DW |
1986 | /* pre-format output fields to sane values */ |
1987 | for (i = 0; i < count; i++) { | |
1988 | same->info[i].bytes_deduped = 0ULL; | |
1989 | same->info[i].status = FILE_DEDUPE_RANGE_SAME; | |
1990 | } | |
1991 | ||
1992 | for (i = 0, info = same->info; i < count; i++, info++) { | |
1993 | struct inode *dst; | |
1994 | struct fd dst_fd = fdget(info->dest_fd); | |
1995 | ||
1996 | dst_file = dst_fd.file; | |
1997 | if (!dst_file) { | |
1998 | info->status = -EBADF; | |
1999 | goto next_loop; | |
2000 | } | |
2001 | dst = file_inode(dst_file); | |
2002 | ||
2003 | ret = mnt_want_write_file(dst_file); | |
2004 | if (ret) { | |
2005 | info->status = ret; | |
2006 | goto next_loop; | |
2007 | } | |
2008 | ||
2009 | dst_off = info->dest_offset; | |
2010 | ret = clone_verify_area(dst_file, dst_off, len, true); | |
2011 | if (ret < 0) { | |
2012 | info->status = ret; | |
2013 | goto next_file; | |
2014 | } | |
2015 | ret = 0; | |
2016 | ||
2017 | if (info->reserved) { | |
2018 | info->status = -EINVAL; | |
2019 | } else if (!(is_admin || (dst_file->f_mode & FMODE_WRITE))) { | |
2020 | info->status = -EINVAL; | |
2021 | } else if (file->f_path.mnt != dst_file->f_path.mnt) { | |
2022 | info->status = -EXDEV; | |
2023 | } else if (S_ISDIR(dst->i_mode)) { | |
2024 | info->status = -EISDIR; | |
2025 | } else if (dst_file->f_op->dedupe_file_range == NULL) { | |
2026 | info->status = -EINVAL; | |
2027 | } else { | |
2028 | deduped = dst_file->f_op->dedupe_file_range(file, off, | |
2029 | len, dst_file, | |
2030 | info->dest_offset); | |
2031 | if (deduped == -EBADE) | |
2032 | info->status = FILE_DEDUPE_RANGE_DIFFERS; | |
2033 | else if (deduped < 0) | |
2034 | info->status = deduped; | |
2035 | else | |
2036 | info->bytes_deduped += deduped; | |
2037 | } | |
2038 | ||
2039 | next_file: | |
2040 | mnt_drop_write_file(dst_file); | |
2041 | next_loop: | |
2042 | fdput(dst_fd); | |
e62e560f DW |
2043 | |
2044 | if (fatal_signal_pending(current)) | |
2045 | goto out; | |
54dbc151 DW |
2046 | } |
2047 | ||
2048 | out: | |
2049 | return ret; | |
2050 | } | |
2051 | EXPORT_SYMBOL(vfs_dedupe_file_range); |