]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/fs/read_write.c | |
3 | * | |
4 | * Copyright (C) 1991, 1992 Linus Torvalds | |
5 | */ | |
6 | ||
b12fb7f4 | 7 | #include <linux/slab.h> |
1da177e4 | 8 | #include <linux/stat.h> |
b12fb7f4 | 9 | #include <linux/sched/xacct.h> |
1da177e4 LT |
10 | #include <linux/fcntl.h> |
11 | #include <linux/file.h> | |
12 | #include <linux/uio.h> | |
0eeca283 | 13 | #include <linux/fsnotify.h> |
1da177e4 | 14 | #include <linux/security.h> |
630d9c47 | 15 | #include <linux/export.h> |
1da177e4 | 16 | #include <linux/syscalls.h> |
e28cc715 | 17 | #include <linux/pagemap.h> |
d6b29d7c | 18 | #include <linux/splice.h> |
561c6731 | 19 | #include <linux/compat.h> |
29732938 | 20 | #include <linux/mount.h> |
2feb55f8 | 21 | #include <linux/fs.h> |
06ae43f3 | 22 | #include "internal.h" |
1da177e4 | 23 | |
7c0f6ba6 | 24 | #include <linux/uaccess.h> |
1da177e4 LT |
25 | #include <asm/unistd.h> |
26 | ||
4b6f5d20 | 27 | const struct file_operations generic_ro_fops = { |
1da177e4 | 28 | .llseek = generic_file_llseek, |
aad4f8bb | 29 | .read_iter = generic_file_read_iter, |
1da177e4 | 30 | .mmap = generic_file_readonly_mmap, |
534f2aaa | 31 | .splice_read = generic_file_splice_read, |
1da177e4 LT |
32 | }; |
33 | ||
34 | EXPORT_SYMBOL(generic_ro_fops); | |
35 | ||
cccb5a1e | 36 | static inline int unsigned_offsets(struct file *file) |
4a3956c7 | 37 | { |
cccb5a1e | 38 | return file->f_mode & FMODE_UNSIGNED_OFFSET; |
4a3956c7 KH |
39 | } |
40 | ||
46a1c2c7 JL |
41 | /** |
42 | * vfs_setpos - update the file offset for lseek | |
43 | * @file: file structure in question | |
44 | * @offset: file offset to seek to | |
45 | * @maxsize: maximum file size | |
46 | * | |
47 | * This is a low-level filesystem helper for updating the file offset to | |
48 | * the value specified by @offset if the given offset is valid and it is | |
49 | * not equal to the current file offset. | |
50 | * | |
51 | * Return the specified offset on success and -EINVAL on invalid offset. | |
52 | */ | |
53 | loff_t vfs_setpos(struct file *file, loff_t offset, loff_t maxsize) | |
ef3d0fd2 AK |
54 | { |
55 | if (offset < 0 && !unsigned_offsets(file)) | |
56 | return -EINVAL; | |
57 | if (offset > maxsize) | |
58 | return -EINVAL; | |
59 | ||
60 | if (offset != file->f_pos) { | |
61 | file->f_pos = offset; | |
62 | file->f_version = 0; | |
63 | } | |
64 | return offset; | |
65 | } | |
46a1c2c7 | 66 | EXPORT_SYMBOL(vfs_setpos); |
ef3d0fd2 | 67 | |
3a8cff4f | 68 | /** |
5760495a | 69 | * generic_file_llseek_size - generic llseek implementation for regular files |
3a8cff4f CH |
70 | * @file: file structure to seek on |
71 | * @offset: file offset to seek to | |
965c8e59 | 72 | * @whence: type of seek |
e8b96eb5 ES |
73 | * @size: max size of this file in file system |
74 | * @eof: offset used for SEEK_END position | |
3a8cff4f | 75 | * |
5760495a | 76 | * This is a variant of generic_file_llseek that allows passing in a custom |
e8b96eb5 | 77 | * maximum file size and a custom EOF position, for e.g. hashed directories |
ef3d0fd2 AK |
78 | * |
79 | * Synchronization: | |
5760495a | 80 | * SEEK_SET and SEEK_END are unsynchronized (but atomic on 64bit platforms) |
ef3d0fd2 AK |
81 | * SEEK_CUR is synchronized against other SEEK_CURs, but not read/writes. |
82 | * read/writes behave like SEEK_SET against seeks. | |
3a8cff4f | 83 | */ |
9465efc9 | 84 | loff_t |
965c8e59 | 85 | generic_file_llseek_size(struct file *file, loff_t offset, int whence, |
e8b96eb5 | 86 | loff_t maxsize, loff_t eof) |
1da177e4 | 87 | { |
965c8e59 | 88 | switch (whence) { |
3a8cff4f | 89 | case SEEK_END: |
e8b96eb5 | 90 | offset += eof; |
3a8cff4f CH |
91 | break; |
92 | case SEEK_CUR: | |
5b6f1eb9 AK |
93 | /* |
94 | * Here we special-case the lseek(fd, 0, SEEK_CUR) | |
95 | * position-querying operation. Avoid rewriting the "same" | |
96 | * f_pos value back to the file because a concurrent read(), | |
97 | * write() or lseek() might have altered it | |
98 | */ | |
99 | if (offset == 0) | |
100 | return file->f_pos; | |
ef3d0fd2 AK |
101 | /* |
102 | * f_lock protects against read/modify/write race with other | |
103 | * SEEK_CURs. Note that parallel writes and reads behave | |
104 | * like SEEK_SET. | |
105 | */ | |
106 | spin_lock(&file->f_lock); | |
46a1c2c7 | 107 | offset = vfs_setpos(file, file->f_pos + offset, maxsize); |
ef3d0fd2 AK |
108 | spin_unlock(&file->f_lock); |
109 | return offset; | |
982d8165 JB |
110 | case SEEK_DATA: |
111 | /* | |
112 | * In the generic case the entire file is data, so as long as | |
113 | * offset isn't at the end of the file then the offset is data. | |
114 | */ | |
e8b96eb5 | 115 | if (offset >= eof) |
982d8165 JB |
116 | return -ENXIO; |
117 | break; | |
118 | case SEEK_HOLE: | |
119 | /* | |
120 | * There is a virtual hole at the end of the file, so as long as | |
121 | * offset isn't i_size or larger, return i_size. | |
122 | */ | |
e8b96eb5 | 123 | if (offset >= eof) |
982d8165 | 124 | return -ENXIO; |
e8b96eb5 | 125 | offset = eof; |
982d8165 | 126 | break; |
1da177e4 | 127 | } |
3a8cff4f | 128 | |
46a1c2c7 | 129 | return vfs_setpos(file, offset, maxsize); |
5760495a AK |
130 | } |
131 | EXPORT_SYMBOL(generic_file_llseek_size); | |
132 | ||
133 | /** | |
134 | * generic_file_llseek - generic llseek implementation for regular files | |
135 | * @file: file structure to seek on | |
136 | * @offset: file offset to seek to | |
965c8e59 | 137 | * @whence: type of seek |
5760495a AK |
138 | * |
139 | * This is a generic implemenation of ->llseek useable for all normal local | |
140 | * filesystems. It just updates the file offset to the value specified by | |
546ae2d2 | 141 | * @offset and @whence. |
5760495a | 142 | */ |
965c8e59 | 143 | loff_t generic_file_llseek(struct file *file, loff_t offset, int whence) |
5760495a AK |
144 | { |
145 | struct inode *inode = file->f_mapping->host; | |
146 | ||
965c8e59 | 147 | return generic_file_llseek_size(file, offset, whence, |
e8b96eb5 ES |
148 | inode->i_sb->s_maxbytes, |
149 | i_size_read(inode)); | |
1da177e4 | 150 | } |
9465efc9 | 151 | EXPORT_SYMBOL(generic_file_llseek); |
1da177e4 | 152 | |
1bf9d14d AV |
153 | /** |
154 | * fixed_size_llseek - llseek implementation for fixed-sized devices | |
155 | * @file: file structure to seek on | |
156 | * @offset: file offset to seek to | |
157 | * @whence: type of seek | |
158 | * @size: size of the file | |
159 | * | |
160 | */ | |
161 | loff_t fixed_size_llseek(struct file *file, loff_t offset, int whence, loff_t size) | |
162 | { | |
163 | switch (whence) { | |
164 | case SEEK_SET: case SEEK_CUR: case SEEK_END: | |
165 | return generic_file_llseek_size(file, offset, whence, | |
166 | size, size); | |
167 | default: | |
168 | return -EINVAL; | |
169 | } | |
170 | } | |
171 | EXPORT_SYMBOL(fixed_size_llseek); | |
172 | ||
b25472f9 AV |
173 | /** |
174 | * no_seek_end_llseek - llseek implementation for fixed-sized devices | |
175 | * @file: file structure to seek on | |
176 | * @offset: file offset to seek to | |
177 | * @whence: type of seek | |
178 | * | |
179 | */ | |
180 | loff_t no_seek_end_llseek(struct file *file, loff_t offset, int whence) | |
181 | { | |
182 | switch (whence) { | |
183 | case SEEK_SET: case SEEK_CUR: | |
184 | return generic_file_llseek_size(file, offset, whence, | |
2feb55f8 | 185 | OFFSET_MAX, 0); |
b25472f9 AV |
186 | default: |
187 | return -EINVAL; | |
188 | } | |
189 | } | |
190 | EXPORT_SYMBOL(no_seek_end_llseek); | |
191 | ||
192 | /** | |
193 | * no_seek_end_llseek_size - llseek implementation for fixed-sized devices | |
194 | * @file: file structure to seek on | |
195 | * @offset: file offset to seek to | |
196 | * @whence: type of seek | |
197 | * @size: maximal offset allowed | |
198 | * | |
199 | */ | |
200 | loff_t no_seek_end_llseek_size(struct file *file, loff_t offset, int whence, loff_t size) | |
201 | { | |
202 | switch (whence) { | |
203 | case SEEK_SET: case SEEK_CUR: | |
204 | return generic_file_llseek_size(file, offset, whence, | |
205 | size, 0); | |
206 | default: | |
207 | return -EINVAL; | |
208 | } | |
209 | } | |
210 | EXPORT_SYMBOL(no_seek_end_llseek_size); | |
211 | ||
ae6afc3f B |
212 | /** |
213 | * noop_llseek - No Operation Performed llseek implementation | |
214 | * @file: file structure to seek on | |
215 | * @offset: file offset to seek to | |
965c8e59 | 216 | * @whence: type of seek |
ae6afc3f B |
217 | * |
218 | * This is an implementation of ->llseek useable for the rare special case when | |
219 | * userspace expects the seek to succeed but the (device) file is actually not | |
220 | * able to perform the seek. In this case you use noop_llseek() instead of | |
221 | * falling back to the default implementation of ->llseek. | |
222 | */ | |
965c8e59 | 223 | loff_t noop_llseek(struct file *file, loff_t offset, int whence) |
ae6afc3f B |
224 | { |
225 | return file->f_pos; | |
226 | } | |
227 | EXPORT_SYMBOL(noop_llseek); | |
228 | ||
965c8e59 | 229 | loff_t no_llseek(struct file *file, loff_t offset, int whence) |
1da177e4 LT |
230 | { |
231 | return -ESPIPE; | |
232 | } | |
233 | EXPORT_SYMBOL(no_llseek); | |
234 | ||
965c8e59 | 235 | loff_t default_llseek(struct file *file, loff_t offset, int whence) |
1da177e4 | 236 | { |
496ad9aa | 237 | struct inode *inode = file_inode(file); |
16abef0e | 238 | loff_t retval; |
1da177e4 | 239 | |
5955102c | 240 | inode_lock(inode); |
965c8e59 | 241 | switch (whence) { |
7b8e8924 | 242 | case SEEK_END: |
982d8165 | 243 | offset += i_size_read(inode); |
1da177e4 | 244 | break; |
7b8e8924 | 245 | case SEEK_CUR: |
5b6f1eb9 AK |
246 | if (offset == 0) { |
247 | retval = file->f_pos; | |
248 | goto out; | |
249 | } | |
1da177e4 | 250 | offset += file->f_pos; |
982d8165 JB |
251 | break; |
252 | case SEEK_DATA: | |
253 | /* | |
254 | * In the generic case the entire file is data, so as | |
255 | * long as offset isn't at the end of the file then the | |
256 | * offset is data. | |
257 | */ | |
bacb2d81 DC |
258 | if (offset >= inode->i_size) { |
259 | retval = -ENXIO; | |
260 | goto out; | |
261 | } | |
982d8165 JB |
262 | break; |
263 | case SEEK_HOLE: | |
264 | /* | |
265 | * There is a virtual hole at the end of the file, so | |
266 | * as long as offset isn't i_size or larger, return | |
267 | * i_size. | |
268 | */ | |
bacb2d81 DC |
269 | if (offset >= inode->i_size) { |
270 | retval = -ENXIO; | |
271 | goto out; | |
272 | } | |
982d8165 JB |
273 | offset = inode->i_size; |
274 | break; | |
1da177e4 LT |
275 | } |
276 | retval = -EINVAL; | |
cccb5a1e | 277 | if (offset >= 0 || unsigned_offsets(file)) { |
1da177e4 LT |
278 | if (offset != file->f_pos) { |
279 | file->f_pos = offset; | |
280 | file->f_version = 0; | |
281 | } | |
282 | retval = offset; | |
283 | } | |
5b6f1eb9 | 284 | out: |
5955102c | 285 | inode_unlock(inode); |
1da177e4 LT |
286 | return retval; |
287 | } | |
288 | EXPORT_SYMBOL(default_llseek); | |
289 | ||
965c8e59 | 290 | loff_t vfs_llseek(struct file *file, loff_t offset, int whence) |
1da177e4 LT |
291 | { |
292 | loff_t (*fn)(struct file *, loff_t, int); | |
293 | ||
294 | fn = no_llseek; | |
295 | if (file->f_mode & FMODE_LSEEK) { | |
72c2d531 | 296 | if (file->f_op->llseek) |
1da177e4 LT |
297 | fn = file->f_op->llseek; |
298 | } | |
965c8e59 | 299 | return fn(file, offset, whence); |
1da177e4 LT |
300 | } |
301 | EXPORT_SYMBOL(vfs_llseek); | |
302 | ||
965c8e59 | 303 | SYSCALL_DEFINE3(lseek, unsigned int, fd, off_t, offset, unsigned int, whence) |
1da177e4 LT |
304 | { |
305 | off_t retval; | |
9c225f26 | 306 | struct fd f = fdget_pos(fd); |
2903ff01 AV |
307 | if (!f.file) |
308 | return -EBADF; | |
1da177e4 LT |
309 | |
310 | retval = -EINVAL; | |
965c8e59 AM |
311 | if (whence <= SEEK_MAX) { |
312 | loff_t res = vfs_llseek(f.file, offset, whence); | |
1da177e4 LT |
313 | retval = res; |
314 | if (res != (loff_t)retval) | |
315 | retval = -EOVERFLOW; /* LFS: should only happen on 32 bit platforms */ | |
316 | } | |
9c225f26 | 317 | fdput_pos(f); |
1da177e4 LT |
318 | return retval; |
319 | } | |
320 | ||
561c6731 AV |
321 | #ifdef CONFIG_COMPAT |
322 | COMPAT_SYSCALL_DEFINE3(lseek, unsigned int, fd, compat_off_t, offset, unsigned int, whence) | |
323 | { | |
324 | return sys_lseek(fd, offset, whence); | |
325 | } | |
326 | #endif | |
327 | ||
1da177e4 | 328 | #ifdef __ARCH_WANT_SYS_LLSEEK |
003d7ab4 HC |
329 | SYSCALL_DEFINE5(llseek, unsigned int, fd, unsigned long, offset_high, |
330 | unsigned long, offset_low, loff_t __user *, result, | |
965c8e59 | 331 | unsigned int, whence) |
1da177e4 LT |
332 | { |
333 | int retval; | |
d7a15f8d | 334 | struct fd f = fdget_pos(fd); |
1da177e4 | 335 | loff_t offset; |
1da177e4 | 336 | |
2903ff01 AV |
337 | if (!f.file) |
338 | return -EBADF; | |
1da177e4 LT |
339 | |
340 | retval = -EINVAL; | |
965c8e59 | 341 | if (whence > SEEK_MAX) |
1da177e4 LT |
342 | goto out_putf; |
343 | ||
2903ff01 | 344 | offset = vfs_llseek(f.file, ((loff_t) offset_high << 32) | offset_low, |
965c8e59 | 345 | whence); |
1da177e4 LT |
346 | |
347 | retval = (int)offset; | |
348 | if (offset >= 0) { | |
349 | retval = -EFAULT; | |
350 | if (!copy_to_user(result, &offset, sizeof(offset))) | |
351 | retval = 0; | |
352 | } | |
353 | out_putf: | |
d7a15f8d | 354 | fdput_pos(f); |
1da177e4 LT |
355 | return retval; |
356 | } | |
357 | #endif | |
358 | ||
68d70d03 | 359 | int rw_verify_area(int read_write, struct file *file, const loff_t *ppos, size_t count) |
1da177e4 LT |
360 | { |
361 | struct inode *inode; | |
362 | loff_t pos; | |
c43e259c | 363 | int retval = -EINVAL; |
1da177e4 | 364 | |
496ad9aa | 365 | inode = file_inode(file); |
e28cc715 | 366 | if (unlikely((ssize_t) count < 0)) |
c43e259c | 367 | return retval; |
1da177e4 | 368 | pos = *ppos; |
cccb5a1e AV |
369 | if (unlikely(pos < 0)) { |
370 | if (!unsigned_offsets(file)) | |
371 | return retval; | |
372 | if (count >= -pos) /* both values are in 0..LLONG_MAX */ | |
373 | return -EOVERFLOW; | |
374 | } else if (unlikely((loff_t) (pos + count) < 0)) { | |
375 | if (!unsigned_offsets(file)) | |
4a3956c7 KH |
376 | return retval; |
377 | } | |
1da177e4 | 378 | |
bd61e0a9 | 379 | if (unlikely(inode->i_flctx && mandatory_lock(inode))) { |
acc15575 CH |
380 | retval = locks_mandatory_area(inode, file, pos, pos + count - 1, |
381 | read_write == READ ? F_RDLCK : F_WRLCK); | |
e28cc715 LT |
382 | if (retval < 0) |
383 | return retval; | |
384 | } | |
bc61384d | 385 | return security_file_permission(file, |
c43e259c | 386 | read_write == READ ? MAY_READ : MAY_WRITE); |
1da177e4 LT |
387 | } |
388 | ||
5d5d5689 | 389 | static ssize_t new_sync_read(struct file *filp, char __user *buf, size_t len, loff_t *ppos) |
293bc982 AV |
390 | { |
391 | struct iovec iov = { .iov_base = buf, .iov_len = len }; | |
392 | struct kiocb kiocb; | |
393 | struct iov_iter iter; | |
394 | ssize_t ret; | |
395 | ||
396 | init_sync_kiocb(&kiocb, filp); | |
397 | kiocb.ki_pos = *ppos; | |
293bc982 AV |
398 | iov_iter_init(&iter, READ, &iov, 1, len); |
399 | ||
bb7462b6 | 400 | ret = call_read_iter(filp, &kiocb, &iter); |
599bd19b | 401 | BUG_ON(ret == -EIOCBQUEUED); |
293bc982 AV |
402 | *ppos = kiocb.ki_pos; |
403 | return ret; | |
404 | } | |
405 | ||
6fb5032e DK |
406 | ssize_t __vfs_read(struct file *file, char __user *buf, size_t count, |
407 | loff_t *pos) | |
408 | { | |
6fb5032e | 409 | if (file->f_op->read) |
3d04c8a1 | 410 | return file->f_op->read(file, buf, count, pos); |
6fb5032e | 411 | else if (file->f_op->read_iter) |
3d04c8a1 | 412 | return new_sync_read(file, buf, count, pos); |
6fb5032e | 413 | else |
3d04c8a1 | 414 | return -EINVAL; |
6fb5032e | 415 | } |
3d04c8a1 | 416 | EXPORT_SYMBOL(__vfs_read); |
6fb5032e | 417 | |
c41fbad0 CH |
418 | int kernel_read(struct file *file, loff_t offset, char *addr, |
419 | unsigned long count) | |
420 | { | |
421 | mm_segment_t old_fs; | |
422 | loff_t pos = offset; | |
423 | int result; | |
424 | ||
425 | old_fs = get_fs(); | |
426 | set_fs(get_ds()); | |
427 | /* The cast to a user pointer is valid due to the set_fs() */ | |
428 | result = vfs_read(file, (void __user *)addr, count, &pos); | |
429 | set_fs(old_fs); | |
430 | return result; | |
431 | } | |
432 | EXPORT_SYMBOL(kernel_read); | |
433 | ||
1da177e4 LT |
434 | ssize_t vfs_read(struct file *file, char __user *buf, size_t count, loff_t *pos) |
435 | { | |
436 | ssize_t ret; | |
437 | ||
438 | if (!(file->f_mode & FMODE_READ)) | |
439 | return -EBADF; | |
7f7f25e8 | 440 | if (!(file->f_mode & FMODE_CAN_READ)) |
1da177e4 LT |
441 | return -EINVAL; |
442 | if (unlikely(!access_ok(VERIFY_WRITE, buf, count))) | |
443 | return -EFAULT; | |
444 | ||
445 | ret = rw_verify_area(READ, file, pos, count); | |
bc61384d AV |
446 | if (!ret) { |
447 | if (count > MAX_RW_COUNT) | |
448 | count = MAX_RW_COUNT; | |
6fb5032e | 449 | ret = __vfs_read(file, buf, count, pos); |
c43e259c | 450 | if (ret > 0) { |
2a12a9d7 | 451 | fsnotify_access(file); |
c43e259c | 452 | add_rchar(current, ret); |
1da177e4 | 453 | } |
c43e259c | 454 | inc_syscr(current); |
1da177e4 LT |
455 | } |
456 | ||
457 | return ret; | |
458 | } | |
459 | ||
460 | EXPORT_SYMBOL(vfs_read); | |
461 | ||
5d5d5689 | 462 | static ssize_t new_sync_write(struct file *filp, const char __user *buf, size_t len, loff_t *ppos) |
293bc982 AV |
463 | { |
464 | struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = len }; | |
465 | struct kiocb kiocb; | |
466 | struct iov_iter iter; | |
467 | ssize_t ret; | |
468 | ||
469 | init_sync_kiocb(&kiocb, filp); | |
470 | kiocb.ki_pos = *ppos; | |
293bc982 AV |
471 | iov_iter_init(&iter, WRITE, &iov, 1, len); |
472 | ||
bb7462b6 | 473 | ret = call_write_iter(filp, &kiocb, &iter); |
599bd19b | 474 | BUG_ON(ret == -EIOCBQUEUED); |
f765b134 AV |
475 | if (ret > 0) |
476 | *ppos = kiocb.ki_pos; | |
293bc982 AV |
477 | return ret; |
478 | } | |
479 | ||
493c84c0 AV |
480 | ssize_t __vfs_write(struct file *file, const char __user *p, size_t count, |
481 | loff_t *pos) | |
482 | { | |
483 | if (file->f_op->write) | |
484 | return file->f_op->write(file, p, count, pos); | |
493c84c0 AV |
485 | else if (file->f_op->write_iter) |
486 | return new_sync_write(file, p, count, pos); | |
487 | else | |
488 | return -EINVAL; | |
489 | } | |
490 | EXPORT_SYMBOL(__vfs_write); | |
491 | ||
06ae43f3 AV |
492 | ssize_t __kernel_write(struct file *file, const char *buf, size_t count, loff_t *pos) |
493 | { | |
494 | mm_segment_t old_fs; | |
495 | const char __user *p; | |
496 | ssize_t ret; | |
497 | ||
7f7f25e8 | 498 | if (!(file->f_mode & FMODE_CAN_WRITE)) |
3e84f48e AV |
499 | return -EINVAL; |
500 | ||
06ae43f3 AV |
501 | old_fs = get_fs(); |
502 | set_fs(get_ds()); | |
503 | p = (__force const char __user *)buf; | |
504 | if (count > MAX_RW_COUNT) | |
505 | count = MAX_RW_COUNT; | |
493c84c0 | 506 | ret = __vfs_write(file, p, count, pos); |
06ae43f3 AV |
507 | set_fs(old_fs); |
508 | if (ret > 0) { | |
509 | fsnotify_modify(file); | |
510 | add_wchar(current, ret); | |
511 | } | |
512 | inc_syscw(current); | |
513 | return ret; | |
514 | } | |
2ec3a12a AV |
515 | EXPORT_SYMBOL(__kernel_write); |
516 | ||
ac452aca CH |
517 | ssize_t kernel_write(struct file *file, const char *buf, size_t count, |
518 | loff_t pos) | |
519 | { | |
520 | mm_segment_t old_fs; | |
521 | ssize_t res; | |
522 | ||
523 | old_fs = get_fs(); | |
524 | set_fs(get_ds()); | |
525 | /* The cast to a user pointer is valid due to the set_fs() */ | |
526 | res = vfs_write(file, (__force const char __user *)buf, count, &pos); | |
527 | set_fs(old_fs); | |
528 | ||
529 | return res; | |
530 | } | |
531 | EXPORT_SYMBOL(kernel_write); | |
532 | ||
1da177e4 LT |
533 | ssize_t vfs_write(struct file *file, const char __user *buf, size_t count, loff_t *pos) |
534 | { | |
535 | ssize_t ret; | |
536 | ||
537 | if (!(file->f_mode & FMODE_WRITE)) | |
538 | return -EBADF; | |
7f7f25e8 | 539 | if (!(file->f_mode & FMODE_CAN_WRITE)) |
1da177e4 LT |
540 | return -EINVAL; |
541 | if (unlikely(!access_ok(VERIFY_READ, buf, count))) | |
542 | return -EFAULT; | |
543 | ||
544 | ret = rw_verify_area(WRITE, file, pos, count); | |
bc61384d AV |
545 | if (!ret) { |
546 | if (count > MAX_RW_COUNT) | |
547 | count = MAX_RW_COUNT; | |
03d95eb2 | 548 | file_start_write(file); |
493c84c0 | 549 | ret = __vfs_write(file, buf, count, pos); |
c43e259c | 550 | if (ret > 0) { |
2a12a9d7 | 551 | fsnotify_modify(file); |
c43e259c | 552 | add_wchar(current, ret); |
1da177e4 | 553 | } |
c43e259c | 554 | inc_syscw(current); |
03d95eb2 | 555 | file_end_write(file); |
1da177e4 LT |
556 | } |
557 | ||
558 | return ret; | |
559 | } | |
560 | ||
561 | EXPORT_SYMBOL(vfs_write); | |
562 | ||
563 | static inline loff_t file_pos_read(struct file *file) | |
564 | { | |
565 | return file->f_pos; | |
566 | } | |
567 | ||
568 | static inline void file_pos_write(struct file *file, loff_t pos) | |
569 | { | |
570 | file->f_pos = pos; | |
571 | } | |
572 | ||
3cdad428 | 573 | SYSCALL_DEFINE3(read, unsigned int, fd, char __user *, buf, size_t, count) |
1da177e4 | 574 | { |
9c225f26 | 575 | struct fd f = fdget_pos(fd); |
1da177e4 | 576 | ssize_t ret = -EBADF; |
1da177e4 | 577 | |
2903ff01 AV |
578 | if (f.file) { |
579 | loff_t pos = file_pos_read(f.file); | |
580 | ret = vfs_read(f.file, buf, count, &pos); | |
5faf153e AV |
581 | if (ret >= 0) |
582 | file_pos_write(f.file, pos); | |
9c225f26 | 583 | fdput_pos(f); |
1da177e4 | 584 | } |
1da177e4 LT |
585 | return ret; |
586 | } | |
1da177e4 | 587 | |
3cdad428 HC |
588 | SYSCALL_DEFINE3(write, unsigned int, fd, const char __user *, buf, |
589 | size_t, count) | |
1da177e4 | 590 | { |
9c225f26 | 591 | struct fd f = fdget_pos(fd); |
1da177e4 | 592 | ssize_t ret = -EBADF; |
1da177e4 | 593 | |
2903ff01 AV |
594 | if (f.file) { |
595 | loff_t pos = file_pos_read(f.file); | |
596 | ret = vfs_write(f.file, buf, count, &pos); | |
5faf153e AV |
597 | if (ret >= 0) |
598 | file_pos_write(f.file, pos); | |
9c225f26 | 599 | fdput_pos(f); |
1da177e4 LT |
600 | } |
601 | ||
602 | return ret; | |
603 | } | |
604 | ||
4a0fd5bf AV |
605 | SYSCALL_DEFINE4(pread64, unsigned int, fd, char __user *, buf, |
606 | size_t, count, loff_t, pos) | |
1da177e4 | 607 | { |
2903ff01 | 608 | struct fd f; |
1da177e4 | 609 | ssize_t ret = -EBADF; |
1da177e4 LT |
610 | |
611 | if (pos < 0) | |
612 | return -EINVAL; | |
613 | ||
2903ff01 AV |
614 | f = fdget(fd); |
615 | if (f.file) { | |
1da177e4 | 616 | ret = -ESPIPE; |
2903ff01 AV |
617 | if (f.file->f_mode & FMODE_PREAD) |
618 | ret = vfs_read(f.file, buf, count, &pos); | |
619 | fdput(f); | |
1da177e4 LT |
620 | } |
621 | ||
622 | return ret; | |
623 | } | |
624 | ||
4a0fd5bf AV |
625 | SYSCALL_DEFINE4(pwrite64, unsigned int, fd, const char __user *, buf, |
626 | size_t, count, loff_t, pos) | |
1da177e4 | 627 | { |
2903ff01 | 628 | struct fd f; |
1da177e4 | 629 | ssize_t ret = -EBADF; |
1da177e4 LT |
630 | |
631 | if (pos < 0) | |
632 | return -EINVAL; | |
633 | ||
2903ff01 AV |
634 | f = fdget(fd); |
635 | if (f.file) { | |
1da177e4 | 636 | ret = -ESPIPE; |
2903ff01 AV |
637 | if (f.file->f_mode & FMODE_PWRITE) |
638 | ret = vfs_write(f.file, buf, count, &pos); | |
639 | fdput(f); | |
1da177e4 LT |
640 | } |
641 | ||
642 | return ret; | |
643 | } | |
644 | ||
645 | /* | |
646 | * Reduce an iovec's length in-place. Return the resulting number of segments | |
647 | */ | |
648 | unsigned long iov_shorten(struct iovec *iov, unsigned long nr_segs, size_t to) | |
649 | { | |
650 | unsigned long seg = 0; | |
651 | size_t len = 0; | |
652 | ||
653 | while (seg < nr_segs) { | |
654 | seg++; | |
655 | if (len + iov->iov_len >= to) { | |
656 | iov->iov_len = to - len; | |
657 | break; | |
658 | } | |
659 | len += iov->iov_len; | |
660 | iov++; | |
661 | } | |
662 | return seg; | |
663 | } | |
19295529 | 664 | EXPORT_SYMBOL(iov_shorten); |
1da177e4 | 665 | |
ac15ac06 | 666 | static ssize_t do_iter_readv_writev(struct file *filp, struct iov_iter *iter, |
0f78d06a | 667 | loff_t *ppos, int type, int flags) |
293bc982 AV |
668 | { |
669 | struct kiocb kiocb; | |
293bc982 AV |
670 | ssize_t ret; |
671 | ||
672 | init_sync_kiocb(&kiocb, filp); | |
fdd2f5b7 GR |
673 | ret = kiocb_set_rw_flags(&kiocb, flags); |
674 | if (ret) | |
675 | return ret; | |
293bc982 | 676 | kiocb.ki_pos = *ppos; |
293bc982 | 677 | |
0f78d06a | 678 | if (type == READ) |
bb7462b6 | 679 | ret = call_read_iter(filp, &kiocb, iter); |
0f78d06a | 680 | else |
bb7462b6 | 681 | ret = call_write_iter(filp, &kiocb, iter); |
599bd19b | 682 | BUG_ON(ret == -EIOCBQUEUED); |
293bc982 AV |
683 | *ppos = kiocb.ki_pos; |
684 | return ret; | |
685 | } | |
686 | ||
ee0b3e67 | 687 | /* Do it by hand, with file-ops */ |
ac15ac06 | 688 | static ssize_t do_loop_readv_writev(struct file *filp, struct iov_iter *iter, |
0f78d06a | 689 | loff_t *ppos, int type, int flags) |
ee0b3e67 | 690 | { |
ee0b3e67 BP |
691 | ssize_t ret = 0; |
692 | ||
97be7ebe | 693 | if (flags & ~RWF_HIPRI) |
793b80ef CH |
694 | return -EOPNOTSUPP; |
695 | ||
ac15ac06 AV |
696 | while (iov_iter_count(iter)) { |
697 | struct iovec iovec = iov_iter_iovec(iter); | |
ee0b3e67 BP |
698 | ssize_t nr; |
699 | ||
0f78d06a MS |
700 | if (type == READ) { |
701 | nr = filp->f_op->read(filp, iovec.iov_base, | |
702 | iovec.iov_len, ppos); | |
703 | } else { | |
704 | nr = filp->f_op->write(filp, iovec.iov_base, | |
705 | iovec.iov_len, ppos); | |
706 | } | |
ee0b3e67 BP |
707 | |
708 | if (nr < 0) { | |
709 | if (!ret) | |
710 | ret = nr; | |
711 | break; | |
712 | } | |
713 | ret += nr; | |
ac15ac06 | 714 | if (nr != iovec.iov_len) |
ee0b3e67 | 715 | break; |
ac15ac06 | 716 | iov_iter_advance(iter, nr); |
ee0b3e67 BP |
717 | } |
718 | ||
719 | return ret; | |
720 | } | |
721 | ||
1da177e4 LT |
722 | /* A write operation does a read from user space and vice versa */ |
723 | #define vrfy_dir(type) ((type) == READ ? VERIFY_WRITE : VERIFY_READ) | |
724 | ||
ffecee4f VN |
725 | /** |
726 | * rw_copy_check_uvector() - Copy an array of &struct iovec from userspace | |
727 | * into the kernel and check that it is valid. | |
728 | * | |
729 | * @type: One of %CHECK_IOVEC_ONLY, %READ, or %WRITE. | |
730 | * @uvector: Pointer to the userspace array. | |
731 | * @nr_segs: Number of elements in userspace array. | |
732 | * @fast_segs: Number of elements in @fast_pointer. | |
733 | * @fast_pointer: Pointer to (usually small on-stack) kernel array. | |
734 | * @ret_pointer: (output parameter) Pointer to a variable that will point to | |
735 | * either @fast_pointer, a newly allocated kernel array, or NULL, | |
736 | * depending on which array was used. | |
737 | * | |
738 | * This function copies an array of &struct iovec of @nr_segs from | |
739 | * userspace into the kernel and checks that each element is valid (e.g. | |
740 | * it does not point to a kernel address or cause overflow by being too | |
741 | * large, etc.). | |
742 | * | |
743 | * As an optimization, the caller may provide a pointer to a small | |
744 | * on-stack array in @fast_pointer, typically %UIO_FASTIOV elements long | |
745 | * (the size of this array, or 0 if unused, should be given in @fast_segs). | |
746 | * | |
747 | * @ret_pointer will always point to the array that was used, so the | |
748 | * caller must take care not to call kfree() on it e.g. in case the | |
749 | * @fast_pointer array was used and it was allocated on the stack. | |
750 | * | |
751 | * Return: The total number of bytes covered by the iovec array on success | |
752 | * or a negative error code on error. | |
753 | */ | |
eed4e51f BP |
754 | ssize_t rw_copy_check_uvector(int type, const struct iovec __user * uvector, |
755 | unsigned long nr_segs, unsigned long fast_segs, | |
756 | struct iovec *fast_pointer, | |
ac34ebb3 | 757 | struct iovec **ret_pointer) |
435f49a5 | 758 | { |
eed4e51f | 759 | unsigned long seg; |
435f49a5 | 760 | ssize_t ret; |
eed4e51f BP |
761 | struct iovec *iov = fast_pointer; |
762 | ||
435f49a5 LT |
763 | /* |
764 | * SuS says "The readv() function *may* fail if the iovcnt argument | |
765 | * was less than or equal to 0, or greater than {IOV_MAX}. Linux has | |
766 | * traditionally returned zero for zero segments, so... | |
767 | */ | |
eed4e51f BP |
768 | if (nr_segs == 0) { |
769 | ret = 0; | |
435f49a5 | 770 | goto out; |
eed4e51f BP |
771 | } |
772 | ||
435f49a5 LT |
773 | /* |
774 | * First get the "struct iovec" from user memory and | |
775 | * verify all the pointers | |
776 | */ | |
eed4e51f BP |
777 | if (nr_segs > UIO_MAXIOV) { |
778 | ret = -EINVAL; | |
435f49a5 | 779 | goto out; |
eed4e51f BP |
780 | } |
781 | if (nr_segs > fast_segs) { | |
435f49a5 | 782 | iov = kmalloc(nr_segs*sizeof(struct iovec), GFP_KERNEL); |
eed4e51f BP |
783 | if (iov == NULL) { |
784 | ret = -ENOMEM; | |
435f49a5 | 785 | goto out; |
eed4e51f | 786 | } |
435f49a5 | 787 | } |
eed4e51f BP |
788 | if (copy_from_user(iov, uvector, nr_segs*sizeof(*uvector))) { |
789 | ret = -EFAULT; | |
435f49a5 | 790 | goto out; |
eed4e51f BP |
791 | } |
792 | ||
435f49a5 | 793 | /* |
eed4e51f BP |
794 | * According to the Single Unix Specification we should return EINVAL |
795 | * if an element length is < 0 when cast to ssize_t or if the | |
796 | * total length would overflow the ssize_t return value of the | |
797 | * system call. | |
435f49a5 LT |
798 | * |
799 | * Linux caps all read/write calls to MAX_RW_COUNT, and avoids the | |
800 | * overflow case. | |
801 | */ | |
eed4e51f | 802 | ret = 0; |
435f49a5 LT |
803 | for (seg = 0; seg < nr_segs; seg++) { |
804 | void __user *buf = iov[seg].iov_base; | |
805 | ssize_t len = (ssize_t)iov[seg].iov_len; | |
eed4e51f BP |
806 | |
807 | /* see if we we're about to use an invalid len or if | |
808 | * it's about to overflow ssize_t */ | |
435f49a5 | 809 | if (len < 0) { |
eed4e51f | 810 | ret = -EINVAL; |
435f49a5 | 811 | goto out; |
eed4e51f | 812 | } |
ac34ebb3 | 813 | if (type >= 0 |
fcf63409 | 814 | && unlikely(!access_ok(vrfy_dir(type), buf, len))) { |
eed4e51f | 815 | ret = -EFAULT; |
435f49a5 LT |
816 | goto out; |
817 | } | |
818 | if (len > MAX_RW_COUNT - ret) { | |
819 | len = MAX_RW_COUNT - ret; | |
820 | iov[seg].iov_len = len; | |
eed4e51f | 821 | } |
eed4e51f | 822 | ret += len; |
435f49a5 | 823 | } |
eed4e51f BP |
824 | out: |
825 | *ret_pointer = iov; | |
826 | return ret; | |
827 | } | |
828 | ||
f5029855 AV |
829 | #ifdef CONFIG_COMPAT |
830 | ssize_t compat_rw_copy_check_uvector(int type, | |
831 | const struct compat_iovec __user *uvector, unsigned long nr_segs, | |
832 | unsigned long fast_segs, struct iovec *fast_pointer, | |
833 | struct iovec **ret_pointer) | |
834 | { | |
835 | compat_ssize_t tot_len; | |
836 | struct iovec *iov = *ret_pointer = fast_pointer; | |
837 | ssize_t ret = 0; | |
838 | int seg; | |
839 | ||
840 | /* | |
841 | * SuS says "The readv() function *may* fail if the iovcnt argument | |
842 | * was less than or equal to 0, or greater than {IOV_MAX}. Linux has | |
843 | * traditionally returned zero for zero segments, so... | |
844 | */ | |
845 | if (nr_segs == 0) | |
846 | goto out; | |
847 | ||
848 | ret = -EINVAL; | |
849 | if (nr_segs > UIO_MAXIOV) | |
850 | goto out; | |
851 | if (nr_segs > fast_segs) { | |
852 | ret = -ENOMEM; | |
853 | iov = kmalloc(nr_segs*sizeof(struct iovec), GFP_KERNEL); | |
854 | if (iov == NULL) | |
855 | goto out; | |
856 | } | |
857 | *ret_pointer = iov; | |
858 | ||
859 | ret = -EFAULT; | |
860 | if (!access_ok(VERIFY_READ, uvector, nr_segs*sizeof(*uvector))) | |
861 | goto out; | |
862 | ||
863 | /* | |
864 | * Single unix specification: | |
865 | * We should -EINVAL if an element length is not >= 0 and fitting an | |
866 | * ssize_t. | |
867 | * | |
868 | * In Linux, the total length is limited to MAX_RW_COUNT, there is | |
869 | * no overflow possibility. | |
870 | */ | |
871 | tot_len = 0; | |
872 | ret = -EINVAL; | |
873 | for (seg = 0; seg < nr_segs; seg++) { | |
874 | compat_uptr_t buf; | |
875 | compat_ssize_t len; | |
876 | ||
877 | if (__get_user(len, &uvector->iov_len) || | |
878 | __get_user(buf, &uvector->iov_base)) { | |
879 | ret = -EFAULT; | |
880 | goto out; | |
881 | } | |
882 | if (len < 0) /* size_t not fitting in compat_ssize_t .. */ | |
883 | goto out; | |
884 | if (type >= 0 && | |
885 | !access_ok(vrfy_dir(type), compat_ptr(buf), len)) { | |
886 | ret = -EFAULT; | |
887 | goto out; | |
888 | } | |
889 | if (len > MAX_RW_COUNT - tot_len) | |
890 | len = MAX_RW_COUNT - tot_len; | |
891 | tot_len += len; | |
892 | iov->iov_base = compat_ptr(buf); | |
893 | iov->iov_len = (compat_size_t) len; | |
894 | uvector++; | |
895 | iov++; | |
896 | } | |
897 | ret = tot_len; | |
898 | ||
899 | out: | |
900 | return ret; | |
901 | } | |
902 | #endif | |
903 | ||
19c73586 CH |
904 | static ssize_t do_iter_read(struct file *file, struct iov_iter *iter, |
905 | loff_t *pos, int flags) | |
1da177e4 | 906 | { |
1da177e4 | 907 | size_t tot_len; |
7687a7a4 | 908 | ssize_t ret = 0; |
1da177e4 | 909 | |
edab5fe3 CH |
910 | if (!(file->f_mode & FMODE_READ)) |
911 | return -EBADF; | |
912 | if (!(file->f_mode & FMODE_CAN_READ)) | |
913 | return -EINVAL; | |
914 | ||
7687a7a4 | 915 | tot_len = iov_iter_count(iter); |
0504c074 AV |
916 | if (!tot_len) |
917 | goto out; | |
19c73586 | 918 | ret = rw_verify_area(READ, file, pos, tot_len); |
e28cc715 | 919 | if (ret < 0) |
19c73586 | 920 | return ret; |
1da177e4 | 921 | |
19c73586 CH |
922 | if (file->f_op->read_iter) |
923 | ret = do_iter_readv_writev(file, iter, pos, READ, flags); | |
ee0b3e67 | 924 | else |
19c73586 | 925 | ret = do_loop_readv_writev(file, iter, pos, READ, flags); |
1da177e4 | 926 | out: |
19c73586 CH |
927 | if (ret >= 0) |
928 | fsnotify_access(file); | |
1da177e4 | 929 | return ret; |
1da177e4 LT |
930 | } |
931 | ||
18e9710e CH |
932 | ssize_t vfs_iter_read(struct file *file, struct iov_iter *iter, loff_t *ppos, |
933 | int flags) | |
7687a7a4 | 934 | { |
18e9710e CH |
935 | if (!file->f_op->read_iter) |
936 | return -EINVAL; | |
937 | return do_iter_read(file, iter, ppos, flags); | |
938 | } | |
939 | EXPORT_SYMBOL(vfs_iter_read); | |
7687a7a4 | 940 | |
19c73586 CH |
941 | static ssize_t do_iter_write(struct file *file, struct iov_iter *iter, |
942 | loff_t *pos, int flags) | |
943 | { | |
944 | size_t tot_len; | |
945 | ssize_t ret = 0; | |
03d95eb2 | 946 | |
edab5fe3 CH |
947 | if (!(file->f_mode & FMODE_WRITE)) |
948 | return -EBADF; | |
949 | if (!(file->f_mode & FMODE_CAN_WRITE)) | |
950 | return -EINVAL; | |
951 | ||
19c73586 CH |
952 | tot_len = iov_iter_count(iter); |
953 | if (!tot_len) | |
954 | return 0; | |
955 | ret = rw_verify_area(WRITE, file, pos, tot_len); | |
7687a7a4 MS |
956 | if (ret < 0) |
957 | return ret; | |
958 | ||
19c73586 CH |
959 | if (file->f_op->write_iter) |
960 | ret = do_iter_readv_writev(file, iter, pos, WRITE, flags); | |
961 | else | |
962 | ret = do_loop_readv_writev(file, iter, pos, WRITE, flags); | |
19c73586 CH |
963 | if (ret > 0) |
964 | fsnotify_modify(file); | |
7687a7a4 MS |
965 | return ret; |
966 | } | |
967 | ||
abbb6589 CH |
968 | ssize_t vfs_iter_write(struct file *file, struct iov_iter *iter, loff_t *ppos, |
969 | int flags) | |
970 | { | |
971 | if (!file->f_op->write_iter) | |
972 | return -EINVAL; | |
973 | return do_iter_write(file, iter, ppos, flags); | |
974 | } | |
975 | EXPORT_SYMBOL(vfs_iter_write); | |
976 | ||
1da177e4 | 977 | ssize_t vfs_readv(struct file *file, const struct iovec __user *vec, |
793b80ef | 978 | unsigned long vlen, loff_t *pos, int flags) |
1da177e4 | 979 | { |
7687a7a4 MS |
980 | struct iovec iovstack[UIO_FASTIOV]; |
981 | struct iovec *iov = iovstack; | |
982 | struct iov_iter iter; | |
983 | ssize_t ret; | |
1da177e4 | 984 | |
251b42a1 | 985 | ret = import_iovec(READ, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter); |
edab5fe3 CH |
986 | if (ret >= 0) { |
987 | ret = do_iter_read(file, &iter, pos, flags); | |
988 | kfree(iov); | |
989 | } | |
1da177e4 | 990 | |
251b42a1 CH |
991 | return ret; |
992 | } | |
1da177e4 LT |
993 | EXPORT_SYMBOL(vfs_readv); |
994 | ||
995 | ssize_t vfs_writev(struct file *file, const struct iovec __user *vec, | |
793b80ef | 996 | unsigned long vlen, loff_t *pos, int flags) |
1da177e4 | 997 | { |
251b42a1 CH |
998 | struct iovec iovstack[UIO_FASTIOV]; |
999 | struct iovec *iov = iovstack; | |
1000 | struct iov_iter iter; | |
1001 | ssize_t ret; | |
1da177e4 | 1002 | |
251b42a1 | 1003 | ret = import_iovec(WRITE, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter); |
edab5fe3 | 1004 | if (ret >= 0) { |
62473a2d | 1005 | file_start_write(file); |
edab5fe3 | 1006 | ret = do_iter_write(file, &iter, pos, flags); |
62473a2d | 1007 | file_end_write(file); |
edab5fe3 CH |
1008 | kfree(iov); |
1009 | } | |
251b42a1 | 1010 | return ret; |
1da177e4 | 1011 | } |
1da177e4 LT |
1012 | EXPORT_SYMBOL(vfs_writev); |
1013 | ||
f17d8b35 MT |
1014 | static ssize_t do_readv(unsigned long fd, const struct iovec __user *vec, |
1015 | unsigned long vlen, int flags) | |
1da177e4 | 1016 | { |
9c225f26 | 1017 | struct fd f = fdget_pos(fd); |
1da177e4 | 1018 | ssize_t ret = -EBADF; |
1da177e4 | 1019 | |
2903ff01 AV |
1020 | if (f.file) { |
1021 | loff_t pos = file_pos_read(f.file); | |
f17d8b35 | 1022 | ret = vfs_readv(f.file, vec, vlen, &pos, flags); |
5faf153e AV |
1023 | if (ret >= 0) |
1024 | file_pos_write(f.file, pos); | |
9c225f26 | 1025 | fdput_pos(f); |
1da177e4 LT |
1026 | } |
1027 | ||
1028 | if (ret > 0) | |
4b98d11b AD |
1029 | add_rchar(current, ret); |
1030 | inc_syscr(current); | |
1da177e4 LT |
1031 | return ret; |
1032 | } | |
1033 | ||
f17d8b35 MT |
1034 | static ssize_t do_writev(unsigned long fd, const struct iovec __user *vec, |
1035 | unsigned long vlen, int flags) | |
1da177e4 | 1036 | { |
9c225f26 | 1037 | struct fd f = fdget_pos(fd); |
1da177e4 | 1038 | ssize_t ret = -EBADF; |
1da177e4 | 1039 | |
2903ff01 AV |
1040 | if (f.file) { |
1041 | loff_t pos = file_pos_read(f.file); | |
f17d8b35 | 1042 | ret = vfs_writev(f.file, vec, vlen, &pos, flags); |
5faf153e AV |
1043 | if (ret >= 0) |
1044 | file_pos_write(f.file, pos); | |
9c225f26 | 1045 | fdput_pos(f); |
1da177e4 LT |
1046 | } |
1047 | ||
1048 | if (ret > 0) | |
4b98d11b AD |
1049 | add_wchar(current, ret); |
1050 | inc_syscw(current); | |
1da177e4 LT |
1051 | return ret; |
1052 | } | |
1053 | ||
601cc11d LT |
1054 | static inline loff_t pos_from_hilo(unsigned long high, unsigned long low) |
1055 | { | |
1056 | #define HALF_LONG_BITS (BITS_PER_LONG / 2) | |
1057 | return (((loff_t)high << HALF_LONG_BITS) << HALF_LONG_BITS) | low; | |
1058 | } | |
1059 | ||
f17d8b35 MT |
1060 | static ssize_t do_preadv(unsigned long fd, const struct iovec __user *vec, |
1061 | unsigned long vlen, loff_t pos, int flags) | |
f3554f4b | 1062 | { |
2903ff01 | 1063 | struct fd f; |
f3554f4b | 1064 | ssize_t ret = -EBADF; |
f3554f4b GH |
1065 | |
1066 | if (pos < 0) | |
1067 | return -EINVAL; | |
1068 | ||
2903ff01 AV |
1069 | f = fdget(fd); |
1070 | if (f.file) { | |
f3554f4b | 1071 | ret = -ESPIPE; |
2903ff01 | 1072 | if (f.file->f_mode & FMODE_PREAD) |
f17d8b35 | 1073 | ret = vfs_readv(f.file, vec, vlen, &pos, flags); |
2903ff01 | 1074 | fdput(f); |
f3554f4b GH |
1075 | } |
1076 | ||
1077 | if (ret > 0) | |
1078 | add_rchar(current, ret); | |
1079 | inc_syscr(current); | |
1080 | return ret; | |
1081 | } | |
1082 | ||
f17d8b35 MT |
1083 | static ssize_t do_pwritev(unsigned long fd, const struct iovec __user *vec, |
1084 | unsigned long vlen, loff_t pos, int flags) | |
f3554f4b | 1085 | { |
2903ff01 | 1086 | struct fd f; |
f3554f4b | 1087 | ssize_t ret = -EBADF; |
f3554f4b GH |
1088 | |
1089 | if (pos < 0) | |
1090 | return -EINVAL; | |
1091 | ||
2903ff01 AV |
1092 | f = fdget(fd); |
1093 | if (f.file) { | |
f3554f4b | 1094 | ret = -ESPIPE; |
2903ff01 | 1095 | if (f.file->f_mode & FMODE_PWRITE) |
f17d8b35 | 1096 | ret = vfs_writev(f.file, vec, vlen, &pos, flags); |
2903ff01 | 1097 | fdput(f); |
f3554f4b GH |
1098 | } |
1099 | ||
1100 | if (ret > 0) | |
1101 | add_wchar(current, ret); | |
1102 | inc_syscw(current); | |
1103 | return ret; | |
1104 | } | |
1105 | ||
f17d8b35 MT |
1106 | SYSCALL_DEFINE3(readv, unsigned long, fd, const struct iovec __user *, vec, |
1107 | unsigned long, vlen) | |
1108 | { | |
1109 | return do_readv(fd, vec, vlen, 0); | |
1110 | } | |
1111 | ||
1112 | SYSCALL_DEFINE3(writev, unsigned long, fd, const struct iovec __user *, vec, | |
1113 | unsigned long, vlen) | |
1114 | { | |
1115 | return do_writev(fd, vec, vlen, 0); | |
1116 | } | |
1117 | ||
1118 | SYSCALL_DEFINE5(preadv, unsigned long, fd, const struct iovec __user *, vec, | |
1119 | unsigned long, vlen, unsigned long, pos_l, unsigned long, pos_h) | |
1120 | { | |
1121 | loff_t pos = pos_from_hilo(pos_h, pos_l); | |
1122 | ||
1123 | return do_preadv(fd, vec, vlen, pos, 0); | |
1124 | } | |
1125 | ||
1126 | SYSCALL_DEFINE6(preadv2, unsigned long, fd, const struct iovec __user *, vec, | |
1127 | unsigned long, vlen, unsigned long, pos_l, unsigned long, pos_h, | |
1128 | int, flags) | |
1129 | { | |
1130 | loff_t pos = pos_from_hilo(pos_h, pos_l); | |
1131 | ||
1132 | if (pos == -1) | |
1133 | return do_readv(fd, vec, vlen, flags); | |
1134 | ||
1135 | return do_preadv(fd, vec, vlen, pos, flags); | |
1136 | } | |
1137 | ||
1138 | SYSCALL_DEFINE5(pwritev, unsigned long, fd, const struct iovec __user *, vec, | |
1139 | unsigned long, vlen, unsigned long, pos_l, unsigned long, pos_h) | |
1140 | { | |
1141 | loff_t pos = pos_from_hilo(pos_h, pos_l); | |
1142 | ||
1143 | return do_pwritev(fd, vec, vlen, pos, 0); | |
1144 | } | |
1145 | ||
1146 | SYSCALL_DEFINE6(pwritev2, unsigned long, fd, const struct iovec __user *, vec, | |
1147 | unsigned long, vlen, unsigned long, pos_l, unsigned long, pos_h, | |
1148 | int, flags) | |
1149 | { | |
1150 | loff_t pos = pos_from_hilo(pos_h, pos_l); | |
1151 | ||
1152 | if (pos == -1) | |
1153 | return do_writev(fd, vec, vlen, flags); | |
1154 | ||
1155 | return do_pwritev(fd, vec, vlen, pos, flags); | |
1156 | } | |
1157 | ||
72ec3516 | 1158 | #ifdef CONFIG_COMPAT |
72ec3516 AV |
1159 | static size_t compat_readv(struct file *file, |
1160 | const struct compat_iovec __user *vec, | |
f17d8b35 | 1161 | unsigned long vlen, loff_t *pos, int flags) |
72ec3516 | 1162 | { |
72ec3516 AV |
1163 | struct iovec iovstack[UIO_FASTIOV]; |
1164 | struct iovec *iov = iovstack; | |
ac15ac06 | 1165 | struct iov_iter iter; |
72ec3516 | 1166 | ssize_t ret; |
72ec3516 | 1167 | |
26c87fb7 | 1168 | ret = compat_import_iovec(READ, vec, vlen, UIO_FASTIOV, &iov, &iter); |
edab5fe3 CH |
1169 | if (ret >= 0) { |
1170 | ret = do_iter_read(file, &iter, pos, flags); | |
1171 | kfree(iov); | |
1172 | } | |
72ec3516 AV |
1173 | if (ret > 0) |
1174 | add_rchar(current, ret); | |
1175 | inc_syscr(current); | |
1176 | return ret; | |
1177 | } | |
1178 | ||
f17d8b35 MT |
1179 | static size_t do_compat_readv(compat_ulong_t fd, |
1180 | const struct compat_iovec __user *vec, | |
1181 | compat_ulong_t vlen, int flags) | |
72ec3516 | 1182 | { |
9c225f26 | 1183 | struct fd f = fdget_pos(fd); |
72ec3516 AV |
1184 | ssize_t ret; |
1185 | loff_t pos; | |
1186 | ||
1187 | if (!f.file) | |
1188 | return -EBADF; | |
1189 | pos = f.file->f_pos; | |
f17d8b35 | 1190 | ret = compat_readv(f.file, vec, vlen, &pos, flags); |
5faf153e AV |
1191 | if (ret >= 0) |
1192 | f.file->f_pos = pos; | |
9c225f26 | 1193 | fdput_pos(f); |
72ec3516 | 1194 | return ret; |
f17d8b35 | 1195 | |
72ec3516 AV |
1196 | } |
1197 | ||
f17d8b35 MT |
1198 | COMPAT_SYSCALL_DEFINE3(readv, compat_ulong_t, fd, |
1199 | const struct compat_iovec __user *,vec, | |
1200 | compat_ulong_t, vlen) | |
1201 | { | |
1202 | return do_compat_readv(fd, vec, vlen, 0); | |
1203 | } | |
1204 | ||
1205 | static long do_compat_preadv64(unsigned long fd, | |
378a10f3 | 1206 | const struct compat_iovec __user *vec, |
f17d8b35 | 1207 | unsigned long vlen, loff_t pos, int flags) |
72ec3516 AV |
1208 | { |
1209 | struct fd f; | |
1210 | ssize_t ret; | |
1211 | ||
1212 | if (pos < 0) | |
1213 | return -EINVAL; | |
1214 | f = fdget(fd); | |
1215 | if (!f.file) | |
1216 | return -EBADF; | |
1217 | ret = -ESPIPE; | |
1218 | if (f.file->f_mode & FMODE_PREAD) | |
f17d8b35 | 1219 | ret = compat_readv(f.file, vec, vlen, &pos, flags); |
72ec3516 AV |
1220 | fdput(f); |
1221 | return ret; | |
1222 | } | |
1223 | ||
378a10f3 HC |
1224 | #ifdef __ARCH_WANT_COMPAT_SYS_PREADV64 |
1225 | COMPAT_SYSCALL_DEFINE4(preadv64, unsigned long, fd, | |
1226 | const struct compat_iovec __user *,vec, | |
1227 | unsigned long, vlen, loff_t, pos) | |
1228 | { | |
f17d8b35 | 1229 | return do_compat_preadv64(fd, vec, vlen, pos, 0); |
378a10f3 HC |
1230 | } |
1231 | #endif | |
1232 | ||
dfd948e3 | 1233 | COMPAT_SYSCALL_DEFINE5(preadv, compat_ulong_t, fd, |
72ec3516 | 1234 | const struct compat_iovec __user *,vec, |
dfd948e3 | 1235 | compat_ulong_t, vlen, u32, pos_low, u32, pos_high) |
72ec3516 AV |
1236 | { |
1237 | loff_t pos = ((loff_t)pos_high << 32) | pos_low; | |
378a10f3 | 1238 | |
f17d8b35 MT |
1239 | return do_compat_preadv64(fd, vec, vlen, pos, 0); |
1240 | } | |
1241 | ||
3ebfd81f L |
1242 | #ifdef __ARCH_WANT_COMPAT_SYS_PREADV64V2 |
1243 | COMPAT_SYSCALL_DEFINE5(preadv64v2, unsigned long, fd, | |
1244 | const struct compat_iovec __user *,vec, | |
1245 | unsigned long, vlen, loff_t, pos, int, flags) | |
1246 | { | |
1247 | return do_compat_preadv64(fd, vec, vlen, pos, flags); | |
1248 | } | |
1249 | #endif | |
1250 | ||
f17d8b35 MT |
1251 | COMPAT_SYSCALL_DEFINE6(preadv2, compat_ulong_t, fd, |
1252 | const struct compat_iovec __user *,vec, | |
1253 | compat_ulong_t, vlen, u32, pos_low, u32, pos_high, | |
1254 | int, flags) | |
1255 | { | |
1256 | loff_t pos = ((loff_t)pos_high << 32) | pos_low; | |
1257 | ||
1258 | if (pos == -1) | |
1259 | return do_compat_readv(fd, vec, vlen, flags); | |
1260 | ||
1261 | return do_compat_preadv64(fd, vec, vlen, pos, flags); | |
72ec3516 AV |
1262 | } |
1263 | ||
1264 | static size_t compat_writev(struct file *file, | |
1265 | const struct compat_iovec __user *vec, | |
f17d8b35 | 1266 | unsigned long vlen, loff_t *pos, int flags) |
72ec3516 | 1267 | { |
26c87fb7 CH |
1268 | struct iovec iovstack[UIO_FASTIOV]; |
1269 | struct iovec *iov = iovstack; | |
1270 | struct iov_iter iter; | |
edab5fe3 | 1271 | ssize_t ret; |
72ec3516 | 1272 | |
26c87fb7 | 1273 | ret = compat_import_iovec(WRITE, vec, vlen, UIO_FASTIOV, &iov, &iter); |
edab5fe3 | 1274 | if (ret >= 0) { |
62473a2d | 1275 | file_start_write(file); |
edab5fe3 | 1276 | ret = do_iter_write(file, &iter, pos, flags); |
62473a2d | 1277 | file_end_write(file); |
edab5fe3 CH |
1278 | kfree(iov); |
1279 | } | |
72ec3516 AV |
1280 | if (ret > 0) |
1281 | add_wchar(current, ret); | |
1282 | inc_syscw(current); | |
1283 | return ret; | |
1284 | } | |
1285 | ||
f17d8b35 MT |
1286 | static size_t do_compat_writev(compat_ulong_t fd, |
1287 | const struct compat_iovec __user* vec, | |
1288 | compat_ulong_t vlen, int flags) | |
72ec3516 | 1289 | { |
9c225f26 | 1290 | struct fd f = fdget_pos(fd); |
72ec3516 AV |
1291 | ssize_t ret; |
1292 | loff_t pos; | |
1293 | ||
1294 | if (!f.file) | |
1295 | return -EBADF; | |
1296 | pos = f.file->f_pos; | |
f17d8b35 | 1297 | ret = compat_writev(f.file, vec, vlen, &pos, flags); |
5faf153e AV |
1298 | if (ret >= 0) |
1299 | f.file->f_pos = pos; | |
9c225f26 | 1300 | fdput_pos(f); |
72ec3516 AV |
1301 | return ret; |
1302 | } | |
1303 | ||
f17d8b35 MT |
1304 | COMPAT_SYSCALL_DEFINE3(writev, compat_ulong_t, fd, |
1305 | const struct compat_iovec __user *, vec, | |
1306 | compat_ulong_t, vlen) | |
1307 | { | |
1308 | return do_compat_writev(fd, vec, vlen, 0); | |
1309 | } | |
1310 | ||
1311 | static long do_compat_pwritev64(unsigned long fd, | |
378a10f3 | 1312 | const struct compat_iovec __user *vec, |
f17d8b35 | 1313 | unsigned long vlen, loff_t pos, int flags) |
72ec3516 AV |
1314 | { |
1315 | struct fd f; | |
1316 | ssize_t ret; | |
1317 | ||
1318 | if (pos < 0) | |
1319 | return -EINVAL; | |
1320 | f = fdget(fd); | |
1321 | if (!f.file) | |
1322 | return -EBADF; | |
1323 | ret = -ESPIPE; | |
1324 | if (f.file->f_mode & FMODE_PWRITE) | |
f17d8b35 | 1325 | ret = compat_writev(f.file, vec, vlen, &pos, flags); |
72ec3516 AV |
1326 | fdput(f); |
1327 | return ret; | |
1328 | } | |
1329 | ||
378a10f3 HC |
1330 | #ifdef __ARCH_WANT_COMPAT_SYS_PWRITEV64 |
1331 | COMPAT_SYSCALL_DEFINE4(pwritev64, unsigned long, fd, | |
1332 | const struct compat_iovec __user *,vec, | |
1333 | unsigned long, vlen, loff_t, pos) | |
1334 | { | |
f17d8b35 | 1335 | return do_compat_pwritev64(fd, vec, vlen, pos, 0); |
378a10f3 HC |
1336 | } |
1337 | #endif | |
1338 | ||
dfd948e3 | 1339 | COMPAT_SYSCALL_DEFINE5(pwritev, compat_ulong_t, fd, |
72ec3516 | 1340 | const struct compat_iovec __user *,vec, |
dfd948e3 | 1341 | compat_ulong_t, vlen, u32, pos_low, u32, pos_high) |
72ec3516 AV |
1342 | { |
1343 | loff_t pos = ((loff_t)pos_high << 32) | pos_low; | |
378a10f3 | 1344 | |
f17d8b35 | 1345 | return do_compat_pwritev64(fd, vec, vlen, pos, 0); |
72ec3516 | 1346 | } |
f17d8b35 | 1347 | |
3ebfd81f L |
1348 | #ifdef __ARCH_WANT_COMPAT_SYS_PWRITEV64V2 |
1349 | COMPAT_SYSCALL_DEFINE5(pwritev64v2, unsigned long, fd, | |
1350 | const struct compat_iovec __user *,vec, | |
1351 | unsigned long, vlen, loff_t, pos, int, flags) | |
1352 | { | |
1353 | return do_compat_pwritev64(fd, vec, vlen, pos, flags); | |
1354 | } | |
1355 | #endif | |
1356 | ||
f17d8b35 MT |
1357 | COMPAT_SYSCALL_DEFINE6(pwritev2, compat_ulong_t, fd, |
1358 | const struct compat_iovec __user *,vec, | |
1359 | compat_ulong_t, vlen, u32, pos_low, u32, pos_high, int, flags) | |
1360 | { | |
1361 | loff_t pos = ((loff_t)pos_high << 32) | pos_low; | |
1362 | ||
1363 | if (pos == -1) | |
1364 | return do_compat_writev(fd, vec, vlen, flags); | |
1365 | ||
1366 | return do_compat_pwritev64(fd, vec, vlen, pos, flags); | |
72ec3516 | 1367 | } |
f17d8b35 | 1368 | |
72ec3516 AV |
1369 | #endif |
1370 | ||
19f4fc3a AV |
1371 | static ssize_t do_sendfile(int out_fd, int in_fd, loff_t *ppos, |
1372 | size_t count, loff_t max) | |
1da177e4 | 1373 | { |
2903ff01 AV |
1374 | struct fd in, out; |
1375 | struct inode *in_inode, *out_inode; | |
1da177e4 | 1376 | loff_t pos; |
7995bd28 | 1377 | loff_t out_pos; |
1da177e4 | 1378 | ssize_t retval; |
2903ff01 | 1379 | int fl; |
1da177e4 LT |
1380 | |
1381 | /* | |
1382 | * Get input file, and verify that it is ok.. | |
1383 | */ | |
1384 | retval = -EBADF; | |
2903ff01 AV |
1385 | in = fdget(in_fd); |
1386 | if (!in.file) | |
1da177e4 | 1387 | goto out; |
2903ff01 | 1388 | if (!(in.file->f_mode & FMODE_READ)) |
1da177e4 | 1389 | goto fput_in; |
1da177e4 | 1390 | retval = -ESPIPE; |
7995bd28 AV |
1391 | if (!ppos) { |
1392 | pos = in.file->f_pos; | |
1393 | } else { | |
1394 | pos = *ppos; | |
2903ff01 | 1395 | if (!(in.file->f_mode & FMODE_PREAD)) |
1da177e4 | 1396 | goto fput_in; |
7995bd28 AV |
1397 | } |
1398 | retval = rw_verify_area(READ, in.file, &pos, count); | |
e28cc715 | 1399 | if (retval < 0) |
1da177e4 | 1400 | goto fput_in; |
bc61384d AV |
1401 | if (count > MAX_RW_COUNT) |
1402 | count = MAX_RW_COUNT; | |
1da177e4 | 1403 | |
1da177e4 LT |
1404 | /* |
1405 | * Get output file, and verify that it is ok.. | |
1406 | */ | |
1407 | retval = -EBADF; | |
2903ff01 AV |
1408 | out = fdget(out_fd); |
1409 | if (!out.file) | |
1da177e4 | 1410 | goto fput_in; |
2903ff01 | 1411 | if (!(out.file->f_mode & FMODE_WRITE)) |
1da177e4 LT |
1412 | goto fput_out; |
1413 | retval = -EINVAL; | |
496ad9aa AV |
1414 | in_inode = file_inode(in.file); |
1415 | out_inode = file_inode(out.file); | |
7995bd28 AV |
1416 | out_pos = out.file->f_pos; |
1417 | retval = rw_verify_area(WRITE, out.file, &out_pos, count); | |
e28cc715 | 1418 | if (retval < 0) |
1da177e4 LT |
1419 | goto fput_out; |
1420 | ||
1da177e4 LT |
1421 | if (!max) |
1422 | max = min(in_inode->i_sb->s_maxbytes, out_inode->i_sb->s_maxbytes); | |
1423 | ||
1da177e4 LT |
1424 | if (unlikely(pos + count > max)) { |
1425 | retval = -EOVERFLOW; | |
1426 | if (pos >= max) | |
1427 | goto fput_out; | |
1428 | count = max - pos; | |
1429 | } | |
1430 | ||
d96e6e71 | 1431 | fl = 0; |
534f2aaa | 1432 | #if 0 |
d96e6e71 JA |
1433 | /* |
1434 | * We need to debate whether we can enable this or not. The | |
1435 | * man page documents EAGAIN return for the output at least, | |
1436 | * and the application is arguably buggy if it doesn't expect | |
1437 | * EAGAIN on a non-blocking file descriptor. | |
1438 | */ | |
2903ff01 | 1439 | if (in.file->f_flags & O_NONBLOCK) |
d96e6e71 | 1440 | fl = SPLICE_F_NONBLOCK; |
534f2aaa | 1441 | #endif |
50cd2c57 | 1442 | file_start_write(out.file); |
7995bd28 | 1443 | retval = do_splice_direct(in.file, &pos, out.file, &out_pos, count, fl); |
50cd2c57 | 1444 | file_end_write(out.file); |
1da177e4 LT |
1445 | |
1446 | if (retval > 0) { | |
4b98d11b AD |
1447 | add_rchar(current, retval); |
1448 | add_wchar(current, retval); | |
a68c2f12 SW |
1449 | fsnotify_access(in.file); |
1450 | fsnotify_modify(out.file); | |
7995bd28 AV |
1451 | out.file->f_pos = out_pos; |
1452 | if (ppos) | |
1453 | *ppos = pos; | |
1454 | else | |
1455 | in.file->f_pos = pos; | |
1da177e4 | 1456 | } |
1da177e4 | 1457 | |
4b98d11b AD |
1458 | inc_syscr(current); |
1459 | inc_syscw(current); | |
7995bd28 | 1460 | if (pos > max) |
1da177e4 LT |
1461 | retval = -EOVERFLOW; |
1462 | ||
1463 | fput_out: | |
2903ff01 | 1464 | fdput(out); |
1da177e4 | 1465 | fput_in: |
2903ff01 | 1466 | fdput(in); |
1da177e4 LT |
1467 | out: |
1468 | return retval; | |
1469 | } | |
1470 | ||
002c8976 | 1471 | SYSCALL_DEFINE4(sendfile, int, out_fd, int, in_fd, off_t __user *, offset, size_t, count) |
1da177e4 LT |
1472 | { |
1473 | loff_t pos; | |
1474 | off_t off; | |
1475 | ssize_t ret; | |
1476 | ||
1477 | if (offset) { | |
1478 | if (unlikely(get_user(off, offset))) | |
1479 | return -EFAULT; | |
1480 | pos = off; | |
1481 | ret = do_sendfile(out_fd, in_fd, &pos, count, MAX_NON_LFS); | |
1482 | if (unlikely(put_user(pos, offset))) | |
1483 | return -EFAULT; | |
1484 | return ret; | |
1485 | } | |
1486 | ||
1487 | return do_sendfile(out_fd, in_fd, NULL, count, 0); | |
1488 | } | |
1489 | ||
002c8976 | 1490 | SYSCALL_DEFINE4(sendfile64, int, out_fd, int, in_fd, loff_t __user *, offset, size_t, count) |
1da177e4 LT |
1491 | { |
1492 | loff_t pos; | |
1493 | ssize_t ret; | |
1494 | ||
1495 | if (offset) { | |
1496 | if (unlikely(copy_from_user(&pos, offset, sizeof(loff_t)))) | |
1497 | return -EFAULT; | |
1498 | ret = do_sendfile(out_fd, in_fd, &pos, count, 0); | |
1499 | if (unlikely(put_user(pos, offset))) | |
1500 | return -EFAULT; | |
1501 | return ret; | |
1502 | } | |
1503 | ||
1504 | return do_sendfile(out_fd, in_fd, NULL, count, 0); | |
1505 | } | |
19f4fc3a AV |
1506 | |
1507 | #ifdef CONFIG_COMPAT | |
1508 | COMPAT_SYSCALL_DEFINE4(sendfile, int, out_fd, int, in_fd, | |
1509 | compat_off_t __user *, offset, compat_size_t, count) | |
1510 | { | |
1511 | loff_t pos; | |
1512 | off_t off; | |
1513 | ssize_t ret; | |
1514 | ||
1515 | if (offset) { | |
1516 | if (unlikely(get_user(off, offset))) | |
1517 | return -EFAULT; | |
1518 | pos = off; | |
1519 | ret = do_sendfile(out_fd, in_fd, &pos, count, MAX_NON_LFS); | |
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 | ||
1528 | COMPAT_SYSCALL_DEFINE4(sendfile64, int, out_fd, int, in_fd, | |
1529 | compat_loff_t __user *, offset, compat_size_t, count) | |
1530 | { | |
1531 | loff_t pos; | |
1532 | ssize_t ret; | |
1533 | ||
1534 | if (offset) { | |
1535 | if (unlikely(copy_from_user(&pos, offset, sizeof(loff_t)))) | |
1536 | return -EFAULT; | |
1537 | ret = do_sendfile(out_fd, in_fd, &pos, count, 0); | |
1538 | if (unlikely(put_user(pos, offset))) | |
1539 | return -EFAULT; | |
1540 | return ret; | |
1541 | } | |
1542 | ||
1543 | return do_sendfile(out_fd, in_fd, NULL, count, 0); | |
1544 | } | |
1545 | #endif | |
29732938 ZB |
1546 | |
1547 | /* | |
1548 | * copy_file_range() differs from regular file read and write in that it | |
1549 | * specifically allows return partial success. When it does so is up to | |
1550 | * the copy_file_range method. | |
1551 | */ | |
1552 | ssize_t vfs_copy_file_range(struct file *file_in, loff_t pos_in, | |
1553 | struct file *file_out, loff_t pos_out, | |
1554 | size_t len, unsigned int flags) | |
1555 | { | |
1556 | struct inode *inode_in = file_inode(file_in); | |
1557 | struct inode *inode_out = file_inode(file_out); | |
1558 | ssize_t ret; | |
1559 | ||
1560 | if (flags != 0) | |
1561 | return -EINVAL; | |
1562 | ||
11cbfb10 AG |
1563 | if (S_ISDIR(inode_in->i_mode) || S_ISDIR(inode_out->i_mode)) |
1564 | return -EISDIR; | |
1565 | if (!S_ISREG(inode_in->i_mode) || !S_ISREG(inode_out->i_mode)) | |
1566 | return -EINVAL; | |
1567 | ||
29732938 | 1568 | ret = rw_verify_area(READ, file_in, &pos_in, len); |
bc61384d AV |
1569 | if (unlikely(ret)) |
1570 | return ret; | |
1571 | ||
1572 | ret = rw_verify_area(WRITE, file_out, &pos_out, len); | |
1573 | if (unlikely(ret)) | |
29732938 ZB |
1574 | return ret; |
1575 | ||
1576 | if (!(file_in->f_mode & FMODE_READ) || | |
1577 | !(file_out->f_mode & FMODE_WRITE) || | |
eac70053 | 1578 | (file_out->f_flags & O_APPEND)) |
29732938 ZB |
1579 | return -EBADF; |
1580 | ||
1581 | /* this could be relaxed once a method supports cross-fs copies */ | |
1582 | if (inode_in->i_sb != inode_out->i_sb) | |
1583 | return -EXDEV; | |
1584 | ||
1585 | if (len == 0) | |
1586 | return 0; | |
1587 | ||
bfe219d3 | 1588 | file_start_write(file_out); |
29732938 | 1589 | |
a76b5b04 CH |
1590 | /* |
1591 | * Try cloning first, this is supported by more file systems, and | |
1592 | * more efficient if both clone and copy are supported (e.g. NFS). | |
1593 | */ | |
1594 | if (file_in->f_op->clone_file_range) { | |
1595 | ret = file_in->f_op->clone_file_range(file_in, pos_in, | |
1596 | file_out, pos_out, len); | |
1597 | if (ret == 0) { | |
1598 | ret = len; | |
1599 | goto done; | |
1600 | } | |
1601 | } | |
1602 | ||
1603 | if (file_out->f_op->copy_file_range) { | |
eac70053 AS |
1604 | ret = file_out->f_op->copy_file_range(file_in, pos_in, file_out, |
1605 | pos_out, len, flags); | |
a76b5b04 CH |
1606 | if (ret != -EOPNOTSUPP) |
1607 | goto done; | |
1608 | } | |
eac70053 | 1609 | |
a76b5b04 CH |
1610 | ret = do_splice_direct(file_in, &pos_in, file_out, &pos_out, |
1611 | len > MAX_RW_COUNT ? MAX_RW_COUNT : len, 0); | |
eac70053 | 1612 | |
a76b5b04 | 1613 | done: |
29732938 ZB |
1614 | if (ret > 0) { |
1615 | fsnotify_access(file_in); | |
1616 | add_rchar(current, ret); | |
1617 | fsnotify_modify(file_out); | |
1618 | add_wchar(current, ret); | |
1619 | } | |
a76b5b04 | 1620 | |
29732938 ZB |
1621 | inc_syscr(current); |
1622 | inc_syscw(current); | |
1623 | ||
bfe219d3 | 1624 | file_end_write(file_out); |
29732938 ZB |
1625 | |
1626 | return ret; | |
1627 | } | |
1628 | EXPORT_SYMBOL(vfs_copy_file_range); | |
1629 | ||
1630 | SYSCALL_DEFINE6(copy_file_range, int, fd_in, loff_t __user *, off_in, | |
1631 | int, fd_out, loff_t __user *, off_out, | |
1632 | size_t, len, unsigned int, flags) | |
1633 | { | |
1634 | loff_t pos_in; | |
1635 | loff_t pos_out; | |
1636 | struct fd f_in; | |
1637 | struct fd f_out; | |
1638 | ssize_t ret = -EBADF; | |
1639 | ||
1640 | f_in = fdget(fd_in); | |
1641 | if (!f_in.file) | |
1642 | goto out2; | |
1643 | ||
1644 | f_out = fdget(fd_out); | |
1645 | if (!f_out.file) | |
1646 | goto out1; | |
1647 | ||
1648 | ret = -EFAULT; | |
1649 | if (off_in) { | |
1650 | if (copy_from_user(&pos_in, off_in, sizeof(loff_t))) | |
1651 | goto out; | |
1652 | } else { | |
1653 | pos_in = f_in.file->f_pos; | |
1654 | } | |
1655 | ||
1656 | if (off_out) { | |
1657 | if (copy_from_user(&pos_out, off_out, sizeof(loff_t))) | |
1658 | goto out; | |
1659 | } else { | |
1660 | pos_out = f_out.file->f_pos; | |
1661 | } | |
1662 | ||
1663 | ret = vfs_copy_file_range(f_in.file, pos_in, f_out.file, pos_out, len, | |
1664 | flags); | |
1665 | if (ret > 0) { | |
1666 | pos_in += ret; | |
1667 | pos_out += ret; | |
1668 | ||
1669 | if (off_in) { | |
1670 | if (copy_to_user(off_in, &pos_in, sizeof(loff_t))) | |
1671 | ret = -EFAULT; | |
1672 | } else { | |
1673 | f_in.file->f_pos = pos_in; | |
1674 | } | |
1675 | ||
1676 | if (off_out) { | |
1677 | if (copy_to_user(off_out, &pos_out, sizeof(loff_t))) | |
1678 | ret = -EFAULT; | |
1679 | } else { | |
1680 | f_out.file->f_pos = pos_out; | |
1681 | } | |
1682 | } | |
1683 | ||
1684 | out: | |
1685 | fdput(f_out); | |
1686 | out1: | |
1687 | fdput(f_in); | |
1688 | out2: | |
1689 | return ret; | |
1690 | } | |
04b38d60 CH |
1691 | |
1692 | static int clone_verify_area(struct file *file, loff_t pos, u64 len, bool write) | |
1693 | { | |
1694 | struct inode *inode = file_inode(file); | |
1695 | ||
1696 | if (unlikely(pos < 0)) | |
1697 | return -EINVAL; | |
1698 | ||
1699 | if (unlikely((loff_t) (pos + len) < 0)) | |
1700 | return -EINVAL; | |
1701 | ||
1702 | if (unlikely(inode->i_flctx && mandatory_lock(inode))) { | |
1703 | loff_t end = len ? pos + len - 1 : OFFSET_MAX; | |
1704 | int retval; | |
1705 | ||
1706 | retval = locks_mandatory_area(inode, file, pos, end, | |
1707 | write ? F_WRLCK : F_RDLCK); | |
1708 | if (retval < 0) | |
1709 | return retval; | |
1710 | } | |
1711 | ||
1712 | return security_file_permission(file, write ? MAY_WRITE : MAY_READ); | |
1713 | } | |
1714 | ||
876bec6f DW |
1715 | /* |
1716 | * Check that the two inodes are eligible for cloning, the ranges make | |
1717 | * sense, and then flush all dirty data. Caller must ensure that the | |
1718 | * inodes have been locked against any other modifications. | |
22725ce4 DW |
1719 | * |
1720 | * Returns: 0 for "nothing to clone", 1 for "something to clone", or | |
1721 | * the usual negative error code. | |
876bec6f DW |
1722 | */ |
1723 | int vfs_clone_file_prep_inodes(struct inode *inode_in, loff_t pos_in, | |
1724 | struct inode *inode_out, loff_t pos_out, | |
1725 | u64 *len, bool is_dedupe) | |
1726 | { | |
1727 | loff_t bs = inode_out->i_sb->s_blocksize; | |
1728 | loff_t blen; | |
1729 | loff_t isize; | |
1730 | bool same_inode = (inode_in == inode_out); | |
1731 | int ret; | |
1732 | ||
1733 | /* Don't touch certain kinds of inodes */ | |
1734 | if (IS_IMMUTABLE(inode_out)) | |
1735 | return -EPERM; | |
1736 | ||
1737 | if (IS_SWAPFILE(inode_in) || IS_SWAPFILE(inode_out)) | |
1738 | return -ETXTBSY; | |
1739 | ||
1740 | /* Don't reflink dirs, pipes, sockets... */ | |
1741 | if (S_ISDIR(inode_in->i_mode) || S_ISDIR(inode_out->i_mode)) | |
1742 | return -EISDIR; | |
1743 | if (!S_ISREG(inode_in->i_mode) || !S_ISREG(inode_out->i_mode)) | |
1744 | return -EINVAL; | |
1745 | ||
1746 | /* Are we going all the way to the end? */ | |
1747 | isize = i_size_read(inode_in); | |
22725ce4 | 1748 | if (isize == 0) |
876bec6f | 1749 | return 0; |
876bec6f DW |
1750 | |
1751 | /* Zero length dedupe exits immediately; reflink goes to EOF. */ | |
1752 | if (*len == 0) { | |
22725ce4 | 1753 | if (is_dedupe || pos_in == isize) |
876bec6f | 1754 | return 0; |
22725ce4 DW |
1755 | if (pos_in > isize) |
1756 | return -EINVAL; | |
876bec6f DW |
1757 | *len = isize - pos_in; |
1758 | } | |
1759 | ||
1760 | /* Ensure offsets don't wrap and the input is inside i_size */ | |
1761 | if (pos_in + *len < pos_in || pos_out + *len < pos_out || | |
1762 | pos_in + *len > isize) | |
1763 | return -EINVAL; | |
1764 | ||
1765 | /* Don't allow dedupe past EOF in the dest file */ | |
1766 | if (is_dedupe) { | |
1767 | loff_t disize; | |
1768 | ||
1769 | disize = i_size_read(inode_out); | |
1770 | if (pos_out >= disize || pos_out + *len > disize) | |
1771 | return -EINVAL; | |
1772 | } | |
1773 | ||
1774 | /* If we're linking to EOF, continue to the block boundary. */ | |
1775 | if (pos_in + *len == isize) | |
1776 | blen = ALIGN(isize, bs) - pos_in; | |
1777 | else | |
1778 | blen = *len; | |
1779 | ||
1780 | /* Only reflink if we're aligned to block boundaries */ | |
1781 | if (!IS_ALIGNED(pos_in, bs) || !IS_ALIGNED(pos_in + blen, bs) || | |
1782 | !IS_ALIGNED(pos_out, bs) || !IS_ALIGNED(pos_out + blen, bs)) | |
1783 | return -EINVAL; | |
1784 | ||
1785 | /* Don't allow overlapped reflink within the same file */ | |
1786 | if (same_inode) { | |
1787 | if (pos_out + blen > pos_in && pos_out < pos_in + blen) | |
1788 | return -EINVAL; | |
1789 | } | |
1790 | ||
1791 | /* Wait for the completion of any pending IOs on both files */ | |
1792 | inode_dio_wait(inode_in); | |
1793 | if (!same_inode) | |
1794 | inode_dio_wait(inode_out); | |
1795 | ||
1796 | ret = filemap_write_and_wait_range(inode_in->i_mapping, | |
1797 | pos_in, pos_in + *len - 1); | |
1798 | if (ret) | |
1799 | return ret; | |
1800 | ||
1801 | ret = filemap_write_and_wait_range(inode_out->i_mapping, | |
1802 | pos_out, pos_out + *len - 1); | |
1803 | if (ret) | |
1804 | return ret; | |
1805 | ||
1806 | /* | |
1807 | * Check that the extents are the same. | |
1808 | */ | |
1809 | if (is_dedupe) { | |
1810 | bool is_same = false; | |
1811 | ||
1812 | ret = vfs_dedupe_file_range_compare(inode_in, pos_in, | |
1813 | inode_out, pos_out, *len, &is_same); | |
1814 | if (ret) | |
1815 | return ret; | |
1816 | if (!is_same) | |
1817 | return -EBADE; | |
1818 | } | |
1819 | ||
22725ce4 | 1820 | return 1; |
876bec6f DW |
1821 | } |
1822 | EXPORT_SYMBOL(vfs_clone_file_prep_inodes); | |
1823 | ||
04b38d60 CH |
1824 | int vfs_clone_file_range(struct file *file_in, loff_t pos_in, |
1825 | struct file *file_out, loff_t pos_out, u64 len) | |
1826 | { | |
1827 | struct inode *inode_in = file_inode(file_in); | |
1828 | struct inode *inode_out = file_inode(file_out); | |
1829 | int ret; | |
1830 | ||
b335e9d9 AG |
1831 | if (S_ISDIR(inode_in->i_mode) || S_ISDIR(inode_out->i_mode)) |
1832 | return -EISDIR; | |
1833 | if (!S_ISREG(inode_in->i_mode) || !S_ISREG(inode_out->i_mode)) | |
1834 | return -EINVAL; | |
1835 | ||
913b86e9 AG |
1836 | /* |
1837 | * FICLONE/FICLONERANGE ioctls enforce that src and dest files are on | |
1838 | * the same mount. Practically, they only need to be on the same file | |
1839 | * system. | |
1840 | */ | |
1841 | if (inode_in->i_sb != inode_out->i_sb) | |
04b38d60 CH |
1842 | return -EXDEV; |
1843 | ||
04b38d60 CH |
1844 | if (!(file_in->f_mode & FMODE_READ) || |
1845 | !(file_out->f_mode & FMODE_WRITE) || | |
0fcbf996 | 1846 | (file_out->f_flags & O_APPEND)) |
04b38d60 CH |
1847 | return -EBADF; |
1848 | ||
0fcbf996 CH |
1849 | if (!file_in->f_op->clone_file_range) |
1850 | return -EOPNOTSUPP; | |
1851 | ||
04b38d60 CH |
1852 | ret = clone_verify_area(file_in, pos_in, len, false); |
1853 | if (ret) | |
1854 | return ret; | |
1855 | ||
1856 | ret = clone_verify_area(file_out, pos_out, len, true); | |
1857 | if (ret) | |
1858 | return ret; | |
1859 | ||
1860 | if (pos_in + len > i_size_read(inode_in)) | |
1861 | return -EINVAL; | |
1862 | ||
04b38d60 CH |
1863 | ret = file_in->f_op->clone_file_range(file_in, pos_in, |
1864 | file_out, pos_out, len); | |
1865 | if (!ret) { | |
1866 | fsnotify_access(file_in); | |
1867 | fsnotify_modify(file_out); | |
1868 | } | |
1869 | ||
04b38d60 CH |
1870 | return ret; |
1871 | } | |
1872 | EXPORT_SYMBOL(vfs_clone_file_range); | |
54dbc151 | 1873 | |
876bec6f DW |
1874 | /* |
1875 | * Read a page's worth of file data into the page cache. Return the page | |
1876 | * locked. | |
1877 | */ | |
1878 | static struct page *vfs_dedupe_get_page(struct inode *inode, loff_t offset) | |
1879 | { | |
1880 | struct address_space *mapping; | |
1881 | struct page *page; | |
1882 | pgoff_t n; | |
1883 | ||
1884 | n = offset >> PAGE_SHIFT; | |
1885 | mapping = inode->i_mapping; | |
1886 | page = read_mapping_page(mapping, n, NULL); | |
1887 | if (IS_ERR(page)) | |
1888 | return page; | |
1889 | if (!PageUptodate(page)) { | |
1890 | put_page(page); | |
1891 | return ERR_PTR(-EIO); | |
1892 | } | |
1893 | lock_page(page); | |
1894 | return page; | |
1895 | } | |
1896 | ||
1897 | /* | |
1898 | * Compare extents of two files to see if they are the same. | |
1899 | * Caller must have locked both inodes to prevent write races. | |
1900 | */ | |
1901 | int vfs_dedupe_file_range_compare(struct inode *src, loff_t srcoff, | |
1902 | struct inode *dest, loff_t destoff, | |
1903 | loff_t len, bool *is_same) | |
1904 | { | |
1905 | loff_t src_poff; | |
1906 | loff_t dest_poff; | |
1907 | void *src_addr; | |
1908 | void *dest_addr; | |
1909 | struct page *src_page; | |
1910 | struct page *dest_page; | |
1911 | loff_t cmp_len; | |
1912 | bool same; | |
1913 | int error; | |
1914 | ||
1915 | error = -EINVAL; | |
1916 | same = true; | |
1917 | while (len) { | |
1918 | src_poff = srcoff & (PAGE_SIZE - 1); | |
1919 | dest_poff = destoff & (PAGE_SIZE - 1); | |
1920 | cmp_len = min(PAGE_SIZE - src_poff, | |
1921 | PAGE_SIZE - dest_poff); | |
1922 | cmp_len = min(cmp_len, len); | |
1923 | if (cmp_len <= 0) | |
1924 | goto out_error; | |
1925 | ||
1926 | src_page = vfs_dedupe_get_page(src, srcoff); | |
1927 | if (IS_ERR(src_page)) { | |
1928 | error = PTR_ERR(src_page); | |
1929 | goto out_error; | |
1930 | } | |
1931 | dest_page = vfs_dedupe_get_page(dest, destoff); | |
1932 | if (IS_ERR(dest_page)) { | |
1933 | error = PTR_ERR(dest_page); | |
1934 | unlock_page(src_page); | |
1935 | put_page(src_page); | |
1936 | goto out_error; | |
1937 | } | |
1938 | src_addr = kmap_atomic(src_page); | |
1939 | dest_addr = kmap_atomic(dest_page); | |
1940 | ||
1941 | flush_dcache_page(src_page); | |
1942 | flush_dcache_page(dest_page); | |
1943 | ||
1944 | if (memcmp(src_addr + src_poff, dest_addr + dest_poff, cmp_len)) | |
1945 | same = false; | |
1946 | ||
1947 | kunmap_atomic(dest_addr); | |
1948 | kunmap_atomic(src_addr); | |
1949 | unlock_page(dest_page); | |
1950 | unlock_page(src_page); | |
1951 | put_page(dest_page); | |
1952 | put_page(src_page); | |
1953 | ||
1954 | if (!same) | |
1955 | break; | |
1956 | ||
1957 | srcoff += cmp_len; | |
1958 | destoff += cmp_len; | |
1959 | len -= cmp_len; | |
1960 | } | |
1961 | ||
1962 | *is_same = same; | |
1963 | return 0; | |
1964 | ||
1965 | out_error: | |
1966 | return error; | |
1967 | } | |
1968 | EXPORT_SYMBOL(vfs_dedupe_file_range_compare); | |
1969 | ||
54dbc151 DW |
1970 | int vfs_dedupe_file_range(struct file *file, struct file_dedupe_range *same) |
1971 | { | |
1972 | struct file_dedupe_range_info *info; | |
1973 | struct inode *src = file_inode(file); | |
1974 | u64 off; | |
1975 | u64 len; | |
1976 | int i; | |
1977 | int ret; | |
1978 | bool is_admin = capable(CAP_SYS_ADMIN); | |
1979 | u16 count = same->dest_count; | |
1980 | struct file *dst_file; | |
1981 | loff_t dst_off; | |
1982 | ssize_t deduped; | |
1983 | ||
1984 | if (!(file->f_mode & FMODE_READ)) | |
1985 | return -EINVAL; | |
1986 | ||
1987 | if (same->reserved1 || same->reserved2) | |
1988 | return -EINVAL; | |
1989 | ||
1990 | off = same->src_offset; | |
1991 | len = same->src_length; | |
1992 | ||
1993 | ret = -EISDIR; | |
1994 | if (S_ISDIR(src->i_mode)) | |
1995 | goto out; | |
1996 | ||
1997 | ret = -EINVAL; | |
1998 | if (!S_ISREG(src->i_mode)) | |
1999 | goto out; | |
2000 | ||
2001 | ret = clone_verify_area(file, off, len, false); | |
2002 | if (ret < 0) | |
2003 | goto out; | |
2004 | ret = 0; | |
2005 | ||
22725ce4 DW |
2006 | if (off + len > i_size_read(src)) |
2007 | return -EINVAL; | |
2008 | ||
54dbc151 DW |
2009 | /* pre-format output fields to sane values */ |
2010 | for (i = 0; i < count; i++) { | |
2011 | same->info[i].bytes_deduped = 0ULL; | |
2012 | same->info[i].status = FILE_DEDUPE_RANGE_SAME; | |
2013 | } | |
2014 | ||
2015 | for (i = 0, info = same->info; i < count; i++, info++) { | |
2016 | struct inode *dst; | |
2017 | struct fd dst_fd = fdget(info->dest_fd); | |
2018 | ||
2019 | dst_file = dst_fd.file; | |
2020 | if (!dst_file) { | |
2021 | info->status = -EBADF; | |
2022 | goto next_loop; | |
2023 | } | |
2024 | dst = file_inode(dst_file); | |
2025 | ||
2026 | ret = mnt_want_write_file(dst_file); | |
2027 | if (ret) { | |
2028 | info->status = ret; | |
2029 | goto next_loop; | |
2030 | } | |
2031 | ||
2032 | dst_off = info->dest_offset; | |
2033 | ret = clone_verify_area(dst_file, dst_off, len, true); | |
2034 | if (ret < 0) { | |
2035 | info->status = ret; | |
2036 | goto next_file; | |
2037 | } | |
2038 | ret = 0; | |
2039 | ||
2040 | if (info->reserved) { | |
2041 | info->status = -EINVAL; | |
2042 | } else if (!(is_admin || (dst_file->f_mode & FMODE_WRITE))) { | |
2043 | info->status = -EINVAL; | |
2044 | } else if (file->f_path.mnt != dst_file->f_path.mnt) { | |
2045 | info->status = -EXDEV; | |
2046 | } else if (S_ISDIR(dst->i_mode)) { | |
2047 | info->status = -EISDIR; | |
2048 | } else if (dst_file->f_op->dedupe_file_range == NULL) { | |
2049 | info->status = -EINVAL; | |
2050 | } else { | |
2051 | deduped = dst_file->f_op->dedupe_file_range(file, off, | |
2052 | len, dst_file, | |
2053 | info->dest_offset); | |
2054 | if (deduped == -EBADE) | |
2055 | info->status = FILE_DEDUPE_RANGE_DIFFERS; | |
2056 | else if (deduped < 0) | |
2057 | info->status = deduped; | |
2058 | else | |
2059 | info->bytes_deduped += deduped; | |
2060 | } | |
2061 | ||
2062 | next_file: | |
2063 | mnt_drop_write_file(dst_file); | |
2064 | next_loop: | |
2065 | fdput(dst_fd); | |
e62e560f DW |
2066 | |
2067 | if (fatal_signal_pending(current)) | |
2068 | goto out; | |
54dbc151 DW |
2069 | } |
2070 | ||
2071 | out: | |
2072 | return ret; | |
2073 | } | |
2074 | EXPORT_SYMBOL(vfs_dedupe_file_range); |