]>
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 | |
bdd1d2d3 | 418 | ssize_t kernel_read(struct file *file, void *buf, size_t count, loff_t *pos) |
c41fbad0 CH |
419 | { |
420 | mm_segment_t old_fs; | |
bdd1d2d3 | 421 | ssize_t result; |
c41fbad0 CH |
422 | |
423 | old_fs = get_fs(); | |
424 | set_fs(get_ds()); | |
425 | /* The cast to a user pointer is valid due to the set_fs() */ | |
bdd1d2d3 | 426 | result = vfs_read(file, (void __user *)buf, count, pos); |
c41fbad0 CH |
427 | set_fs(old_fs); |
428 | return result; | |
429 | } | |
430 | EXPORT_SYMBOL(kernel_read); | |
431 | ||
1da177e4 LT |
432 | ssize_t vfs_read(struct file *file, char __user *buf, size_t count, loff_t *pos) |
433 | { | |
434 | ssize_t ret; | |
435 | ||
436 | if (!(file->f_mode & FMODE_READ)) | |
437 | return -EBADF; | |
7f7f25e8 | 438 | if (!(file->f_mode & FMODE_CAN_READ)) |
1da177e4 LT |
439 | return -EINVAL; |
440 | if (unlikely(!access_ok(VERIFY_WRITE, buf, count))) | |
441 | return -EFAULT; | |
442 | ||
443 | ret = rw_verify_area(READ, file, pos, count); | |
bc61384d AV |
444 | if (!ret) { |
445 | if (count > MAX_RW_COUNT) | |
446 | count = MAX_RW_COUNT; | |
6fb5032e | 447 | ret = __vfs_read(file, buf, count, pos); |
c43e259c | 448 | if (ret > 0) { |
2a12a9d7 | 449 | fsnotify_access(file); |
c43e259c | 450 | add_rchar(current, ret); |
1da177e4 | 451 | } |
c43e259c | 452 | inc_syscr(current); |
1da177e4 LT |
453 | } |
454 | ||
455 | return ret; | |
456 | } | |
457 | ||
458 | EXPORT_SYMBOL(vfs_read); | |
459 | ||
5d5d5689 | 460 | static ssize_t new_sync_write(struct file *filp, const char __user *buf, size_t len, loff_t *ppos) |
293bc982 AV |
461 | { |
462 | struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = len }; | |
463 | struct kiocb kiocb; | |
464 | struct iov_iter iter; | |
465 | ssize_t ret; | |
466 | ||
467 | init_sync_kiocb(&kiocb, filp); | |
468 | kiocb.ki_pos = *ppos; | |
293bc982 AV |
469 | iov_iter_init(&iter, WRITE, &iov, 1, len); |
470 | ||
bb7462b6 | 471 | ret = call_write_iter(filp, &kiocb, &iter); |
599bd19b | 472 | BUG_ON(ret == -EIOCBQUEUED); |
f765b134 AV |
473 | if (ret > 0) |
474 | *ppos = kiocb.ki_pos; | |
293bc982 AV |
475 | return ret; |
476 | } | |
477 | ||
493c84c0 AV |
478 | ssize_t __vfs_write(struct file *file, const char __user *p, size_t count, |
479 | loff_t *pos) | |
480 | { | |
481 | if (file->f_op->write) | |
482 | return file->f_op->write(file, p, count, pos); | |
493c84c0 AV |
483 | else if (file->f_op->write_iter) |
484 | return new_sync_write(file, p, count, pos); | |
485 | else | |
486 | return -EINVAL; | |
487 | } | |
488 | EXPORT_SYMBOL(__vfs_write); | |
489 | ||
06ae43f3 AV |
490 | ssize_t __kernel_write(struct file *file, const char *buf, size_t count, loff_t *pos) |
491 | { | |
492 | mm_segment_t old_fs; | |
493 | const char __user *p; | |
494 | ssize_t ret; | |
495 | ||
7f7f25e8 | 496 | if (!(file->f_mode & FMODE_CAN_WRITE)) |
3e84f48e AV |
497 | return -EINVAL; |
498 | ||
06ae43f3 AV |
499 | old_fs = get_fs(); |
500 | set_fs(get_ds()); | |
501 | p = (__force const char __user *)buf; | |
502 | if (count > MAX_RW_COUNT) | |
503 | count = MAX_RW_COUNT; | |
493c84c0 | 504 | ret = __vfs_write(file, p, count, pos); |
06ae43f3 AV |
505 | set_fs(old_fs); |
506 | if (ret > 0) { | |
507 | fsnotify_modify(file); | |
508 | add_wchar(current, ret); | |
509 | } | |
510 | inc_syscw(current); | |
511 | return ret; | |
512 | } | |
2ec3a12a AV |
513 | EXPORT_SYMBOL(__kernel_write); |
514 | ||
e13ec939 CH |
515 | ssize_t kernel_write(struct file *file, const void *buf, size_t count, |
516 | loff_t *pos) | |
ac452aca CH |
517 | { |
518 | mm_segment_t old_fs; | |
519 | ssize_t res; | |
520 | ||
521 | old_fs = get_fs(); | |
522 | set_fs(get_ds()); | |
523 | /* The cast to a user pointer is valid due to the set_fs() */ | |
e13ec939 | 524 | res = vfs_write(file, (__force const char __user *)buf, count, pos); |
ac452aca CH |
525 | set_fs(old_fs); |
526 | ||
527 | return res; | |
528 | } | |
529 | EXPORT_SYMBOL(kernel_write); | |
530 | ||
1da177e4 LT |
531 | ssize_t vfs_write(struct file *file, const char __user *buf, size_t count, loff_t *pos) |
532 | { | |
533 | ssize_t ret; | |
534 | ||
535 | if (!(file->f_mode & FMODE_WRITE)) | |
536 | return -EBADF; | |
7f7f25e8 | 537 | if (!(file->f_mode & FMODE_CAN_WRITE)) |
1da177e4 LT |
538 | return -EINVAL; |
539 | if (unlikely(!access_ok(VERIFY_READ, buf, count))) | |
540 | return -EFAULT; | |
541 | ||
542 | ret = rw_verify_area(WRITE, file, pos, count); | |
bc61384d AV |
543 | if (!ret) { |
544 | if (count > MAX_RW_COUNT) | |
545 | count = MAX_RW_COUNT; | |
03d95eb2 | 546 | file_start_write(file); |
493c84c0 | 547 | ret = __vfs_write(file, buf, count, pos); |
c43e259c | 548 | if (ret > 0) { |
2a12a9d7 | 549 | fsnotify_modify(file); |
c43e259c | 550 | add_wchar(current, ret); |
1da177e4 | 551 | } |
c43e259c | 552 | inc_syscw(current); |
03d95eb2 | 553 | file_end_write(file); |
1da177e4 LT |
554 | } |
555 | ||
556 | return ret; | |
557 | } | |
558 | ||
559 | EXPORT_SYMBOL(vfs_write); | |
560 | ||
561 | static inline loff_t file_pos_read(struct file *file) | |
562 | { | |
563 | return file->f_pos; | |
564 | } | |
565 | ||
566 | static inline void file_pos_write(struct file *file, loff_t pos) | |
567 | { | |
568 | file->f_pos = pos; | |
569 | } | |
570 | ||
3cdad428 | 571 | SYSCALL_DEFINE3(read, unsigned int, fd, char __user *, buf, size_t, count) |
1da177e4 | 572 | { |
9c225f26 | 573 | struct fd f = fdget_pos(fd); |
1da177e4 | 574 | ssize_t ret = -EBADF; |
1da177e4 | 575 | |
2903ff01 AV |
576 | if (f.file) { |
577 | loff_t pos = file_pos_read(f.file); | |
578 | ret = vfs_read(f.file, buf, count, &pos); | |
5faf153e AV |
579 | if (ret >= 0) |
580 | file_pos_write(f.file, pos); | |
9c225f26 | 581 | fdput_pos(f); |
1da177e4 | 582 | } |
1da177e4 LT |
583 | return ret; |
584 | } | |
1da177e4 | 585 | |
3cdad428 HC |
586 | SYSCALL_DEFINE3(write, unsigned int, fd, const char __user *, buf, |
587 | size_t, count) | |
1da177e4 | 588 | { |
9c225f26 | 589 | struct fd f = fdget_pos(fd); |
1da177e4 | 590 | ssize_t ret = -EBADF; |
1da177e4 | 591 | |
2903ff01 AV |
592 | if (f.file) { |
593 | loff_t pos = file_pos_read(f.file); | |
594 | ret = vfs_write(f.file, buf, count, &pos); | |
5faf153e AV |
595 | if (ret >= 0) |
596 | file_pos_write(f.file, pos); | |
9c225f26 | 597 | fdput_pos(f); |
1da177e4 LT |
598 | } |
599 | ||
600 | return ret; | |
601 | } | |
602 | ||
4a0fd5bf AV |
603 | SYSCALL_DEFINE4(pread64, unsigned int, fd, char __user *, buf, |
604 | size_t, count, loff_t, pos) | |
1da177e4 | 605 | { |
2903ff01 | 606 | struct fd f; |
1da177e4 | 607 | ssize_t ret = -EBADF; |
1da177e4 LT |
608 | |
609 | if (pos < 0) | |
610 | return -EINVAL; | |
611 | ||
2903ff01 AV |
612 | f = fdget(fd); |
613 | if (f.file) { | |
1da177e4 | 614 | ret = -ESPIPE; |
2903ff01 AV |
615 | if (f.file->f_mode & FMODE_PREAD) |
616 | ret = vfs_read(f.file, buf, count, &pos); | |
617 | fdput(f); | |
1da177e4 LT |
618 | } |
619 | ||
620 | return ret; | |
621 | } | |
622 | ||
4a0fd5bf AV |
623 | SYSCALL_DEFINE4(pwrite64, unsigned int, fd, const char __user *, buf, |
624 | size_t, count, loff_t, pos) | |
1da177e4 | 625 | { |
2903ff01 | 626 | struct fd f; |
1da177e4 | 627 | ssize_t ret = -EBADF; |
1da177e4 LT |
628 | |
629 | if (pos < 0) | |
630 | return -EINVAL; | |
631 | ||
2903ff01 AV |
632 | f = fdget(fd); |
633 | if (f.file) { | |
1da177e4 | 634 | ret = -ESPIPE; |
2903ff01 AV |
635 | if (f.file->f_mode & FMODE_PWRITE) |
636 | ret = vfs_write(f.file, buf, count, &pos); | |
637 | fdput(f); | |
1da177e4 LT |
638 | } |
639 | ||
640 | return ret; | |
641 | } | |
642 | ||
643 | /* | |
644 | * Reduce an iovec's length in-place. Return the resulting number of segments | |
645 | */ | |
646 | unsigned long iov_shorten(struct iovec *iov, unsigned long nr_segs, size_t to) | |
647 | { | |
648 | unsigned long seg = 0; | |
649 | size_t len = 0; | |
650 | ||
651 | while (seg < nr_segs) { | |
652 | seg++; | |
653 | if (len + iov->iov_len >= to) { | |
654 | iov->iov_len = to - len; | |
655 | break; | |
656 | } | |
657 | len += iov->iov_len; | |
658 | iov++; | |
659 | } | |
660 | return seg; | |
661 | } | |
19295529 | 662 | EXPORT_SYMBOL(iov_shorten); |
1da177e4 | 663 | |
ac15ac06 | 664 | static ssize_t do_iter_readv_writev(struct file *filp, struct iov_iter *iter, |
0f78d06a | 665 | loff_t *ppos, int type, int flags) |
293bc982 AV |
666 | { |
667 | struct kiocb kiocb; | |
293bc982 AV |
668 | ssize_t ret; |
669 | ||
670 | init_sync_kiocb(&kiocb, filp); | |
fdd2f5b7 GR |
671 | ret = kiocb_set_rw_flags(&kiocb, flags); |
672 | if (ret) | |
673 | return ret; | |
293bc982 | 674 | kiocb.ki_pos = *ppos; |
293bc982 | 675 | |
0f78d06a | 676 | if (type == READ) |
bb7462b6 | 677 | ret = call_read_iter(filp, &kiocb, iter); |
0f78d06a | 678 | else |
bb7462b6 | 679 | ret = call_write_iter(filp, &kiocb, iter); |
599bd19b | 680 | BUG_ON(ret == -EIOCBQUEUED); |
293bc982 AV |
681 | *ppos = kiocb.ki_pos; |
682 | return ret; | |
683 | } | |
684 | ||
ee0b3e67 | 685 | /* Do it by hand, with file-ops */ |
ac15ac06 | 686 | static ssize_t do_loop_readv_writev(struct file *filp, struct iov_iter *iter, |
0f78d06a | 687 | loff_t *ppos, int type, int flags) |
ee0b3e67 | 688 | { |
ee0b3e67 BP |
689 | ssize_t ret = 0; |
690 | ||
97be7ebe | 691 | if (flags & ~RWF_HIPRI) |
793b80ef CH |
692 | return -EOPNOTSUPP; |
693 | ||
ac15ac06 AV |
694 | while (iov_iter_count(iter)) { |
695 | struct iovec iovec = iov_iter_iovec(iter); | |
ee0b3e67 BP |
696 | ssize_t nr; |
697 | ||
0f78d06a MS |
698 | if (type == READ) { |
699 | nr = filp->f_op->read(filp, iovec.iov_base, | |
700 | iovec.iov_len, ppos); | |
701 | } else { | |
702 | nr = filp->f_op->write(filp, iovec.iov_base, | |
703 | iovec.iov_len, ppos); | |
704 | } | |
ee0b3e67 BP |
705 | |
706 | if (nr < 0) { | |
707 | if (!ret) | |
708 | ret = nr; | |
709 | break; | |
710 | } | |
711 | ret += nr; | |
ac15ac06 | 712 | if (nr != iovec.iov_len) |
ee0b3e67 | 713 | break; |
ac15ac06 | 714 | iov_iter_advance(iter, nr); |
ee0b3e67 BP |
715 | } |
716 | ||
717 | return ret; | |
718 | } | |
719 | ||
1da177e4 LT |
720 | /* A write operation does a read from user space and vice versa */ |
721 | #define vrfy_dir(type) ((type) == READ ? VERIFY_WRITE : VERIFY_READ) | |
722 | ||
ffecee4f VN |
723 | /** |
724 | * rw_copy_check_uvector() - Copy an array of &struct iovec from userspace | |
725 | * into the kernel and check that it is valid. | |
726 | * | |
727 | * @type: One of %CHECK_IOVEC_ONLY, %READ, or %WRITE. | |
728 | * @uvector: Pointer to the userspace array. | |
729 | * @nr_segs: Number of elements in userspace array. | |
730 | * @fast_segs: Number of elements in @fast_pointer. | |
731 | * @fast_pointer: Pointer to (usually small on-stack) kernel array. | |
732 | * @ret_pointer: (output parameter) Pointer to a variable that will point to | |
733 | * either @fast_pointer, a newly allocated kernel array, or NULL, | |
734 | * depending on which array was used. | |
735 | * | |
736 | * This function copies an array of &struct iovec of @nr_segs from | |
737 | * userspace into the kernel and checks that each element is valid (e.g. | |
738 | * it does not point to a kernel address or cause overflow by being too | |
739 | * large, etc.). | |
740 | * | |
741 | * As an optimization, the caller may provide a pointer to a small | |
742 | * on-stack array in @fast_pointer, typically %UIO_FASTIOV elements long | |
743 | * (the size of this array, or 0 if unused, should be given in @fast_segs). | |
744 | * | |
745 | * @ret_pointer will always point to the array that was used, so the | |
746 | * caller must take care not to call kfree() on it e.g. in case the | |
747 | * @fast_pointer array was used and it was allocated on the stack. | |
748 | * | |
749 | * Return: The total number of bytes covered by the iovec array on success | |
750 | * or a negative error code on error. | |
751 | */ | |
eed4e51f BP |
752 | ssize_t rw_copy_check_uvector(int type, const struct iovec __user * uvector, |
753 | unsigned long nr_segs, unsigned long fast_segs, | |
754 | struct iovec *fast_pointer, | |
ac34ebb3 | 755 | struct iovec **ret_pointer) |
435f49a5 | 756 | { |
eed4e51f | 757 | unsigned long seg; |
435f49a5 | 758 | ssize_t ret; |
eed4e51f BP |
759 | struct iovec *iov = fast_pointer; |
760 | ||
435f49a5 LT |
761 | /* |
762 | * SuS says "The readv() function *may* fail if the iovcnt argument | |
763 | * was less than or equal to 0, or greater than {IOV_MAX}. Linux has | |
764 | * traditionally returned zero for zero segments, so... | |
765 | */ | |
eed4e51f BP |
766 | if (nr_segs == 0) { |
767 | ret = 0; | |
435f49a5 | 768 | goto out; |
eed4e51f BP |
769 | } |
770 | ||
435f49a5 LT |
771 | /* |
772 | * First get the "struct iovec" from user memory and | |
773 | * verify all the pointers | |
774 | */ | |
eed4e51f BP |
775 | if (nr_segs > UIO_MAXIOV) { |
776 | ret = -EINVAL; | |
435f49a5 | 777 | goto out; |
eed4e51f BP |
778 | } |
779 | if (nr_segs > fast_segs) { | |
435f49a5 | 780 | iov = kmalloc(nr_segs*sizeof(struct iovec), GFP_KERNEL); |
eed4e51f BP |
781 | if (iov == NULL) { |
782 | ret = -ENOMEM; | |
435f49a5 | 783 | goto out; |
eed4e51f | 784 | } |
435f49a5 | 785 | } |
eed4e51f BP |
786 | if (copy_from_user(iov, uvector, nr_segs*sizeof(*uvector))) { |
787 | ret = -EFAULT; | |
435f49a5 | 788 | goto out; |
eed4e51f BP |
789 | } |
790 | ||
435f49a5 | 791 | /* |
eed4e51f BP |
792 | * According to the Single Unix Specification we should return EINVAL |
793 | * if an element length is < 0 when cast to ssize_t or if the | |
794 | * total length would overflow the ssize_t return value of the | |
795 | * system call. | |
435f49a5 LT |
796 | * |
797 | * Linux caps all read/write calls to MAX_RW_COUNT, and avoids the | |
798 | * overflow case. | |
799 | */ | |
eed4e51f | 800 | ret = 0; |
435f49a5 LT |
801 | for (seg = 0; seg < nr_segs; seg++) { |
802 | void __user *buf = iov[seg].iov_base; | |
803 | ssize_t len = (ssize_t)iov[seg].iov_len; | |
eed4e51f BP |
804 | |
805 | /* see if we we're about to use an invalid len or if | |
806 | * it's about to overflow ssize_t */ | |
435f49a5 | 807 | if (len < 0) { |
eed4e51f | 808 | ret = -EINVAL; |
435f49a5 | 809 | goto out; |
eed4e51f | 810 | } |
ac34ebb3 | 811 | if (type >= 0 |
fcf63409 | 812 | && unlikely(!access_ok(vrfy_dir(type), buf, len))) { |
eed4e51f | 813 | ret = -EFAULT; |
435f49a5 LT |
814 | goto out; |
815 | } | |
816 | if (len > MAX_RW_COUNT - ret) { | |
817 | len = MAX_RW_COUNT - ret; | |
818 | iov[seg].iov_len = len; | |
eed4e51f | 819 | } |
eed4e51f | 820 | ret += len; |
435f49a5 | 821 | } |
eed4e51f BP |
822 | out: |
823 | *ret_pointer = iov; | |
824 | return ret; | |
825 | } | |
826 | ||
f5029855 AV |
827 | #ifdef CONFIG_COMPAT |
828 | ssize_t compat_rw_copy_check_uvector(int type, | |
829 | const struct compat_iovec __user *uvector, unsigned long nr_segs, | |
830 | unsigned long fast_segs, struct iovec *fast_pointer, | |
831 | struct iovec **ret_pointer) | |
832 | { | |
833 | compat_ssize_t tot_len; | |
834 | struct iovec *iov = *ret_pointer = fast_pointer; | |
835 | ssize_t ret = 0; | |
836 | int seg; | |
837 | ||
838 | /* | |
839 | * SuS says "The readv() function *may* fail if the iovcnt argument | |
840 | * was less than or equal to 0, or greater than {IOV_MAX}. Linux has | |
841 | * traditionally returned zero for zero segments, so... | |
842 | */ | |
843 | if (nr_segs == 0) | |
844 | goto out; | |
845 | ||
846 | ret = -EINVAL; | |
847 | if (nr_segs > UIO_MAXIOV) | |
848 | goto out; | |
849 | if (nr_segs > fast_segs) { | |
850 | ret = -ENOMEM; | |
851 | iov = kmalloc(nr_segs*sizeof(struct iovec), GFP_KERNEL); | |
852 | if (iov == NULL) | |
853 | goto out; | |
854 | } | |
855 | *ret_pointer = iov; | |
856 | ||
857 | ret = -EFAULT; | |
858 | if (!access_ok(VERIFY_READ, uvector, nr_segs*sizeof(*uvector))) | |
859 | goto out; | |
860 | ||
861 | /* | |
862 | * Single unix specification: | |
863 | * We should -EINVAL if an element length is not >= 0 and fitting an | |
864 | * ssize_t. | |
865 | * | |
866 | * In Linux, the total length is limited to MAX_RW_COUNT, there is | |
867 | * no overflow possibility. | |
868 | */ | |
869 | tot_len = 0; | |
870 | ret = -EINVAL; | |
871 | for (seg = 0; seg < nr_segs; seg++) { | |
872 | compat_uptr_t buf; | |
873 | compat_ssize_t len; | |
874 | ||
875 | if (__get_user(len, &uvector->iov_len) || | |
876 | __get_user(buf, &uvector->iov_base)) { | |
877 | ret = -EFAULT; | |
878 | goto out; | |
879 | } | |
880 | if (len < 0) /* size_t not fitting in compat_ssize_t .. */ | |
881 | goto out; | |
882 | if (type >= 0 && | |
883 | !access_ok(vrfy_dir(type), compat_ptr(buf), len)) { | |
884 | ret = -EFAULT; | |
885 | goto out; | |
886 | } | |
887 | if (len > MAX_RW_COUNT - tot_len) | |
888 | len = MAX_RW_COUNT - tot_len; | |
889 | tot_len += len; | |
890 | iov->iov_base = compat_ptr(buf); | |
891 | iov->iov_len = (compat_size_t) len; | |
892 | uvector++; | |
893 | iov++; | |
894 | } | |
895 | ret = tot_len; | |
896 | ||
897 | out: | |
898 | return ret; | |
899 | } | |
900 | #endif | |
901 | ||
19c73586 CH |
902 | static ssize_t do_iter_read(struct file *file, struct iov_iter *iter, |
903 | loff_t *pos, int flags) | |
1da177e4 | 904 | { |
1da177e4 | 905 | size_t tot_len; |
7687a7a4 | 906 | ssize_t ret = 0; |
1da177e4 | 907 | |
edab5fe3 CH |
908 | if (!(file->f_mode & FMODE_READ)) |
909 | return -EBADF; | |
910 | if (!(file->f_mode & FMODE_CAN_READ)) | |
911 | return -EINVAL; | |
912 | ||
7687a7a4 | 913 | tot_len = iov_iter_count(iter); |
0504c074 AV |
914 | if (!tot_len) |
915 | goto out; | |
19c73586 | 916 | ret = rw_verify_area(READ, file, pos, tot_len); |
e28cc715 | 917 | if (ret < 0) |
19c73586 | 918 | return ret; |
1da177e4 | 919 | |
19c73586 CH |
920 | if (file->f_op->read_iter) |
921 | ret = do_iter_readv_writev(file, iter, pos, READ, flags); | |
ee0b3e67 | 922 | else |
19c73586 | 923 | ret = do_loop_readv_writev(file, iter, pos, READ, flags); |
1da177e4 | 924 | out: |
19c73586 CH |
925 | if (ret >= 0) |
926 | fsnotify_access(file); | |
1da177e4 | 927 | return ret; |
1da177e4 LT |
928 | } |
929 | ||
18e9710e CH |
930 | ssize_t vfs_iter_read(struct file *file, struct iov_iter *iter, loff_t *ppos, |
931 | int flags) | |
7687a7a4 | 932 | { |
18e9710e CH |
933 | if (!file->f_op->read_iter) |
934 | return -EINVAL; | |
935 | return do_iter_read(file, iter, ppos, flags); | |
936 | } | |
937 | EXPORT_SYMBOL(vfs_iter_read); | |
7687a7a4 | 938 | |
19c73586 CH |
939 | static ssize_t do_iter_write(struct file *file, struct iov_iter *iter, |
940 | loff_t *pos, int flags) | |
941 | { | |
942 | size_t tot_len; | |
943 | ssize_t ret = 0; | |
03d95eb2 | 944 | |
edab5fe3 CH |
945 | if (!(file->f_mode & FMODE_WRITE)) |
946 | return -EBADF; | |
947 | if (!(file->f_mode & FMODE_CAN_WRITE)) | |
948 | return -EINVAL; | |
949 | ||
19c73586 CH |
950 | tot_len = iov_iter_count(iter); |
951 | if (!tot_len) | |
952 | return 0; | |
953 | ret = rw_verify_area(WRITE, file, pos, tot_len); | |
7687a7a4 MS |
954 | if (ret < 0) |
955 | return ret; | |
956 | ||
19c73586 CH |
957 | if (file->f_op->write_iter) |
958 | ret = do_iter_readv_writev(file, iter, pos, WRITE, flags); | |
959 | else | |
960 | ret = do_loop_readv_writev(file, iter, pos, WRITE, flags); | |
19c73586 CH |
961 | if (ret > 0) |
962 | fsnotify_modify(file); | |
7687a7a4 MS |
963 | return ret; |
964 | } | |
965 | ||
abbb6589 CH |
966 | ssize_t vfs_iter_write(struct file *file, struct iov_iter *iter, loff_t *ppos, |
967 | int flags) | |
968 | { | |
969 | if (!file->f_op->write_iter) | |
970 | return -EINVAL; | |
971 | return do_iter_write(file, iter, ppos, flags); | |
972 | } | |
973 | EXPORT_SYMBOL(vfs_iter_write); | |
974 | ||
1da177e4 | 975 | ssize_t vfs_readv(struct file *file, const struct iovec __user *vec, |
793b80ef | 976 | unsigned long vlen, loff_t *pos, int flags) |
1da177e4 | 977 | { |
7687a7a4 MS |
978 | struct iovec iovstack[UIO_FASTIOV]; |
979 | struct iovec *iov = iovstack; | |
980 | struct iov_iter iter; | |
981 | ssize_t ret; | |
1da177e4 | 982 | |
251b42a1 | 983 | ret = import_iovec(READ, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter); |
edab5fe3 CH |
984 | if (ret >= 0) { |
985 | ret = do_iter_read(file, &iter, pos, flags); | |
986 | kfree(iov); | |
987 | } | |
1da177e4 | 988 | |
251b42a1 CH |
989 | return ret; |
990 | } | |
1da177e4 LT |
991 | EXPORT_SYMBOL(vfs_readv); |
992 | ||
993 | ssize_t vfs_writev(struct file *file, const struct iovec __user *vec, | |
793b80ef | 994 | unsigned long vlen, loff_t *pos, int flags) |
1da177e4 | 995 | { |
251b42a1 CH |
996 | struct iovec iovstack[UIO_FASTIOV]; |
997 | struct iovec *iov = iovstack; | |
998 | struct iov_iter iter; | |
999 | ssize_t ret; | |
1da177e4 | 1000 | |
251b42a1 | 1001 | ret = import_iovec(WRITE, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter); |
edab5fe3 | 1002 | if (ret >= 0) { |
62473a2d | 1003 | file_start_write(file); |
edab5fe3 | 1004 | ret = do_iter_write(file, &iter, pos, flags); |
62473a2d | 1005 | file_end_write(file); |
edab5fe3 CH |
1006 | kfree(iov); |
1007 | } | |
251b42a1 | 1008 | return ret; |
1da177e4 | 1009 | } |
1da177e4 LT |
1010 | EXPORT_SYMBOL(vfs_writev); |
1011 | ||
f17d8b35 MT |
1012 | static ssize_t do_readv(unsigned long fd, const struct iovec __user *vec, |
1013 | unsigned long vlen, int flags) | |
1da177e4 | 1014 | { |
9c225f26 | 1015 | struct fd f = fdget_pos(fd); |
1da177e4 | 1016 | ssize_t ret = -EBADF; |
1da177e4 | 1017 | |
2903ff01 AV |
1018 | if (f.file) { |
1019 | loff_t pos = file_pos_read(f.file); | |
f17d8b35 | 1020 | ret = vfs_readv(f.file, vec, vlen, &pos, flags); |
5faf153e AV |
1021 | if (ret >= 0) |
1022 | file_pos_write(f.file, pos); | |
9c225f26 | 1023 | fdput_pos(f); |
1da177e4 LT |
1024 | } |
1025 | ||
1026 | if (ret > 0) | |
4b98d11b AD |
1027 | add_rchar(current, ret); |
1028 | inc_syscr(current); | |
1da177e4 LT |
1029 | return ret; |
1030 | } | |
1031 | ||
f17d8b35 MT |
1032 | static ssize_t do_writev(unsigned long fd, const struct iovec __user *vec, |
1033 | unsigned long vlen, int flags) | |
1da177e4 | 1034 | { |
9c225f26 | 1035 | struct fd f = fdget_pos(fd); |
1da177e4 | 1036 | ssize_t ret = -EBADF; |
1da177e4 | 1037 | |
2903ff01 AV |
1038 | if (f.file) { |
1039 | loff_t pos = file_pos_read(f.file); | |
f17d8b35 | 1040 | ret = vfs_writev(f.file, vec, vlen, &pos, flags); |
5faf153e AV |
1041 | if (ret >= 0) |
1042 | file_pos_write(f.file, pos); | |
9c225f26 | 1043 | fdput_pos(f); |
1da177e4 LT |
1044 | } |
1045 | ||
1046 | if (ret > 0) | |
4b98d11b AD |
1047 | add_wchar(current, ret); |
1048 | inc_syscw(current); | |
1da177e4 LT |
1049 | return ret; |
1050 | } | |
1051 | ||
601cc11d LT |
1052 | static inline loff_t pos_from_hilo(unsigned long high, unsigned long low) |
1053 | { | |
1054 | #define HALF_LONG_BITS (BITS_PER_LONG / 2) | |
1055 | return (((loff_t)high << HALF_LONG_BITS) << HALF_LONG_BITS) | low; | |
1056 | } | |
1057 | ||
f17d8b35 MT |
1058 | static ssize_t do_preadv(unsigned long fd, const struct iovec __user *vec, |
1059 | unsigned long vlen, loff_t pos, int flags) | |
f3554f4b | 1060 | { |
2903ff01 | 1061 | struct fd f; |
f3554f4b | 1062 | ssize_t ret = -EBADF; |
f3554f4b GH |
1063 | |
1064 | if (pos < 0) | |
1065 | return -EINVAL; | |
1066 | ||
2903ff01 AV |
1067 | f = fdget(fd); |
1068 | if (f.file) { | |
f3554f4b | 1069 | ret = -ESPIPE; |
2903ff01 | 1070 | if (f.file->f_mode & FMODE_PREAD) |
f17d8b35 | 1071 | ret = vfs_readv(f.file, vec, vlen, &pos, flags); |
2903ff01 | 1072 | fdput(f); |
f3554f4b GH |
1073 | } |
1074 | ||
1075 | if (ret > 0) | |
1076 | add_rchar(current, ret); | |
1077 | inc_syscr(current); | |
1078 | return ret; | |
1079 | } | |
1080 | ||
f17d8b35 MT |
1081 | static ssize_t do_pwritev(unsigned long fd, const struct iovec __user *vec, |
1082 | unsigned long vlen, loff_t pos, int flags) | |
f3554f4b | 1083 | { |
2903ff01 | 1084 | struct fd f; |
f3554f4b | 1085 | ssize_t ret = -EBADF; |
f3554f4b GH |
1086 | |
1087 | if (pos < 0) | |
1088 | return -EINVAL; | |
1089 | ||
2903ff01 AV |
1090 | f = fdget(fd); |
1091 | if (f.file) { | |
f3554f4b | 1092 | ret = -ESPIPE; |
2903ff01 | 1093 | if (f.file->f_mode & FMODE_PWRITE) |
f17d8b35 | 1094 | ret = vfs_writev(f.file, vec, vlen, &pos, flags); |
2903ff01 | 1095 | fdput(f); |
f3554f4b GH |
1096 | } |
1097 | ||
1098 | if (ret > 0) | |
1099 | add_wchar(current, ret); | |
1100 | inc_syscw(current); | |
1101 | return ret; | |
1102 | } | |
1103 | ||
f17d8b35 MT |
1104 | SYSCALL_DEFINE3(readv, unsigned long, fd, const struct iovec __user *, vec, |
1105 | unsigned long, vlen) | |
1106 | { | |
1107 | return do_readv(fd, vec, vlen, 0); | |
1108 | } | |
1109 | ||
1110 | SYSCALL_DEFINE3(writev, unsigned long, fd, const struct iovec __user *, vec, | |
1111 | unsigned long, vlen) | |
1112 | { | |
1113 | return do_writev(fd, vec, vlen, 0); | |
1114 | } | |
1115 | ||
1116 | SYSCALL_DEFINE5(preadv, unsigned long, fd, const struct iovec __user *, vec, | |
1117 | unsigned long, vlen, unsigned long, pos_l, unsigned long, pos_h) | |
1118 | { | |
1119 | loff_t pos = pos_from_hilo(pos_h, pos_l); | |
1120 | ||
1121 | return do_preadv(fd, vec, vlen, pos, 0); | |
1122 | } | |
1123 | ||
1124 | SYSCALL_DEFINE6(preadv2, unsigned long, fd, const struct iovec __user *, vec, | |
1125 | unsigned long, vlen, unsigned long, pos_l, unsigned long, pos_h, | |
1126 | int, flags) | |
1127 | { | |
1128 | loff_t pos = pos_from_hilo(pos_h, pos_l); | |
1129 | ||
1130 | if (pos == -1) | |
1131 | return do_readv(fd, vec, vlen, flags); | |
1132 | ||
1133 | return do_preadv(fd, vec, vlen, pos, flags); | |
1134 | } | |
1135 | ||
1136 | SYSCALL_DEFINE5(pwritev, unsigned long, fd, const struct iovec __user *, vec, | |
1137 | unsigned long, vlen, unsigned long, pos_l, unsigned long, pos_h) | |
1138 | { | |
1139 | loff_t pos = pos_from_hilo(pos_h, pos_l); | |
1140 | ||
1141 | return do_pwritev(fd, vec, vlen, pos, 0); | |
1142 | } | |
1143 | ||
1144 | SYSCALL_DEFINE6(pwritev2, unsigned long, fd, const struct iovec __user *, vec, | |
1145 | unsigned long, vlen, unsigned long, pos_l, unsigned long, pos_h, | |
1146 | int, flags) | |
1147 | { | |
1148 | loff_t pos = pos_from_hilo(pos_h, pos_l); | |
1149 | ||
1150 | if (pos == -1) | |
1151 | return do_writev(fd, vec, vlen, flags); | |
1152 | ||
1153 | return do_pwritev(fd, vec, vlen, pos, flags); | |
1154 | } | |
1155 | ||
72ec3516 | 1156 | #ifdef CONFIG_COMPAT |
72ec3516 AV |
1157 | static size_t compat_readv(struct file *file, |
1158 | const struct compat_iovec __user *vec, | |
f17d8b35 | 1159 | unsigned long vlen, loff_t *pos, int flags) |
72ec3516 | 1160 | { |
72ec3516 AV |
1161 | struct iovec iovstack[UIO_FASTIOV]; |
1162 | struct iovec *iov = iovstack; | |
ac15ac06 | 1163 | struct iov_iter iter; |
72ec3516 | 1164 | ssize_t ret; |
72ec3516 | 1165 | |
26c87fb7 | 1166 | ret = compat_import_iovec(READ, vec, vlen, UIO_FASTIOV, &iov, &iter); |
edab5fe3 CH |
1167 | if (ret >= 0) { |
1168 | ret = do_iter_read(file, &iter, pos, flags); | |
1169 | kfree(iov); | |
1170 | } | |
72ec3516 AV |
1171 | if (ret > 0) |
1172 | add_rchar(current, ret); | |
1173 | inc_syscr(current); | |
1174 | return ret; | |
1175 | } | |
1176 | ||
f17d8b35 MT |
1177 | static size_t do_compat_readv(compat_ulong_t fd, |
1178 | const struct compat_iovec __user *vec, | |
1179 | compat_ulong_t vlen, int flags) | |
72ec3516 | 1180 | { |
9c225f26 | 1181 | struct fd f = fdget_pos(fd); |
72ec3516 AV |
1182 | ssize_t ret; |
1183 | loff_t pos; | |
1184 | ||
1185 | if (!f.file) | |
1186 | return -EBADF; | |
1187 | pos = f.file->f_pos; | |
f17d8b35 | 1188 | ret = compat_readv(f.file, vec, vlen, &pos, flags); |
5faf153e AV |
1189 | if (ret >= 0) |
1190 | f.file->f_pos = pos; | |
9c225f26 | 1191 | fdput_pos(f); |
72ec3516 | 1192 | return ret; |
f17d8b35 | 1193 | |
72ec3516 AV |
1194 | } |
1195 | ||
f17d8b35 MT |
1196 | COMPAT_SYSCALL_DEFINE3(readv, compat_ulong_t, fd, |
1197 | const struct compat_iovec __user *,vec, | |
1198 | compat_ulong_t, vlen) | |
1199 | { | |
1200 | return do_compat_readv(fd, vec, vlen, 0); | |
1201 | } | |
1202 | ||
1203 | static long do_compat_preadv64(unsigned long fd, | |
378a10f3 | 1204 | const struct compat_iovec __user *vec, |
f17d8b35 | 1205 | unsigned long vlen, loff_t pos, int flags) |
72ec3516 AV |
1206 | { |
1207 | struct fd f; | |
1208 | ssize_t ret; | |
1209 | ||
1210 | if (pos < 0) | |
1211 | return -EINVAL; | |
1212 | f = fdget(fd); | |
1213 | if (!f.file) | |
1214 | return -EBADF; | |
1215 | ret = -ESPIPE; | |
1216 | if (f.file->f_mode & FMODE_PREAD) | |
f17d8b35 | 1217 | ret = compat_readv(f.file, vec, vlen, &pos, flags); |
72ec3516 AV |
1218 | fdput(f); |
1219 | return ret; | |
1220 | } | |
1221 | ||
378a10f3 HC |
1222 | #ifdef __ARCH_WANT_COMPAT_SYS_PREADV64 |
1223 | COMPAT_SYSCALL_DEFINE4(preadv64, unsigned long, fd, | |
1224 | const struct compat_iovec __user *,vec, | |
1225 | unsigned long, vlen, loff_t, pos) | |
1226 | { | |
f17d8b35 | 1227 | return do_compat_preadv64(fd, vec, vlen, pos, 0); |
378a10f3 HC |
1228 | } |
1229 | #endif | |
1230 | ||
dfd948e3 | 1231 | COMPAT_SYSCALL_DEFINE5(preadv, compat_ulong_t, fd, |
72ec3516 | 1232 | const struct compat_iovec __user *,vec, |
dfd948e3 | 1233 | compat_ulong_t, vlen, u32, pos_low, u32, pos_high) |
72ec3516 AV |
1234 | { |
1235 | loff_t pos = ((loff_t)pos_high << 32) | pos_low; | |
378a10f3 | 1236 | |
f17d8b35 MT |
1237 | return do_compat_preadv64(fd, vec, vlen, pos, 0); |
1238 | } | |
1239 | ||
3ebfd81f L |
1240 | #ifdef __ARCH_WANT_COMPAT_SYS_PREADV64V2 |
1241 | COMPAT_SYSCALL_DEFINE5(preadv64v2, unsigned long, fd, | |
1242 | const struct compat_iovec __user *,vec, | |
1243 | unsigned long, vlen, loff_t, pos, int, flags) | |
1244 | { | |
1245 | return do_compat_preadv64(fd, vec, vlen, pos, flags); | |
1246 | } | |
1247 | #endif | |
1248 | ||
f17d8b35 MT |
1249 | COMPAT_SYSCALL_DEFINE6(preadv2, compat_ulong_t, fd, |
1250 | const struct compat_iovec __user *,vec, | |
1251 | compat_ulong_t, vlen, u32, pos_low, u32, pos_high, | |
1252 | int, flags) | |
1253 | { | |
1254 | loff_t pos = ((loff_t)pos_high << 32) | pos_low; | |
1255 | ||
1256 | if (pos == -1) | |
1257 | return do_compat_readv(fd, vec, vlen, flags); | |
1258 | ||
1259 | return do_compat_preadv64(fd, vec, vlen, pos, flags); | |
72ec3516 AV |
1260 | } |
1261 | ||
1262 | static size_t compat_writev(struct file *file, | |
1263 | const struct compat_iovec __user *vec, | |
f17d8b35 | 1264 | unsigned long vlen, loff_t *pos, int flags) |
72ec3516 | 1265 | { |
26c87fb7 CH |
1266 | struct iovec iovstack[UIO_FASTIOV]; |
1267 | struct iovec *iov = iovstack; | |
1268 | struct iov_iter iter; | |
edab5fe3 | 1269 | ssize_t ret; |
72ec3516 | 1270 | |
26c87fb7 | 1271 | ret = compat_import_iovec(WRITE, vec, vlen, UIO_FASTIOV, &iov, &iter); |
edab5fe3 | 1272 | if (ret >= 0) { |
62473a2d | 1273 | file_start_write(file); |
edab5fe3 | 1274 | ret = do_iter_write(file, &iter, pos, flags); |
62473a2d | 1275 | file_end_write(file); |
edab5fe3 CH |
1276 | kfree(iov); |
1277 | } | |
72ec3516 AV |
1278 | if (ret > 0) |
1279 | add_wchar(current, ret); | |
1280 | inc_syscw(current); | |
1281 | return ret; | |
1282 | } | |
1283 | ||
f17d8b35 MT |
1284 | static size_t do_compat_writev(compat_ulong_t fd, |
1285 | const struct compat_iovec __user* vec, | |
1286 | compat_ulong_t vlen, int flags) | |
72ec3516 | 1287 | { |
9c225f26 | 1288 | struct fd f = fdget_pos(fd); |
72ec3516 AV |
1289 | ssize_t ret; |
1290 | loff_t pos; | |
1291 | ||
1292 | if (!f.file) | |
1293 | return -EBADF; | |
1294 | pos = f.file->f_pos; | |
f17d8b35 | 1295 | ret = compat_writev(f.file, vec, vlen, &pos, flags); |
5faf153e AV |
1296 | if (ret >= 0) |
1297 | f.file->f_pos = pos; | |
9c225f26 | 1298 | fdput_pos(f); |
72ec3516 AV |
1299 | return ret; |
1300 | } | |
1301 | ||
f17d8b35 MT |
1302 | COMPAT_SYSCALL_DEFINE3(writev, compat_ulong_t, fd, |
1303 | const struct compat_iovec __user *, vec, | |
1304 | compat_ulong_t, vlen) | |
1305 | { | |
1306 | return do_compat_writev(fd, vec, vlen, 0); | |
1307 | } | |
1308 | ||
1309 | static long do_compat_pwritev64(unsigned long fd, | |
378a10f3 | 1310 | const struct compat_iovec __user *vec, |
f17d8b35 | 1311 | unsigned long vlen, loff_t pos, int flags) |
72ec3516 AV |
1312 | { |
1313 | struct fd f; | |
1314 | ssize_t ret; | |
1315 | ||
1316 | if (pos < 0) | |
1317 | return -EINVAL; | |
1318 | f = fdget(fd); | |
1319 | if (!f.file) | |
1320 | return -EBADF; | |
1321 | ret = -ESPIPE; | |
1322 | if (f.file->f_mode & FMODE_PWRITE) | |
f17d8b35 | 1323 | ret = compat_writev(f.file, vec, vlen, &pos, flags); |
72ec3516 AV |
1324 | fdput(f); |
1325 | return ret; | |
1326 | } | |
1327 | ||
378a10f3 HC |
1328 | #ifdef __ARCH_WANT_COMPAT_SYS_PWRITEV64 |
1329 | COMPAT_SYSCALL_DEFINE4(pwritev64, unsigned long, fd, | |
1330 | const struct compat_iovec __user *,vec, | |
1331 | unsigned long, vlen, loff_t, pos) | |
1332 | { | |
f17d8b35 | 1333 | return do_compat_pwritev64(fd, vec, vlen, pos, 0); |
378a10f3 HC |
1334 | } |
1335 | #endif | |
1336 | ||
dfd948e3 | 1337 | COMPAT_SYSCALL_DEFINE5(pwritev, compat_ulong_t, fd, |
72ec3516 | 1338 | const struct compat_iovec __user *,vec, |
dfd948e3 | 1339 | compat_ulong_t, vlen, u32, pos_low, u32, pos_high) |
72ec3516 AV |
1340 | { |
1341 | loff_t pos = ((loff_t)pos_high << 32) | pos_low; | |
378a10f3 | 1342 | |
f17d8b35 | 1343 | return do_compat_pwritev64(fd, vec, vlen, pos, 0); |
72ec3516 | 1344 | } |
f17d8b35 | 1345 | |
3ebfd81f L |
1346 | #ifdef __ARCH_WANT_COMPAT_SYS_PWRITEV64V2 |
1347 | COMPAT_SYSCALL_DEFINE5(pwritev64v2, unsigned long, fd, | |
1348 | const struct compat_iovec __user *,vec, | |
1349 | unsigned long, vlen, loff_t, pos, int, flags) | |
1350 | { | |
1351 | return do_compat_pwritev64(fd, vec, vlen, pos, flags); | |
1352 | } | |
1353 | #endif | |
1354 | ||
f17d8b35 MT |
1355 | COMPAT_SYSCALL_DEFINE6(pwritev2, compat_ulong_t, fd, |
1356 | const struct compat_iovec __user *,vec, | |
1357 | compat_ulong_t, vlen, u32, pos_low, u32, pos_high, int, flags) | |
1358 | { | |
1359 | loff_t pos = ((loff_t)pos_high << 32) | pos_low; | |
1360 | ||
1361 | if (pos == -1) | |
1362 | return do_compat_writev(fd, vec, vlen, flags); | |
1363 | ||
1364 | return do_compat_pwritev64(fd, vec, vlen, pos, flags); | |
72ec3516 | 1365 | } |
f17d8b35 | 1366 | |
72ec3516 AV |
1367 | #endif |
1368 | ||
19f4fc3a AV |
1369 | static ssize_t do_sendfile(int out_fd, int in_fd, loff_t *ppos, |
1370 | size_t count, loff_t max) | |
1da177e4 | 1371 | { |
2903ff01 AV |
1372 | struct fd in, out; |
1373 | struct inode *in_inode, *out_inode; | |
1da177e4 | 1374 | loff_t pos; |
7995bd28 | 1375 | loff_t out_pos; |
1da177e4 | 1376 | ssize_t retval; |
2903ff01 | 1377 | int fl; |
1da177e4 LT |
1378 | |
1379 | /* | |
1380 | * Get input file, and verify that it is ok.. | |
1381 | */ | |
1382 | retval = -EBADF; | |
2903ff01 AV |
1383 | in = fdget(in_fd); |
1384 | if (!in.file) | |
1da177e4 | 1385 | goto out; |
2903ff01 | 1386 | if (!(in.file->f_mode & FMODE_READ)) |
1da177e4 | 1387 | goto fput_in; |
1da177e4 | 1388 | retval = -ESPIPE; |
7995bd28 AV |
1389 | if (!ppos) { |
1390 | pos = in.file->f_pos; | |
1391 | } else { | |
1392 | pos = *ppos; | |
2903ff01 | 1393 | if (!(in.file->f_mode & FMODE_PREAD)) |
1da177e4 | 1394 | goto fput_in; |
7995bd28 AV |
1395 | } |
1396 | retval = rw_verify_area(READ, in.file, &pos, count); | |
e28cc715 | 1397 | if (retval < 0) |
1da177e4 | 1398 | goto fput_in; |
bc61384d AV |
1399 | if (count > MAX_RW_COUNT) |
1400 | count = MAX_RW_COUNT; | |
1da177e4 | 1401 | |
1da177e4 LT |
1402 | /* |
1403 | * Get output file, and verify that it is ok.. | |
1404 | */ | |
1405 | retval = -EBADF; | |
2903ff01 AV |
1406 | out = fdget(out_fd); |
1407 | if (!out.file) | |
1da177e4 | 1408 | goto fput_in; |
2903ff01 | 1409 | if (!(out.file->f_mode & FMODE_WRITE)) |
1da177e4 LT |
1410 | goto fput_out; |
1411 | retval = -EINVAL; | |
496ad9aa AV |
1412 | in_inode = file_inode(in.file); |
1413 | out_inode = file_inode(out.file); | |
7995bd28 AV |
1414 | out_pos = out.file->f_pos; |
1415 | retval = rw_verify_area(WRITE, out.file, &out_pos, count); | |
e28cc715 | 1416 | if (retval < 0) |
1da177e4 LT |
1417 | goto fput_out; |
1418 | ||
1da177e4 LT |
1419 | if (!max) |
1420 | max = min(in_inode->i_sb->s_maxbytes, out_inode->i_sb->s_maxbytes); | |
1421 | ||
1da177e4 LT |
1422 | if (unlikely(pos + count > max)) { |
1423 | retval = -EOVERFLOW; | |
1424 | if (pos >= max) | |
1425 | goto fput_out; | |
1426 | count = max - pos; | |
1427 | } | |
1428 | ||
d96e6e71 | 1429 | fl = 0; |
534f2aaa | 1430 | #if 0 |
d96e6e71 JA |
1431 | /* |
1432 | * We need to debate whether we can enable this or not. The | |
1433 | * man page documents EAGAIN return for the output at least, | |
1434 | * and the application is arguably buggy if it doesn't expect | |
1435 | * EAGAIN on a non-blocking file descriptor. | |
1436 | */ | |
2903ff01 | 1437 | if (in.file->f_flags & O_NONBLOCK) |
d96e6e71 | 1438 | fl = SPLICE_F_NONBLOCK; |
534f2aaa | 1439 | #endif |
50cd2c57 | 1440 | file_start_write(out.file); |
7995bd28 | 1441 | retval = do_splice_direct(in.file, &pos, out.file, &out_pos, count, fl); |
50cd2c57 | 1442 | file_end_write(out.file); |
1da177e4 LT |
1443 | |
1444 | if (retval > 0) { | |
4b98d11b AD |
1445 | add_rchar(current, retval); |
1446 | add_wchar(current, retval); | |
a68c2f12 SW |
1447 | fsnotify_access(in.file); |
1448 | fsnotify_modify(out.file); | |
7995bd28 AV |
1449 | out.file->f_pos = out_pos; |
1450 | if (ppos) | |
1451 | *ppos = pos; | |
1452 | else | |
1453 | in.file->f_pos = pos; | |
1da177e4 | 1454 | } |
1da177e4 | 1455 | |
4b98d11b AD |
1456 | inc_syscr(current); |
1457 | inc_syscw(current); | |
7995bd28 | 1458 | if (pos > max) |
1da177e4 LT |
1459 | retval = -EOVERFLOW; |
1460 | ||
1461 | fput_out: | |
2903ff01 | 1462 | fdput(out); |
1da177e4 | 1463 | fput_in: |
2903ff01 | 1464 | fdput(in); |
1da177e4 LT |
1465 | out: |
1466 | return retval; | |
1467 | } | |
1468 | ||
002c8976 | 1469 | SYSCALL_DEFINE4(sendfile, int, out_fd, int, in_fd, off_t __user *, offset, size_t, count) |
1da177e4 LT |
1470 | { |
1471 | loff_t pos; | |
1472 | off_t off; | |
1473 | ssize_t ret; | |
1474 | ||
1475 | if (offset) { | |
1476 | if (unlikely(get_user(off, offset))) | |
1477 | return -EFAULT; | |
1478 | pos = off; | |
1479 | ret = do_sendfile(out_fd, in_fd, &pos, count, MAX_NON_LFS); | |
1480 | if (unlikely(put_user(pos, offset))) | |
1481 | return -EFAULT; | |
1482 | return ret; | |
1483 | } | |
1484 | ||
1485 | return do_sendfile(out_fd, in_fd, NULL, count, 0); | |
1486 | } | |
1487 | ||
002c8976 | 1488 | SYSCALL_DEFINE4(sendfile64, int, out_fd, int, in_fd, loff_t __user *, offset, size_t, count) |
1da177e4 LT |
1489 | { |
1490 | loff_t pos; | |
1491 | ssize_t ret; | |
1492 | ||
1493 | if (offset) { | |
1494 | if (unlikely(copy_from_user(&pos, offset, sizeof(loff_t)))) | |
1495 | return -EFAULT; | |
1496 | ret = do_sendfile(out_fd, in_fd, &pos, count, 0); | |
1497 | if (unlikely(put_user(pos, offset))) | |
1498 | return -EFAULT; | |
1499 | return ret; | |
1500 | } | |
1501 | ||
1502 | return do_sendfile(out_fd, in_fd, NULL, count, 0); | |
1503 | } | |
19f4fc3a AV |
1504 | |
1505 | #ifdef CONFIG_COMPAT | |
1506 | COMPAT_SYSCALL_DEFINE4(sendfile, int, out_fd, int, in_fd, | |
1507 | compat_off_t __user *, offset, compat_size_t, count) | |
1508 | { | |
1509 | loff_t pos; | |
1510 | off_t off; | |
1511 | ssize_t ret; | |
1512 | ||
1513 | if (offset) { | |
1514 | if (unlikely(get_user(off, offset))) | |
1515 | return -EFAULT; | |
1516 | pos = off; | |
1517 | ret = do_sendfile(out_fd, in_fd, &pos, count, MAX_NON_LFS); | |
1518 | if (unlikely(put_user(pos, offset))) | |
1519 | return -EFAULT; | |
1520 | return ret; | |
1521 | } | |
1522 | ||
1523 | return do_sendfile(out_fd, in_fd, NULL, count, 0); | |
1524 | } | |
1525 | ||
1526 | COMPAT_SYSCALL_DEFINE4(sendfile64, int, out_fd, int, in_fd, | |
1527 | compat_loff_t __user *, offset, compat_size_t, count) | |
1528 | { | |
1529 | loff_t pos; | |
1530 | ssize_t ret; | |
1531 | ||
1532 | if (offset) { | |
1533 | if (unlikely(copy_from_user(&pos, offset, sizeof(loff_t)))) | |
1534 | return -EFAULT; | |
1535 | ret = do_sendfile(out_fd, in_fd, &pos, count, 0); | |
1536 | if (unlikely(put_user(pos, offset))) | |
1537 | return -EFAULT; | |
1538 | return ret; | |
1539 | } | |
1540 | ||
1541 | return do_sendfile(out_fd, in_fd, NULL, count, 0); | |
1542 | } | |
1543 | #endif | |
29732938 ZB |
1544 | |
1545 | /* | |
1546 | * copy_file_range() differs from regular file read and write in that it | |
1547 | * specifically allows return partial success. When it does so is up to | |
1548 | * the copy_file_range method. | |
1549 | */ | |
1550 | ssize_t vfs_copy_file_range(struct file *file_in, loff_t pos_in, | |
1551 | struct file *file_out, loff_t pos_out, | |
1552 | size_t len, unsigned int flags) | |
1553 | { | |
1554 | struct inode *inode_in = file_inode(file_in); | |
1555 | struct inode *inode_out = file_inode(file_out); | |
1556 | ssize_t ret; | |
1557 | ||
1558 | if (flags != 0) | |
1559 | return -EINVAL; | |
1560 | ||
11cbfb10 AG |
1561 | if (S_ISDIR(inode_in->i_mode) || S_ISDIR(inode_out->i_mode)) |
1562 | return -EISDIR; | |
1563 | if (!S_ISREG(inode_in->i_mode) || !S_ISREG(inode_out->i_mode)) | |
1564 | return -EINVAL; | |
1565 | ||
29732938 | 1566 | ret = rw_verify_area(READ, file_in, &pos_in, len); |
bc61384d AV |
1567 | if (unlikely(ret)) |
1568 | return ret; | |
1569 | ||
1570 | ret = rw_verify_area(WRITE, file_out, &pos_out, len); | |
1571 | if (unlikely(ret)) | |
29732938 ZB |
1572 | return ret; |
1573 | ||
1574 | if (!(file_in->f_mode & FMODE_READ) || | |
1575 | !(file_out->f_mode & FMODE_WRITE) || | |
eac70053 | 1576 | (file_out->f_flags & O_APPEND)) |
29732938 ZB |
1577 | return -EBADF; |
1578 | ||
1579 | /* this could be relaxed once a method supports cross-fs copies */ | |
1580 | if (inode_in->i_sb != inode_out->i_sb) | |
1581 | return -EXDEV; | |
1582 | ||
1583 | if (len == 0) | |
1584 | return 0; | |
1585 | ||
bfe219d3 | 1586 | file_start_write(file_out); |
29732938 | 1587 | |
a76b5b04 CH |
1588 | /* |
1589 | * Try cloning first, this is supported by more file systems, and | |
1590 | * more efficient if both clone and copy are supported (e.g. NFS). | |
1591 | */ | |
1592 | if (file_in->f_op->clone_file_range) { | |
1593 | ret = file_in->f_op->clone_file_range(file_in, pos_in, | |
1594 | file_out, pos_out, len); | |
1595 | if (ret == 0) { | |
1596 | ret = len; | |
1597 | goto done; | |
1598 | } | |
1599 | } | |
1600 | ||
1601 | if (file_out->f_op->copy_file_range) { | |
eac70053 AS |
1602 | ret = file_out->f_op->copy_file_range(file_in, pos_in, file_out, |
1603 | pos_out, len, flags); | |
a76b5b04 CH |
1604 | if (ret != -EOPNOTSUPP) |
1605 | goto done; | |
1606 | } | |
eac70053 | 1607 | |
a76b5b04 CH |
1608 | ret = do_splice_direct(file_in, &pos_in, file_out, &pos_out, |
1609 | len > MAX_RW_COUNT ? MAX_RW_COUNT : len, 0); | |
eac70053 | 1610 | |
a76b5b04 | 1611 | done: |
29732938 ZB |
1612 | if (ret > 0) { |
1613 | fsnotify_access(file_in); | |
1614 | add_rchar(current, ret); | |
1615 | fsnotify_modify(file_out); | |
1616 | add_wchar(current, ret); | |
1617 | } | |
a76b5b04 | 1618 | |
29732938 ZB |
1619 | inc_syscr(current); |
1620 | inc_syscw(current); | |
1621 | ||
bfe219d3 | 1622 | file_end_write(file_out); |
29732938 ZB |
1623 | |
1624 | return ret; | |
1625 | } | |
1626 | EXPORT_SYMBOL(vfs_copy_file_range); | |
1627 | ||
1628 | SYSCALL_DEFINE6(copy_file_range, int, fd_in, loff_t __user *, off_in, | |
1629 | int, fd_out, loff_t __user *, off_out, | |
1630 | size_t, len, unsigned int, flags) | |
1631 | { | |
1632 | loff_t pos_in; | |
1633 | loff_t pos_out; | |
1634 | struct fd f_in; | |
1635 | struct fd f_out; | |
1636 | ssize_t ret = -EBADF; | |
1637 | ||
1638 | f_in = fdget(fd_in); | |
1639 | if (!f_in.file) | |
1640 | goto out2; | |
1641 | ||
1642 | f_out = fdget(fd_out); | |
1643 | if (!f_out.file) | |
1644 | goto out1; | |
1645 | ||
1646 | ret = -EFAULT; | |
1647 | if (off_in) { | |
1648 | if (copy_from_user(&pos_in, off_in, sizeof(loff_t))) | |
1649 | goto out; | |
1650 | } else { | |
1651 | pos_in = f_in.file->f_pos; | |
1652 | } | |
1653 | ||
1654 | if (off_out) { | |
1655 | if (copy_from_user(&pos_out, off_out, sizeof(loff_t))) | |
1656 | goto out; | |
1657 | } else { | |
1658 | pos_out = f_out.file->f_pos; | |
1659 | } | |
1660 | ||
1661 | ret = vfs_copy_file_range(f_in.file, pos_in, f_out.file, pos_out, len, | |
1662 | flags); | |
1663 | if (ret > 0) { | |
1664 | pos_in += ret; | |
1665 | pos_out += ret; | |
1666 | ||
1667 | if (off_in) { | |
1668 | if (copy_to_user(off_in, &pos_in, sizeof(loff_t))) | |
1669 | ret = -EFAULT; | |
1670 | } else { | |
1671 | f_in.file->f_pos = pos_in; | |
1672 | } | |
1673 | ||
1674 | if (off_out) { | |
1675 | if (copy_to_user(off_out, &pos_out, sizeof(loff_t))) | |
1676 | ret = -EFAULT; | |
1677 | } else { | |
1678 | f_out.file->f_pos = pos_out; | |
1679 | } | |
1680 | } | |
1681 | ||
1682 | out: | |
1683 | fdput(f_out); | |
1684 | out1: | |
1685 | fdput(f_in); | |
1686 | out2: | |
1687 | return ret; | |
1688 | } | |
04b38d60 CH |
1689 | |
1690 | static int clone_verify_area(struct file *file, loff_t pos, u64 len, bool write) | |
1691 | { | |
1692 | struct inode *inode = file_inode(file); | |
1693 | ||
1694 | if (unlikely(pos < 0)) | |
1695 | return -EINVAL; | |
1696 | ||
1697 | if (unlikely((loff_t) (pos + len) < 0)) | |
1698 | return -EINVAL; | |
1699 | ||
1700 | if (unlikely(inode->i_flctx && mandatory_lock(inode))) { | |
1701 | loff_t end = len ? pos + len - 1 : OFFSET_MAX; | |
1702 | int retval; | |
1703 | ||
1704 | retval = locks_mandatory_area(inode, file, pos, end, | |
1705 | write ? F_WRLCK : F_RDLCK); | |
1706 | if (retval < 0) | |
1707 | return retval; | |
1708 | } | |
1709 | ||
1710 | return security_file_permission(file, write ? MAY_WRITE : MAY_READ); | |
1711 | } | |
1712 | ||
876bec6f DW |
1713 | /* |
1714 | * Check that the two inodes are eligible for cloning, the ranges make | |
1715 | * sense, and then flush all dirty data. Caller must ensure that the | |
1716 | * inodes have been locked against any other modifications. | |
22725ce4 DW |
1717 | * |
1718 | * Returns: 0 for "nothing to clone", 1 for "something to clone", or | |
1719 | * the usual negative error code. | |
876bec6f DW |
1720 | */ |
1721 | int vfs_clone_file_prep_inodes(struct inode *inode_in, loff_t pos_in, | |
1722 | struct inode *inode_out, loff_t pos_out, | |
1723 | u64 *len, bool is_dedupe) | |
1724 | { | |
1725 | loff_t bs = inode_out->i_sb->s_blocksize; | |
1726 | loff_t blen; | |
1727 | loff_t isize; | |
1728 | bool same_inode = (inode_in == inode_out); | |
1729 | int ret; | |
1730 | ||
1731 | /* Don't touch certain kinds of inodes */ | |
1732 | if (IS_IMMUTABLE(inode_out)) | |
1733 | return -EPERM; | |
1734 | ||
1735 | if (IS_SWAPFILE(inode_in) || IS_SWAPFILE(inode_out)) | |
1736 | return -ETXTBSY; | |
1737 | ||
1738 | /* Don't reflink dirs, pipes, sockets... */ | |
1739 | if (S_ISDIR(inode_in->i_mode) || S_ISDIR(inode_out->i_mode)) | |
1740 | return -EISDIR; | |
1741 | if (!S_ISREG(inode_in->i_mode) || !S_ISREG(inode_out->i_mode)) | |
1742 | return -EINVAL; | |
1743 | ||
1744 | /* Are we going all the way to the end? */ | |
1745 | isize = i_size_read(inode_in); | |
22725ce4 | 1746 | if (isize == 0) |
876bec6f | 1747 | return 0; |
876bec6f DW |
1748 | |
1749 | /* Zero length dedupe exits immediately; reflink goes to EOF. */ | |
1750 | if (*len == 0) { | |
22725ce4 | 1751 | if (is_dedupe || pos_in == isize) |
876bec6f | 1752 | return 0; |
22725ce4 DW |
1753 | if (pos_in > isize) |
1754 | return -EINVAL; | |
876bec6f DW |
1755 | *len = isize - pos_in; |
1756 | } | |
1757 | ||
1758 | /* Ensure offsets don't wrap and the input is inside i_size */ | |
1759 | if (pos_in + *len < pos_in || pos_out + *len < pos_out || | |
1760 | pos_in + *len > isize) | |
1761 | return -EINVAL; | |
1762 | ||
1763 | /* Don't allow dedupe past EOF in the dest file */ | |
1764 | if (is_dedupe) { | |
1765 | loff_t disize; | |
1766 | ||
1767 | disize = i_size_read(inode_out); | |
1768 | if (pos_out >= disize || pos_out + *len > disize) | |
1769 | return -EINVAL; | |
1770 | } | |
1771 | ||
1772 | /* If we're linking to EOF, continue to the block boundary. */ | |
1773 | if (pos_in + *len == isize) | |
1774 | blen = ALIGN(isize, bs) - pos_in; | |
1775 | else | |
1776 | blen = *len; | |
1777 | ||
1778 | /* Only reflink if we're aligned to block boundaries */ | |
1779 | if (!IS_ALIGNED(pos_in, bs) || !IS_ALIGNED(pos_in + blen, bs) || | |
1780 | !IS_ALIGNED(pos_out, bs) || !IS_ALIGNED(pos_out + blen, bs)) | |
1781 | return -EINVAL; | |
1782 | ||
1783 | /* Don't allow overlapped reflink within the same file */ | |
1784 | if (same_inode) { | |
1785 | if (pos_out + blen > pos_in && pos_out < pos_in + blen) | |
1786 | return -EINVAL; | |
1787 | } | |
1788 | ||
1789 | /* Wait for the completion of any pending IOs on both files */ | |
1790 | inode_dio_wait(inode_in); | |
1791 | if (!same_inode) | |
1792 | inode_dio_wait(inode_out); | |
1793 | ||
1794 | ret = filemap_write_and_wait_range(inode_in->i_mapping, | |
1795 | pos_in, pos_in + *len - 1); | |
1796 | if (ret) | |
1797 | return ret; | |
1798 | ||
1799 | ret = filemap_write_and_wait_range(inode_out->i_mapping, | |
1800 | pos_out, pos_out + *len - 1); | |
1801 | if (ret) | |
1802 | return ret; | |
1803 | ||
1804 | /* | |
1805 | * Check that the extents are the same. | |
1806 | */ | |
1807 | if (is_dedupe) { | |
1808 | bool is_same = false; | |
1809 | ||
1810 | ret = vfs_dedupe_file_range_compare(inode_in, pos_in, | |
1811 | inode_out, pos_out, *len, &is_same); | |
1812 | if (ret) | |
1813 | return ret; | |
1814 | if (!is_same) | |
1815 | return -EBADE; | |
1816 | } | |
1817 | ||
22725ce4 | 1818 | return 1; |
876bec6f DW |
1819 | } |
1820 | EXPORT_SYMBOL(vfs_clone_file_prep_inodes); | |
1821 | ||
04b38d60 CH |
1822 | int vfs_clone_file_range(struct file *file_in, loff_t pos_in, |
1823 | struct file *file_out, loff_t pos_out, u64 len) | |
1824 | { | |
1825 | struct inode *inode_in = file_inode(file_in); | |
1826 | struct inode *inode_out = file_inode(file_out); | |
1827 | int ret; | |
1828 | ||
b335e9d9 AG |
1829 | if (S_ISDIR(inode_in->i_mode) || S_ISDIR(inode_out->i_mode)) |
1830 | return -EISDIR; | |
1831 | if (!S_ISREG(inode_in->i_mode) || !S_ISREG(inode_out->i_mode)) | |
1832 | return -EINVAL; | |
1833 | ||
913b86e9 AG |
1834 | /* |
1835 | * FICLONE/FICLONERANGE ioctls enforce that src and dest files are on | |
1836 | * the same mount. Practically, they only need to be on the same file | |
1837 | * system. | |
1838 | */ | |
1839 | if (inode_in->i_sb != inode_out->i_sb) | |
04b38d60 CH |
1840 | return -EXDEV; |
1841 | ||
04b38d60 CH |
1842 | if (!(file_in->f_mode & FMODE_READ) || |
1843 | !(file_out->f_mode & FMODE_WRITE) || | |
0fcbf996 | 1844 | (file_out->f_flags & O_APPEND)) |
04b38d60 CH |
1845 | return -EBADF; |
1846 | ||
0fcbf996 CH |
1847 | if (!file_in->f_op->clone_file_range) |
1848 | return -EOPNOTSUPP; | |
1849 | ||
04b38d60 CH |
1850 | ret = clone_verify_area(file_in, pos_in, len, false); |
1851 | if (ret) | |
1852 | return ret; | |
1853 | ||
1854 | ret = clone_verify_area(file_out, pos_out, len, true); | |
1855 | if (ret) | |
1856 | return ret; | |
1857 | ||
1858 | if (pos_in + len > i_size_read(inode_in)) | |
1859 | return -EINVAL; | |
1860 | ||
04b38d60 CH |
1861 | ret = file_in->f_op->clone_file_range(file_in, pos_in, |
1862 | file_out, pos_out, len); | |
1863 | if (!ret) { | |
1864 | fsnotify_access(file_in); | |
1865 | fsnotify_modify(file_out); | |
1866 | } | |
1867 | ||
04b38d60 CH |
1868 | return ret; |
1869 | } | |
1870 | EXPORT_SYMBOL(vfs_clone_file_range); | |
54dbc151 | 1871 | |
876bec6f DW |
1872 | /* |
1873 | * Read a page's worth of file data into the page cache. Return the page | |
1874 | * locked. | |
1875 | */ | |
1876 | static struct page *vfs_dedupe_get_page(struct inode *inode, loff_t offset) | |
1877 | { | |
1878 | struct address_space *mapping; | |
1879 | struct page *page; | |
1880 | pgoff_t n; | |
1881 | ||
1882 | n = offset >> PAGE_SHIFT; | |
1883 | mapping = inode->i_mapping; | |
1884 | page = read_mapping_page(mapping, n, NULL); | |
1885 | if (IS_ERR(page)) | |
1886 | return page; | |
1887 | if (!PageUptodate(page)) { | |
1888 | put_page(page); | |
1889 | return ERR_PTR(-EIO); | |
1890 | } | |
1891 | lock_page(page); | |
1892 | return page; | |
1893 | } | |
1894 | ||
1895 | /* | |
1896 | * Compare extents of two files to see if they are the same. | |
1897 | * Caller must have locked both inodes to prevent write races. | |
1898 | */ | |
1899 | int vfs_dedupe_file_range_compare(struct inode *src, loff_t srcoff, | |
1900 | struct inode *dest, loff_t destoff, | |
1901 | loff_t len, bool *is_same) | |
1902 | { | |
1903 | loff_t src_poff; | |
1904 | loff_t dest_poff; | |
1905 | void *src_addr; | |
1906 | void *dest_addr; | |
1907 | struct page *src_page; | |
1908 | struct page *dest_page; | |
1909 | loff_t cmp_len; | |
1910 | bool same; | |
1911 | int error; | |
1912 | ||
1913 | error = -EINVAL; | |
1914 | same = true; | |
1915 | while (len) { | |
1916 | src_poff = srcoff & (PAGE_SIZE - 1); | |
1917 | dest_poff = destoff & (PAGE_SIZE - 1); | |
1918 | cmp_len = min(PAGE_SIZE - src_poff, | |
1919 | PAGE_SIZE - dest_poff); | |
1920 | cmp_len = min(cmp_len, len); | |
1921 | if (cmp_len <= 0) | |
1922 | goto out_error; | |
1923 | ||
1924 | src_page = vfs_dedupe_get_page(src, srcoff); | |
1925 | if (IS_ERR(src_page)) { | |
1926 | error = PTR_ERR(src_page); | |
1927 | goto out_error; | |
1928 | } | |
1929 | dest_page = vfs_dedupe_get_page(dest, destoff); | |
1930 | if (IS_ERR(dest_page)) { | |
1931 | error = PTR_ERR(dest_page); | |
1932 | unlock_page(src_page); | |
1933 | put_page(src_page); | |
1934 | goto out_error; | |
1935 | } | |
1936 | src_addr = kmap_atomic(src_page); | |
1937 | dest_addr = kmap_atomic(dest_page); | |
1938 | ||
1939 | flush_dcache_page(src_page); | |
1940 | flush_dcache_page(dest_page); | |
1941 | ||
1942 | if (memcmp(src_addr + src_poff, dest_addr + dest_poff, cmp_len)) | |
1943 | same = false; | |
1944 | ||
1945 | kunmap_atomic(dest_addr); | |
1946 | kunmap_atomic(src_addr); | |
1947 | unlock_page(dest_page); | |
1948 | unlock_page(src_page); | |
1949 | put_page(dest_page); | |
1950 | put_page(src_page); | |
1951 | ||
1952 | if (!same) | |
1953 | break; | |
1954 | ||
1955 | srcoff += cmp_len; | |
1956 | destoff += cmp_len; | |
1957 | len -= cmp_len; | |
1958 | } | |
1959 | ||
1960 | *is_same = same; | |
1961 | return 0; | |
1962 | ||
1963 | out_error: | |
1964 | return error; | |
1965 | } | |
1966 | EXPORT_SYMBOL(vfs_dedupe_file_range_compare); | |
1967 | ||
54dbc151 DW |
1968 | int vfs_dedupe_file_range(struct file *file, struct file_dedupe_range *same) |
1969 | { | |
1970 | struct file_dedupe_range_info *info; | |
1971 | struct inode *src = file_inode(file); | |
1972 | u64 off; | |
1973 | u64 len; | |
1974 | int i; | |
1975 | int ret; | |
1976 | bool is_admin = capable(CAP_SYS_ADMIN); | |
1977 | u16 count = same->dest_count; | |
1978 | struct file *dst_file; | |
1979 | loff_t dst_off; | |
1980 | ssize_t deduped; | |
1981 | ||
1982 | if (!(file->f_mode & FMODE_READ)) | |
1983 | return -EINVAL; | |
1984 | ||
1985 | if (same->reserved1 || same->reserved2) | |
1986 | return -EINVAL; | |
1987 | ||
1988 | off = same->src_offset; | |
1989 | len = same->src_length; | |
1990 | ||
1991 | ret = -EISDIR; | |
1992 | if (S_ISDIR(src->i_mode)) | |
1993 | goto out; | |
1994 | ||
1995 | ret = -EINVAL; | |
1996 | if (!S_ISREG(src->i_mode)) | |
1997 | goto out; | |
1998 | ||
1999 | ret = clone_verify_area(file, off, len, false); | |
2000 | if (ret < 0) | |
2001 | goto out; | |
2002 | ret = 0; | |
2003 | ||
22725ce4 DW |
2004 | if (off + len > i_size_read(src)) |
2005 | return -EINVAL; | |
2006 | ||
54dbc151 DW |
2007 | /* pre-format output fields to sane values */ |
2008 | for (i = 0; i < count; i++) { | |
2009 | same->info[i].bytes_deduped = 0ULL; | |
2010 | same->info[i].status = FILE_DEDUPE_RANGE_SAME; | |
2011 | } | |
2012 | ||
2013 | for (i = 0, info = same->info; i < count; i++, info++) { | |
2014 | struct inode *dst; | |
2015 | struct fd dst_fd = fdget(info->dest_fd); | |
2016 | ||
2017 | dst_file = dst_fd.file; | |
2018 | if (!dst_file) { | |
2019 | info->status = -EBADF; | |
2020 | goto next_loop; | |
2021 | } | |
2022 | dst = file_inode(dst_file); | |
2023 | ||
2024 | ret = mnt_want_write_file(dst_file); | |
2025 | if (ret) { | |
2026 | info->status = ret; | |
2027 | goto next_loop; | |
2028 | } | |
2029 | ||
2030 | dst_off = info->dest_offset; | |
2031 | ret = clone_verify_area(dst_file, dst_off, len, true); | |
2032 | if (ret < 0) { | |
2033 | info->status = ret; | |
2034 | goto next_file; | |
2035 | } | |
2036 | ret = 0; | |
2037 | ||
2038 | if (info->reserved) { | |
2039 | info->status = -EINVAL; | |
2040 | } else if (!(is_admin || (dst_file->f_mode & FMODE_WRITE))) { | |
2041 | info->status = -EINVAL; | |
2042 | } else if (file->f_path.mnt != dst_file->f_path.mnt) { | |
2043 | info->status = -EXDEV; | |
2044 | } else if (S_ISDIR(dst->i_mode)) { | |
2045 | info->status = -EISDIR; | |
2046 | } else if (dst_file->f_op->dedupe_file_range == NULL) { | |
2047 | info->status = -EINVAL; | |
2048 | } else { | |
2049 | deduped = dst_file->f_op->dedupe_file_range(file, off, | |
2050 | len, dst_file, | |
2051 | info->dest_offset); | |
2052 | if (deduped == -EBADE) | |
2053 | info->status = FILE_DEDUPE_RANGE_DIFFERS; | |
2054 | else if (deduped < 0) | |
2055 | info->status = deduped; | |
2056 | else | |
2057 | info->bytes_deduped += deduped; | |
2058 | } | |
2059 | ||
2060 | next_file: | |
2061 | mnt_drop_write_file(dst_file); | |
2062 | next_loop: | |
2063 | fdput(dst_fd); | |
e62e560f DW |
2064 | |
2065 | if (fatal_signal_pending(current)) | |
2066 | goto out; | |
54dbc151 DW |
2067 | } |
2068 | ||
2069 | out: | |
2070 | return ret; | |
2071 | } | |
2072 | EXPORT_SYMBOL(vfs_dedupe_file_range); |