]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/fs/read_write.c | |
3 | * | |
4 | * Copyright (C) 1991, 1992 Linus Torvalds | |
5 | */ | |
6 | ||
7 | #include <linux/slab.h> | |
8 | #include <linux/stat.h> | |
9 | #include <linux/fcntl.h> | |
10 | #include <linux/file.h> | |
11 | #include <linux/uio.h> | |
0eeca283 | 12 | #include <linux/fsnotify.h> |
1da177e4 | 13 | #include <linux/security.h> |
630d9c47 | 14 | #include <linux/export.h> |
1da177e4 | 15 | #include <linux/syscalls.h> |
e28cc715 | 16 | #include <linux/pagemap.h> |
d6b29d7c | 17 | #include <linux/splice.h> |
561c6731 | 18 | #include <linux/compat.h> |
29732938 | 19 | #include <linux/mount.h> |
06ae43f3 | 20 | #include "internal.h" |
1da177e4 LT |
21 | |
22 | #include <asm/uaccess.h> | |
23 | #include <asm/unistd.h> | |
24 | ||
c0bd14af | 25 | typedef ssize_t (*io_fn_t)(struct file *, char __user *, size_t, loff_t *); |
293bc982 | 26 | typedef ssize_t (*iter_fn_t)(struct kiocb *, struct iov_iter *); |
c0bd14af | 27 | |
4b6f5d20 | 28 | const struct file_operations generic_ro_fops = { |
1da177e4 | 29 | .llseek = generic_file_llseek, |
aad4f8bb | 30 | .read_iter = generic_file_read_iter, |
1da177e4 | 31 | .mmap = generic_file_readonly_mmap, |
534f2aaa | 32 | .splice_read = generic_file_splice_read, |
1da177e4 LT |
33 | }; |
34 | ||
35 | EXPORT_SYMBOL(generic_ro_fops); | |
36 | ||
cccb5a1e | 37 | static inline int unsigned_offsets(struct file *file) |
4a3956c7 | 38 | { |
cccb5a1e | 39 | return file->f_mode & FMODE_UNSIGNED_OFFSET; |
4a3956c7 KH |
40 | } |
41 | ||
46a1c2c7 JL |
42 | /** |
43 | * vfs_setpos - update the file offset for lseek | |
44 | * @file: file structure in question | |
45 | * @offset: file offset to seek to | |
46 | * @maxsize: maximum file size | |
47 | * | |
48 | * This is a low-level filesystem helper for updating the file offset to | |
49 | * the value specified by @offset if the given offset is valid and it is | |
50 | * not equal to the current file offset. | |
51 | * | |
52 | * Return the specified offset on success and -EINVAL on invalid offset. | |
53 | */ | |
54 | loff_t vfs_setpos(struct file *file, loff_t offset, loff_t maxsize) | |
ef3d0fd2 AK |
55 | { |
56 | if (offset < 0 && !unsigned_offsets(file)) | |
57 | return -EINVAL; | |
58 | if (offset > maxsize) | |
59 | return -EINVAL; | |
60 | ||
61 | if (offset != file->f_pos) { | |
62 | file->f_pos = offset; | |
63 | file->f_version = 0; | |
64 | } | |
65 | return offset; | |
66 | } | |
46a1c2c7 | 67 | EXPORT_SYMBOL(vfs_setpos); |
ef3d0fd2 | 68 | |
3a8cff4f | 69 | /** |
5760495a | 70 | * generic_file_llseek_size - generic llseek implementation for regular files |
3a8cff4f CH |
71 | * @file: file structure to seek on |
72 | * @offset: file offset to seek to | |
965c8e59 | 73 | * @whence: type of seek |
e8b96eb5 ES |
74 | * @size: max size of this file in file system |
75 | * @eof: offset used for SEEK_END position | |
3a8cff4f | 76 | * |
5760495a | 77 | * This is a variant of generic_file_llseek that allows passing in a custom |
e8b96eb5 | 78 | * maximum file size and a custom EOF position, for e.g. hashed directories |
ef3d0fd2 AK |
79 | * |
80 | * Synchronization: | |
5760495a | 81 | * SEEK_SET and SEEK_END are unsynchronized (but atomic on 64bit platforms) |
ef3d0fd2 AK |
82 | * SEEK_CUR is synchronized against other SEEK_CURs, but not read/writes. |
83 | * read/writes behave like SEEK_SET against seeks. | |
3a8cff4f | 84 | */ |
9465efc9 | 85 | loff_t |
965c8e59 | 86 | generic_file_llseek_size(struct file *file, loff_t offset, int whence, |
e8b96eb5 | 87 | loff_t maxsize, loff_t eof) |
1da177e4 | 88 | { |
965c8e59 | 89 | switch (whence) { |
3a8cff4f | 90 | case SEEK_END: |
e8b96eb5 | 91 | offset += eof; |
3a8cff4f CH |
92 | break; |
93 | case SEEK_CUR: | |
5b6f1eb9 AK |
94 | /* |
95 | * Here we special-case the lseek(fd, 0, SEEK_CUR) | |
96 | * position-querying operation. Avoid rewriting the "same" | |
97 | * f_pos value back to the file because a concurrent read(), | |
98 | * write() or lseek() might have altered it | |
99 | */ | |
100 | if (offset == 0) | |
101 | return file->f_pos; | |
ef3d0fd2 AK |
102 | /* |
103 | * f_lock protects against read/modify/write race with other | |
104 | * SEEK_CURs. Note that parallel writes and reads behave | |
105 | * like SEEK_SET. | |
106 | */ | |
107 | spin_lock(&file->f_lock); | |
46a1c2c7 | 108 | offset = vfs_setpos(file, file->f_pos + offset, maxsize); |
ef3d0fd2 AK |
109 | spin_unlock(&file->f_lock); |
110 | return offset; | |
982d8165 JB |
111 | case SEEK_DATA: |
112 | /* | |
113 | * In the generic case the entire file is data, so as long as | |
114 | * offset isn't at the end of the file then the offset is data. | |
115 | */ | |
e8b96eb5 | 116 | if (offset >= eof) |
982d8165 JB |
117 | return -ENXIO; |
118 | break; | |
119 | case SEEK_HOLE: | |
120 | /* | |
121 | * There is a virtual hole at the end of the file, so as long as | |
122 | * offset isn't i_size or larger, return i_size. | |
123 | */ | |
e8b96eb5 | 124 | if (offset >= eof) |
982d8165 | 125 | return -ENXIO; |
e8b96eb5 | 126 | offset = eof; |
982d8165 | 127 | break; |
1da177e4 | 128 | } |
3a8cff4f | 129 | |
46a1c2c7 | 130 | return vfs_setpos(file, offset, maxsize); |
5760495a AK |
131 | } |
132 | EXPORT_SYMBOL(generic_file_llseek_size); | |
133 | ||
134 | /** | |
135 | * generic_file_llseek - generic llseek implementation for regular files | |
136 | * @file: file structure to seek on | |
137 | * @offset: file offset to seek to | |
965c8e59 | 138 | * @whence: type of seek |
5760495a AK |
139 | * |
140 | * This is a generic implemenation of ->llseek useable for all normal local | |
141 | * filesystems. It just updates the file offset to the value specified by | |
546ae2d2 | 142 | * @offset and @whence. |
5760495a | 143 | */ |
965c8e59 | 144 | loff_t generic_file_llseek(struct file *file, loff_t offset, int whence) |
5760495a AK |
145 | { |
146 | struct inode *inode = file->f_mapping->host; | |
147 | ||
965c8e59 | 148 | return generic_file_llseek_size(file, offset, whence, |
e8b96eb5 ES |
149 | inode->i_sb->s_maxbytes, |
150 | i_size_read(inode)); | |
1da177e4 | 151 | } |
9465efc9 | 152 | EXPORT_SYMBOL(generic_file_llseek); |
1da177e4 | 153 | |
1bf9d14d AV |
154 | /** |
155 | * fixed_size_llseek - llseek implementation for fixed-sized devices | |
156 | * @file: file structure to seek on | |
157 | * @offset: file offset to seek to | |
158 | * @whence: type of seek | |
159 | * @size: size of the file | |
160 | * | |
161 | */ | |
162 | loff_t fixed_size_llseek(struct file *file, loff_t offset, int whence, loff_t size) | |
163 | { | |
164 | switch (whence) { | |
165 | case SEEK_SET: case SEEK_CUR: case SEEK_END: | |
166 | return generic_file_llseek_size(file, offset, whence, | |
167 | size, size); | |
168 | default: | |
169 | return -EINVAL; | |
170 | } | |
171 | } | |
172 | EXPORT_SYMBOL(fixed_size_llseek); | |
173 | ||
ae6afc3f B |
174 | /** |
175 | * noop_llseek - No Operation Performed llseek implementation | |
176 | * @file: file structure to seek on | |
177 | * @offset: file offset to seek to | |
965c8e59 | 178 | * @whence: type of seek |
ae6afc3f B |
179 | * |
180 | * This is an implementation of ->llseek useable for the rare special case when | |
181 | * userspace expects the seek to succeed but the (device) file is actually not | |
182 | * able to perform the seek. In this case you use noop_llseek() instead of | |
183 | * falling back to the default implementation of ->llseek. | |
184 | */ | |
965c8e59 | 185 | loff_t noop_llseek(struct file *file, loff_t offset, int whence) |
ae6afc3f B |
186 | { |
187 | return file->f_pos; | |
188 | } | |
189 | EXPORT_SYMBOL(noop_llseek); | |
190 | ||
965c8e59 | 191 | loff_t no_llseek(struct file *file, loff_t offset, int whence) |
1da177e4 LT |
192 | { |
193 | return -ESPIPE; | |
194 | } | |
195 | EXPORT_SYMBOL(no_llseek); | |
196 | ||
965c8e59 | 197 | loff_t default_llseek(struct file *file, loff_t offset, int whence) |
1da177e4 | 198 | { |
496ad9aa | 199 | struct inode *inode = file_inode(file); |
16abef0e | 200 | loff_t retval; |
1da177e4 | 201 | |
982d8165 | 202 | mutex_lock(&inode->i_mutex); |
965c8e59 | 203 | switch (whence) { |
7b8e8924 | 204 | case SEEK_END: |
982d8165 | 205 | offset += i_size_read(inode); |
1da177e4 | 206 | break; |
7b8e8924 | 207 | case SEEK_CUR: |
5b6f1eb9 AK |
208 | if (offset == 0) { |
209 | retval = file->f_pos; | |
210 | goto out; | |
211 | } | |
1da177e4 | 212 | offset += file->f_pos; |
982d8165 JB |
213 | break; |
214 | case SEEK_DATA: | |
215 | /* | |
216 | * In the generic case the entire file is data, so as | |
217 | * long as offset isn't at the end of the file then the | |
218 | * offset is data. | |
219 | */ | |
bacb2d81 DC |
220 | if (offset >= inode->i_size) { |
221 | retval = -ENXIO; | |
222 | goto out; | |
223 | } | |
982d8165 JB |
224 | break; |
225 | case SEEK_HOLE: | |
226 | /* | |
227 | * There is a virtual hole at the end of the file, so | |
228 | * as long as offset isn't i_size or larger, return | |
229 | * i_size. | |
230 | */ | |
bacb2d81 DC |
231 | if (offset >= inode->i_size) { |
232 | retval = -ENXIO; | |
233 | goto out; | |
234 | } | |
982d8165 JB |
235 | offset = inode->i_size; |
236 | break; | |
1da177e4 LT |
237 | } |
238 | retval = -EINVAL; | |
cccb5a1e | 239 | if (offset >= 0 || unsigned_offsets(file)) { |
1da177e4 LT |
240 | if (offset != file->f_pos) { |
241 | file->f_pos = offset; | |
242 | file->f_version = 0; | |
243 | } | |
244 | retval = offset; | |
245 | } | |
5b6f1eb9 | 246 | out: |
982d8165 | 247 | mutex_unlock(&inode->i_mutex); |
1da177e4 LT |
248 | return retval; |
249 | } | |
250 | EXPORT_SYMBOL(default_llseek); | |
251 | ||
965c8e59 | 252 | loff_t vfs_llseek(struct file *file, loff_t offset, int whence) |
1da177e4 LT |
253 | { |
254 | loff_t (*fn)(struct file *, loff_t, int); | |
255 | ||
256 | fn = no_llseek; | |
257 | if (file->f_mode & FMODE_LSEEK) { | |
72c2d531 | 258 | if (file->f_op->llseek) |
1da177e4 LT |
259 | fn = file->f_op->llseek; |
260 | } | |
965c8e59 | 261 | return fn(file, offset, whence); |
1da177e4 LT |
262 | } |
263 | EXPORT_SYMBOL(vfs_llseek); | |
264 | ||
9c225f26 LT |
265 | static inline struct fd fdget_pos(int fd) |
266 | { | |
bd2a31d5 | 267 | return __to_fd(__fdget_pos(fd)); |
9c225f26 LT |
268 | } |
269 | ||
270 | static inline void fdput_pos(struct fd f) | |
271 | { | |
272 | if (f.flags & FDPUT_POS_UNLOCK) | |
273 | mutex_unlock(&f.file->f_pos_lock); | |
274 | fdput(f); | |
275 | } | |
276 | ||
965c8e59 | 277 | SYSCALL_DEFINE3(lseek, unsigned int, fd, off_t, offset, unsigned int, whence) |
1da177e4 LT |
278 | { |
279 | off_t retval; | |
9c225f26 | 280 | struct fd f = fdget_pos(fd); |
2903ff01 AV |
281 | if (!f.file) |
282 | return -EBADF; | |
1da177e4 LT |
283 | |
284 | retval = -EINVAL; | |
965c8e59 AM |
285 | if (whence <= SEEK_MAX) { |
286 | loff_t res = vfs_llseek(f.file, offset, whence); | |
1da177e4 LT |
287 | retval = res; |
288 | if (res != (loff_t)retval) | |
289 | retval = -EOVERFLOW; /* LFS: should only happen on 32 bit platforms */ | |
290 | } | |
9c225f26 | 291 | fdput_pos(f); |
1da177e4 LT |
292 | return retval; |
293 | } | |
294 | ||
561c6731 AV |
295 | #ifdef CONFIG_COMPAT |
296 | COMPAT_SYSCALL_DEFINE3(lseek, unsigned int, fd, compat_off_t, offset, unsigned int, whence) | |
297 | { | |
298 | return sys_lseek(fd, offset, whence); | |
299 | } | |
300 | #endif | |
301 | ||
1da177e4 | 302 | #ifdef __ARCH_WANT_SYS_LLSEEK |
003d7ab4 HC |
303 | SYSCALL_DEFINE5(llseek, unsigned int, fd, unsigned long, offset_high, |
304 | unsigned long, offset_low, loff_t __user *, result, | |
965c8e59 | 305 | unsigned int, whence) |
1da177e4 LT |
306 | { |
307 | int retval; | |
d7a15f8d | 308 | struct fd f = fdget_pos(fd); |
1da177e4 | 309 | loff_t offset; |
1da177e4 | 310 | |
2903ff01 AV |
311 | if (!f.file) |
312 | return -EBADF; | |
1da177e4 LT |
313 | |
314 | retval = -EINVAL; | |
965c8e59 | 315 | if (whence > SEEK_MAX) |
1da177e4 LT |
316 | goto out_putf; |
317 | ||
2903ff01 | 318 | offset = vfs_llseek(f.file, ((loff_t) offset_high << 32) | offset_low, |
965c8e59 | 319 | whence); |
1da177e4 LT |
320 | |
321 | retval = (int)offset; | |
322 | if (offset >= 0) { | |
323 | retval = -EFAULT; | |
324 | if (!copy_to_user(result, &offset, sizeof(offset))) | |
325 | retval = 0; | |
326 | } | |
327 | out_putf: | |
d7a15f8d | 328 | fdput_pos(f); |
1da177e4 LT |
329 | return retval; |
330 | } | |
331 | #endif | |
332 | ||
dbe4e192 CH |
333 | ssize_t vfs_iter_read(struct file *file, struct iov_iter *iter, loff_t *ppos) |
334 | { | |
335 | struct kiocb kiocb; | |
336 | ssize_t ret; | |
337 | ||
338 | if (!file->f_op->read_iter) | |
339 | return -EINVAL; | |
340 | ||
341 | init_sync_kiocb(&kiocb, file); | |
342 | kiocb.ki_pos = *ppos; | |
dbe4e192 CH |
343 | |
344 | iter->type |= READ; | |
345 | ret = file->f_op->read_iter(&kiocb, iter); | |
599bd19b | 346 | BUG_ON(ret == -EIOCBQUEUED); |
dbe4e192 CH |
347 | if (ret > 0) |
348 | *ppos = kiocb.ki_pos; | |
349 | return ret; | |
350 | } | |
351 | EXPORT_SYMBOL(vfs_iter_read); | |
352 | ||
353 | ssize_t vfs_iter_write(struct file *file, struct iov_iter *iter, loff_t *ppos) | |
354 | { | |
355 | struct kiocb kiocb; | |
356 | ssize_t ret; | |
357 | ||
358 | if (!file->f_op->write_iter) | |
359 | return -EINVAL; | |
360 | ||
361 | init_sync_kiocb(&kiocb, file); | |
362 | kiocb.ki_pos = *ppos; | |
dbe4e192 CH |
363 | |
364 | iter->type |= WRITE; | |
365 | ret = file->f_op->write_iter(&kiocb, iter); | |
599bd19b | 366 | BUG_ON(ret == -EIOCBQUEUED); |
dbe4e192 CH |
367 | if (ret > 0) |
368 | *ppos = kiocb.ki_pos; | |
369 | return ret; | |
370 | } | |
371 | EXPORT_SYMBOL(vfs_iter_write); | |
372 | ||
e28cc715 LT |
373 | /* |
374 | * rw_verify_area doesn't like huge counts. We limit | |
375 | * them to something that fits in "int" so that others | |
376 | * won't have to do range checks all the time. | |
377 | */ | |
68d70d03 | 378 | int rw_verify_area(int read_write, struct file *file, const loff_t *ppos, size_t count) |
1da177e4 LT |
379 | { |
380 | struct inode *inode; | |
381 | loff_t pos; | |
c43e259c | 382 | int retval = -EINVAL; |
1da177e4 | 383 | |
496ad9aa | 384 | inode = file_inode(file); |
e28cc715 | 385 | if (unlikely((ssize_t) count < 0)) |
c43e259c | 386 | return retval; |
1da177e4 | 387 | pos = *ppos; |
cccb5a1e AV |
388 | if (unlikely(pos < 0)) { |
389 | if (!unsigned_offsets(file)) | |
390 | return retval; | |
391 | if (count >= -pos) /* both values are in 0..LLONG_MAX */ | |
392 | return -EOVERFLOW; | |
393 | } else if (unlikely((loff_t) (pos + count) < 0)) { | |
394 | if (!unsigned_offsets(file)) | |
4a3956c7 KH |
395 | return retval; |
396 | } | |
1da177e4 | 397 | |
bd61e0a9 | 398 | if (unlikely(inode->i_flctx && mandatory_lock(inode))) { |
c43e259c | 399 | retval = locks_mandatory_area( |
e28cc715 LT |
400 | read_write == READ ? FLOCK_VERIFY_READ : FLOCK_VERIFY_WRITE, |
401 | inode, file, pos, count); | |
402 | if (retval < 0) | |
403 | return retval; | |
404 | } | |
c43e259c JM |
405 | retval = security_file_permission(file, |
406 | read_write == READ ? MAY_READ : MAY_WRITE); | |
407 | if (retval) | |
408 | return retval; | |
e28cc715 | 409 | return count > MAX_RW_COUNT ? MAX_RW_COUNT : count; |
1da177e4 LT |
410 | } |
411 | ||
5d5d5689 | 412 | static ssize_t new_sync_read(struct file *filp, char __user *buf, size_t len, loff_t *ppos) |
293bc982 AV |
413 | { |
414 | struct iovec iov = { .iov_base = buf, .iov_len = len }; | |
415 | struct kiocb kiocb; | |
416 | struct iov_iter iter; | |
417 | ssize_t ret; | |
418 | ||
419 | init_sync_kiocb(&kiocb, filp); | |
420 | kiocb.ki_pos = *ppos; | |
293bc982 AV |
421 | iov_iter_init(&iter, READ, &iov, 1, len); |
422 | ||
423 | ret = filp->f_op->read_iter(&kiocb, &iter); | |
599bd19b | 424 | BUG_ON(ret == -EIOCBQUEUED); |
293bc982 AV |
425 | *ppos = kiocb.ki_pos; |
426 | return ret; | |
427 | } | |
428 | ||
6fb5032e DK |
429 | ssize_t __vfs_read(struct file *file, char __user *buf, size_t count, |
430 | loff_t *pos) | |
431 | { | |
6fb5032e | 432 | if (file->f_op->read) |
3d04c8a1 | 433 | return file->f_op->read(file, buf, count, pos); |
6fb5032e | 434 | else if (file->f_op->read_iter) |
3d04c8a1 | 435 | return new_sync_read(file, buf, count, pos); |
6fb5032e | 436 | else |
3d04c8a1 | 437 | return -EINVAL; |
6fb5032e | 438 | } |
3d04c8a1 | 439 | EXPORT_SYMBOL(__vfs_read); |
6fb5032e | 440 | |
1da177e4 LT |
441 | ssize_t vfs_read(struct file *file, char __user *buf, size_t count, loff_t *pos) |
442 | { | |
443 | ssize_t ret; | |
444 | ||
445 | if (!(file->f_mode & FMODE_READ)) | |
446 | return -EBADF; | |
7f7f25e8 | 447 | if (!(file->f_mode & FMODE_CAN_READ)) |
1da177e4 LT |
448 | return -EINVAL; |
449 | if (unlikely(!access_ok(VERIFY_WRITE, buf, count))) | |
450 | return -EFAULT; | |
451 | ||
452 | ret = rw_verify_area(READ, file, pos, count); | |
e28cc715 LT |
453 | if (ret >= 0) { |
454 | count = ret; | |
6fb5032e | 455 | ret = __vfs_read(file, buf, count, pos); |
c43e259c | 456 | if (ret > 0) { |
2a12a9d7 | 457 | fsnotify_access(file); |
c43e259c | 458 | add_rchar(current, ret); |
1da177e4 | 459 | } |
c43e259c | 460 | inc_syscr(current); |
1da177e4 LT |
461 | } |
462 | ||
463 | return ret; | |
464 | } | |
465 | ||
466 | EXPORT_SYMBOL(vfs_read); | |
467 | ||
5d5d5689 | 468 | static ssize_t new_sync_write(struct file *filp, const char __user *buf, size_t len, loff_t *ppos) |
293bc982 AV |
469 | { |
470 | struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = len }; | |
471 | struct kiocb kiocb; | |
472 | struct iov_iter iter; | |
473 | ssize_t ret; | |
474 | ||
475 | init_sync_kiocb(&kiocb, filp); | |
476 | kiocb.ki_pos = *ppos; | |
293bc982 AV |
477 | iov_iter_init(&iter, WRITE, &iov, 1, len); |
478 | ||
479 | ret = filp->f_op->write_iter(&kiocb, &iter); | |
599bd19b | 480 | BUG_ON(ret == -EIOCBQUEUED); |
f765b134 AV |
481 | if (ret > 0) |
482 | *ppos = kiocb.ki_pos; | |
293bc982 AV |
483 | return ret; |
484 | } | |
485 | ||
493c84c0 AV |
486 | ssize_t __vfs_write(struct file *file, const char __user *p, size_t count, |
487 | loff_t *pos) | |
488 | { | |
489 | if (file->f_op->write) | |
490 | return file->f_op->write(file, p, count, pos); | |
493c84c0 AV |
491 | else if (file->f_op->write_iter) |
492 | return new_sync_write(file, p, count, pos); | |
493 | else | |
494 | return -EINVAL; | |
495 | } | |
496 | EXPORT_SYMBOL(__vfs_write); | |
497 | ||
06ae43f3 AV |
498 | ssize_t __kernel_write(struct file *file, const char *buf, size_t count, loff_t *pos) |
499 | { | |
500 | mm_segment_t old_fs; | |
501 | const char __user *p; | |
502 | ssize_t ret; | |
503 | ||
7f7f25e8 | 504 | if (!(file->f_mode & FMODE_CAN_WRITE)) |
3e84f48e AV |
505 | return -EINVAL; |
506 | ||
06ae43f3 AV |
507 | old_fs = get_fs(); |
508 | set_fs(get_ds()); | |
509 | p = (__force const char __user *)buf; | |
510 | if (count > MAX_RW_COUNT) | |
511 | count = MAX_RW_COUNT; | |
493c84c0 | 512 | ret = __vfs_write(file, p, count, pos); |
06ae43f3 AV |
513 | set_fs(old_fs); |
514 | if (ret > 0) { | |
515 | fsnotify_modify(file); | |
516 | add_wchar(current, ret); | |
517 | } | |
518 | inc_syscw(current); | |
519 | return ret; | |
520 | } | |
521 | ||
2ec3a12a AV |
522 | EXPORT_SYMBOL(__kernel_write); |
523 | ||
1da177e4 LT |
524 | ssize_t vfs_write(struct file *file, const char __user *buf, size_t count, loff_t *pos) |
525 | { | |
526 | ssize_t ret; | |
527 | ||
528 | if (!(file->f_mode & FMODE_WRITE)) | |
529 | return -EBADF; | |
7f7f25e8 | 530 | if (!(file->f_mode & FMODE_CAN_WRITE)) |
1da177e4 LT |
531 | return -EINVAL; |
532 | if (unlikely(!access_ok(VERIFY_READ, buf, count))) | |
533 | return -EFAULT; | |
534 | ||
535 | ret = rw_verify_area(WRITE, file, pos, count); | |
e28cc715 LT |
536 | if (ret >= 0) { |
537 | count = ret; | |
03d95eb2 | 538 | file_start_write(file); |
493c84c0 | 539 | ret = __vfs_write(file, buf, count, pos); |
c43e259c | 540 | if (ret > 0) { |
2a12a9d7 | 541 | fsnotify_modify(file); |
c43e259c | 542 | add_wchar(current, ret); |
1da177e4 | 543 | } |
c43e259c | 544 | inc_syscw(current); |
03d95eb2 | 545 | file_end_write(file); |
1da177e4 LT |
546 | } |
547 | ||
548 | return ret; | |
549 | } | |
550 | ||
551 | EXPORT_SYMBOL(vfs_write); | |
552 | ||
553 | static inline loff_t file_pos_read(struct file *file) | |
554 | { | |
555 | return file->f_pos; | |
556 | } | |
557 | ||
558 | static inline void file_pos_write(struct file *file, loff_t pos) | |
559 | { | |
560 | file->f_pos = pos; | |
561 | } | |
562 | ||
3cdad428 | 563 | SYSCALL_DEFINE3(read, unsigned int, fd, char __user *, buf, size_t, count) |
1da177e4 | 564 | { |
9c225f26 | 565 | struct fd f = fdget_pos(fd); |
1da177e4 | 566 | ssize_t ret = -EBADF; |
1da177e4 | 567 | |
2903ff01 AV |
568 | if (f.file) { |
569 | loff_t pos = file_pos_read(f.file); | |
570 | ret = vfs_read(f.file, buf, count, &pos); | |
5faf153e AV |
571 | if (ret >= 0) |
572 | file_pos_write(f.file, pos); | |
9c225f26 | 573 | fdput_pos(f); |
1da177e4 | 574 | } |
1da177e4 LT |
575 | return ret; |
576 | } | |
1da177e4 | 577 | |
3cdad428 HC |
578 | SYSCALL_DEFINE3(write, unsigned int, fd, const char __user *, buf, |
579 | size_t, count) | |
1da177e4 | 580 | { |
9c225f26 | 581 | struct fd f = fdget_pos(fd); |
1da177e4 | 582 | ssize_t ret = -EBADF; |
1da177e4 | 583 | |
2903ff01 AV |
584 | if (f.file) { |
585 | loff_t pos = file_pos_read(f.file); | |
586 | ret = vfs_write(f.file, buf, count, &pos); | |
5faf153e AV |
587 | if (ret >= 0) |
588 | file_pos_write(f.file, pos); | |
9c225f26 | 589 | fdput_pos(f); |
1da177e4 LT |
590 | } |
591 | ||
592 | return ret; | |
593 | } | |
594 | ||
4a0fd5bf AV |
595 | SYSCALL_DEFINE4(pread64, unsigned int, fd, char __user *, buf, |
596 | size_t, count, loff_t, pos) | |
1da177e4 | 597 | { |
2903ff01 | 598 | struct fd f; |
1da177e4 | 599 | ssize_t ret = -EBADF; |
1da177e4 LT |
600 | |
601 | if (pos < 0) | |
602 | return -EINVAL; | |
603 | ||
2903ff01 AV |
604 | f = fdget(fd); |
605 | if (f.file) { | |
1da177e4 | 606 | ret = -ESPIPE; |
2903ff01 AV |
607 | if (f.file->f_mode & FMODE_PREAD) |
608 | ret = vfs_read(f.file, buf, count, &pos); | |
609 | fdput(f); | |
1da177e4 LT |
610 | } |
611 | ||
612 | return ret; | |
613 | } | |
614 | ||
4a0fd5bf AV |
615 | SYSCALL_DEFINE4(pwrite64, unsigned int, fd, const char __user *, buf, |
616 | size_t, count, loff_t, pos) | |
1da177e4 | 617 | { |
2903ff01 | 618 | struct fd f; |
1da177e4 | 619 | ssize_t ret = -EBADF; |
1da177e4 LT |
620 | |
621 | if (pos < 0) | |
622 | return -EINVAL; | |
623 | ||
2903ff01 AV |
624 | f = fdget(fd); |
625 | if (f.file) { | |
1da177e4 | 626 | ret = -ESPIPE; |
2903ff01 AV |
627 | if (f.file->f_mode & FMODE_PWRITE) |
628 | ret = vfs_write(f.file, buf, count, &pos); | |
629 | fdput(f); | |
1da177e4 LT |
630 | } |
631 | ||
632 | return ret; | |
633 | } | |
634 | ||
635 | /* | |
636 | * Reduce an iovec's length in-place. Return the resulting number of segments | |
637 | */ | |
638 | unsigned long iov_shorten(struct iovec *iov, unsigned long nr_segs, size_t to) | |
639 | { | |
640 | unsigned long seg = 0; | |
641 | size_t len = 0; | |
642 | ||
643 | while (seg < nr_segs) { | |
644 | seg++; | |
645 | if (len + iov->iov_len >= to) { | |
646 | iov->iov_len = to - len; | |
647 | break; | |
648 | } | |
649 | len += iov->iov_len; | |
650 | iov++; | |
651 | } | |
652 | return seg; | |
653 | } | |
19295529 | 654 | EXPORT_SYMBOL(iov_shorten); |
1da177e4 | 655 | |
ac15ac06 AV |
656 | static ssize_t do_iter_readv_writev(struct file *filp, struct iov_iter *iter, |
657 | loff_t *ppos, iter_fn_t fn) | |
293bc982 AV |
658 | { |
659 | struct kiocb kiocb; | |
293bc982 AV |
660 | ssize_t ret; |
661 | ||
662 | init_sync_kiocb(&kiocb, filp); | |
663 | kiocb.ki_pos = *ppos; | |
293bc982 | 664 | |
ac15ac06 | 665 | ret = fn(&kiocb, iter); |
599bd19b | 666 | BUG_ON(ret == -EIOCBQUEUED); |
293bc982 AV |
667 | *ppos = kiocb.ki_pos; |
668 | return ret; | |
669 | } | |
670 | ||
ee0b3e67 | 671 | /* Do it by hand, with file-ops */ |
ac15ac06 AV |
672 | static ssize_t do_loop_readv_writev(struct file *filp, struct iov_iter *iter, |
673 | loff_t *ppos, io_fn_t fn) | |
ee0b3e67 | 674 | { |
ee0b3e67 BP |
675 | ssize_t ret = 0; |
676 | ||
ac15ac06 AV |
677 | while (iov_iter_count(iter)) { |
678 | struct iovec iovec = iov_iter_iovec(iter); | |
ee0b3e67 BP |
679 | ssize_t nr; |
680 | ||
ac15ac06 | 681 | nr = fn(filp, iovec.iov_base, iovec.iov_len, ppos); |
ee0b3e67 BP |
682 | |
683 | if (nr < 0) { | |
684 | if (!ret) | |
685 | ret = nr; | |
686 | break; | |
687 | } | |
688 | ret += nr; | |
ac15ac06 | 689 | if (nr != iovec.iov_len) |
ee0b3e67 | 690 | break; |
ac15ac06 | 691 | iov_iter_advance(iter, nr); |
ee0b3e67 BP |
692 | } |
693 | ||
694 | return ret; | |
695 | } | |
696 | ||
1da177e4 LT |
697 | /* A write operation does a read from user space and vice versa */ |
698 | #define vrfy_dir(type) ((type) == READ ? VERIFY_WRITE : VERIFY_READ) | |
699 | ||
eed4e51f BP |
700 | ssize_t rw_copy_check_uvector(int type, const struct iovec __user * uvector, |
701 | unsigned long nr_segs, unsigned long fast_segs, | |
702 | struct iovec *fast_pointer, | |
ac34ebb3 | 703 | struct iovec **ret_pointer) |
435f49a5 | 704 | { |
eed4e51f | 705 | unsigned long seg; |
435f49a5 | 706 | ssize_t ret; |
eed4e51f BP |
707 | struct iovec *iov = fast_pointer; |
708 | ||
435f49a5 LT |
709 | /* |
710 | * SuS says "The readv() function *may* fail if the iovcnt argument | |
711 | * was less than or equal to 0, or greater than {IOV_MAX}. Linux has | |
712 | * traditionally returned zero for zero segments, so... | |
713 | */ | |
eed4e51f BP |
714 | if (nr_segs == 0) { |
715 | ret = 0; | |
435f49a5 | 716 | goto out; |
eed4e51f BP |
717 | } |
718 | ||
435f49a5 LT |
719 | /* |
720 | * First get the "struct iovec" from user memory and | |
721 | * verify all the pointers | |
722 | */ | |
eed4e51f BP |
723 | if (nr_segs > UIO_MAXIOV) { |
724 | ret = -EINVAL; | |
435f49a5 | 725 | goto out; |
eed4e51f BP |
726 | } |
727 | if (nr_segs > fast_segs) { | |
435f49a5 | 728 | iov = kmalloc(nr_segs*sizeof(struct iovec), GFP_KERNEL); |
eed4e51f BP |
729 | if (iov == NULL) { |
730 | ret = -ENOMEM; | |
435f49a5 | 731 | goto out; |
eed4e51f | 732 | } |
435f49a5 | 733 | } |
eed4e51f BP |
734 | if (copy_from_user(iov, uvector, nr_segs*sizeof(*uvector))) { |
735 | ret = -EFAULT; | |
435f49a5 | 736 | goto out; |
eed4e51f BP |
737 | } |
738 | ||
435f49a5 | 739 | /* |
eed4e51f BP |
740 | * According to the Single Unix Specification we should return EINVAL |
741 | * if an element length is < 0 when cast to ssize_t or if the | |
742 | * total length would overflow the ssize_t return value of the | |
743 | * system call. | |
435f49a5 LT |
744 | * |
745 | * Linux caps all read/write calls to MAX_RW_COUNT, and avoids the | |
746 | * overflow case. | |
747 | */ | |
eed4e51f | 748 | ret = 0; |
435f49a5 LT |
749 | for (seg = 0; seg < nr_segs; seg++) { |
750 | void __user *buf = iov[seg].iov_base; | |
751 | ssize_t len = (ssize_t)iov[seg].iov_len; | |
eed4e51f BP |
752 | |
753 | /* see if we we're about to use an invalid len or if | |
754 | * it's about to overflow ssize_t */ | |
435f49a5 | 755 | if (len < 0) { |
eed4e51f | 756 | ret = -EINVAL; |
435f49a5 | 757 | goto out; |
eed4e51f | 758 | } |
ac34ebb3 | 759 | if (type >= 0 |
fcf63409 | 760 | && unlikely(!access_ok(vrfy_dir(type), buf, len))) { |
eed4e51f | 761 | ret = -EFAULT; |
435f49a5 LT |
762 | goto out; |
763 | } | |
764 | if (len > MAX_RW_COUNT - ret) { | |
765 | len = MAX_RW_COUNT - ret; | |
766 | iov[seg].iov_len = len; | |
eed4e51f | 767 | } |
eed4e51f | 768 | ret += len; |
435f49a5 | 769 | } |
eed4e51f BP |
770 | out: |
771 | *ret_pointer = iov; | |
772 | return ret; | |
773 | } | |
774 | ||
1da177e4 LT |
775 | static ssize_t do_readv_writev(int type, struct file *file, |
776 | const struct iovec __user * uvector, | |
777 | unsigned long nr_segs, loff_t *pos) | |
778 | { | |
1da177e4 LT |
779 | size_t tot_len; |
780 | struct iovec iovstack[UIO_FASTIOV]; | |
ee0b3e67 | 781 | struct iovec *iov = iovstack; |
ac15ac06 | 782 | struct iov_iter iter; |
1da177e4 | 783 | ssize_t ret; |
1da177e4 | 784 | io_fn_t fn; |
293bc982 | 785 | iter_fn_t iter_fn; |
1da177e4 | 786 | |
0504c074 AV |
787 | ret = import_iovec(type, uvector, nr_segs, |
788 | ARRAY_SIZE(iovstack), &iov, &iter); | |
789 | if (ret < 0) | |
790 | return ret; | |
1da177e4 | 791 | |
0504c074 AV |
792 | tot_len = iov_iter_count(&iter); |
793 | if (!tot_len) | |
794 | goto out; | |
1da177e4 | 795 | ret = rw_verify_area(type, file, pos, tot_len); |
e28cc715 | 796 | if (ret < 0) |
411b67b4 | 797 | goto out; |
1da177e4 | 798 | |
1da177e4 LT |
799 | if (type == READ) { |
800 | fn = file->f_op->read; | |
293bc982 | 801 | iter_fn = file->f_op->read_iter; |
1da177e4 LT |
802 | } else { |
803 | fn = (io_fn_t)file->f_op->write; | |
293bc982 | 804 | iter_fn = file->f_op->write_iter; |
03d95eb2 | 805 | file_start_write(file); |
1da177e4 LT |
806 | } |
807 | ||
293bc982 | 808 | if (iter_fn) |
ac15ac06 | 809 | ret = do_iter_readv_writev(file, &iter, pos, iter_fn); |
ee0b3e67 | 810 | else |
ac15ac06 | 811 | ret = do_loop_readv_writev(file, &iter, pos, fn); |
1da177e4 | 812 | |
03d95eb2 AV |
813 | if (type != READ) |
814 | file_end_write(file); | |
815 | ||
1da177e4 | 816 | out: |
0504c074 | 817 | kfree(iov); |
0eeca283 RL |
818 | if ((ret + (type == READ)) > 0) { |
819 | if (type == READ) | |
2a12a9d7 | 820 | fsnotify_access(file); |
0eeca283 | 821 | else |
2a12a9d7 | 822 | fsnotify_modify(file); |
0eeca283 | 823 | } |
1da177e4 | 824 | return ret; |
1da177e4 LT |
825 | } |
826 | ||
827 | ssize_t vfs_readv(struct file *file, const struct iovec __user *vec, | |
828 | unsigned long vlen, loff_t *pos) | |
829 | { | |
830 | if (!(file->f_mode & FMODE_READ)) | |
831 | return -EBADF; | |
7f7f25e8 | 832 | if (!(file->f_mode & FMODE_CAN_READ)) |
1da177e4 LT |
833 | return -EINVAL; |
834 | ||
835 | return do_readv_writev(READ, file, vec, vlen, pos); | |
836 | } | |
837 | ||
838 | EXPORT_SYMBOL(vfs_readv); | |
839 | ||
840 | ssize_t vfs_writev(struct file *file, const struct iovec __user *vec, | |
841 | unsigned long vlen, loff_t *pos) | |
842 | { | |
843 | if (!(file->f_mode & FMODE_WRITE)) | |
844 | return -EBADF; | |
7f7f25e8 | 845 | if (!(file->f_mode & FMODE_CAN_WRITE)) |
1da177e4 LT |
846 | return -EINVAL; |
847 | ||
848 | return do_readv_writev(WRITE, file, vec, vlen, pos); | |
849 | } | |
850 | ||
851 | EXPORT_SYMBOL(vfs_writev); | |
852 | ||
3cdad428 HC |
853 | SYSCALL_DEFINE3(readv, unsigned long, fd, const struct iovec __user *, vec, |
854 | unsigned long, vlen) | |
1da177e4 | 855 | { |
9c225f26 | 856 | struct fd f = fdget_pos(fd); |
1da177e4 | 857 | ssize_t ret = -EBADF; |
1da177e4 | 858 | |
2903ff01 AV |
859 | if (f.file) { |
860 | loff_t pos = file_pos_read(f.file); | |
861 | ret = vfs_readv(f.file, vec, vlen, &pos); | |
5faf153e AV |
862 | if (ret >= 0) |
863 | file_pos_write(f.file, pos); | |
9c225f26 | 864 | fdput_pos(f); |
1da177e4 LT |
865 | } |
866 | ||
867 | if (ret > 0) | |
4b98d11b AD |
868 | add_rchar(current, ret); |
869 | inc_syscr(current); | |
1da177e4 LT |
870 | return ret; |
871 | } | |
872 | ||
3cdad428 HC |
873 | SYSCALL_DEFINE3(writev, unsigned long, fd, const struct iovec __user *, vec, |
874 | unsigned long, vlen) | |
1da177e4 | 875 | { |
9c225f26 | 876 | struct fd f = fdget_pos(fd); |
1da177e4 | 877 | ssize_t ret = -EBADF; |
1da177e4 | 878 | |
2903ff01 AV |
879 | if (f.file) { |
880 | loff_t pos = file_pos_read(f.file); | |
881 | ret = vfs_writev(f.file, vec, vlen, &pos); | |
5faf153e AV |
882 | if (ret >= 0) |
883 | file_pos_write(f.file, pos); | |
9c225f26 | 884 | fdput_pos(f); |
1da177e4 LT |
885 | } |
886 | ||
887 | if (ret > 0) | |
4b98d11b AD |
888 | add_wchar(current, ret); |
889 | inc_syscw(current); | |
1da177e4 LT |
890 | return ret; |
891 | } | |
892 | ||
601cc11d LT |
893 | static inline loff_t pos_from_hilo(unsigned long high, unsigned long low) |
894 | { | |
895 | #define HALF_LONG_BITS (BITS_PER_LONG / 2) | |
896 | return (((loff_t)high << HALF_LONG_BITS) << HALF_LONG_BITS) | low; | |
897 | } | |
898 | ||
f3554f4b | 899 | SYSCALL_DEFINE5(preadv, unsigned long, fd, const struct iovec __user *, vec, |
601cc11d | 900 | unsigned long, vlen, unsigned long, pos_l, unsigned long, pos_h) |
f3554f4b | 901 | { |
601cc11d | 902 | loff_t pos = pos_from_hilo(pos_h, pos_l); |
2903ff01 | 903 | struct fd f; |
f3554f4b | 904 | ssize_t ret = -EBADF; |
f3554f4b GH |
905 | |
906 | if (pos < 0) | |
907 | return -EINVAL; | |
908 | ||
2903ff01 AV |
909 | f = fdget(fd); |
910 | if (f.file) { | |
f3554f4b | 911 | ret = -ESPIPE; |
2903ff01 AV |
912 | if (f.file->f_mode & FMODE_PREAD) |
913 | ret = vfs_readv(f.file, vec, vlen, &pos); | |
914 | fdput(f); | |
f3554f4b GH |
915 | } |
916 | ||
917 | if (ret > 0) | |
918 | add_rchar(current, ret); | |
919 | inc_syscr(current); | |
920 | return ret; | |
921 | } | |
922 | ||
923 | SYSCALL_DEFINE5(pwritev, unsigned long, fd, const struct iovec __user *, vec, | |
601cc11d | 924 | unsigned long, vlen, unsigned long, pos_l, unsigned long, pos_h) |
f3554f4b | 925 | { |
601cc11d | 926 | loff_t pos = pos_from_hilo(pos_h, pos_l); |
2903ff01 | 927 | struct fd f; |
f3554f4b | 928 | ssize_t ret = -EBADF; |
f3554f4b GH |
929 | |
930 | if (pos < 0) | |
931 | return -EINVAL; | |
932 | ||
2903ff01 AV |
933 | f = fdget(fd); |
934 | if (f.file) { | |
f3554f4b | 935 | ret = -ESPIPE; |
2903ff01 AV |
936 | if (f.file->f_mode & FMODE_PWRITE) |
937 | ret = vfs_writev(f.file, vec, vlen, &pos); | |
938 | fdput(f); | |
f3554f4b GH |
939 | } |
940 | ||
941 | if (ret > 0) | |
942 | add_wchar(current, ret); | |
943 | inc_syscw(current); | |
944 | return ret; | |
945 | } | |
946 | ||
72ec3516 AV |
947 | #ifdef CONFIG_COMPAT |
948 | ||
949 | static ssize_t compat_do_readv_writev(int type, struct file *file, | |
950 | const struct compat_iovec __user *uvector, | |
951 | unsigned long nr_segs, loff_t *pos) | |
952 | { | |
953 | compat_ssize_t tot_len; | |
954 | struct iovec iovstack[UIO_FASTIOV]; | |
955 | struct iovec *iov = iovstack; | |
ac15ac06 | 956 | struct iov_iter iter; |
72ec3516 AV |
957 | ssize_t ret; |
958 | io_fn_t fn; | |
293bc982 | 959 | iter_fn_t iter_fn; |
72ec3516 | 960 | |
0504c074 AV |
961 | ret = compat_import_iovec(type, uvector, nr_segs, |
962 | UIO_FASTIOV, &iov, &iter); | |
963 | if (ret < 0) | |
964 | return ret; | |
72ec3516 | 965 | |
0504c074 AV |
966 | tot_len = iov_iter_count(&iter); |
967 | if (!tot_len) | |
968 | goto out; | |
72ec3516 AV |
969 | ret = rw_verify_area(type, file, pos, tot_len); |
970 | if (ret < 0) | |
971 | goto out; | |
972 | ||
72ec3516 AV |
973 | if (type == READ) { |
974 | fn = file->f_op->read; | |
293bc982 | 975 | iter_fn = file->f_op->read_iter; |
72ec3516 AV |
976 | } else { |
977 | fn = (io_fn_t)file->f_op->write; | |
293bc982 | 978 | iter_fn = file->f_op->write_iter; |
03d95eb2 | 979 | file_start_write(file); |
72ec3516 AV |
980 | } |
981 | ||
293bc982 | 982 | if (iter_fn) |
ac15ac06 | 983 | ret = do_iter_readv_writev(file, &iter, pos, iter_fn); |
03d95eb2 | 984 | else |
ac15ac06 | 985 | ret = do_loop_readv_writev(file, &iter, pos, fn); |
72ec3516 | 986 | |
03d95eb2 AV |
987 | if (type != READ) |
988 | file_end_write(file); | |
989 | ||
72ec3516 | 990 | out: |
0504c074 | 991 | kfree(iov); |
72ec3516 AV |
992 | if ((ret + (type == READ)) > 0) { |
993 | if (type == READ) | |
994 | fsnotify_access(file); | |
995 | else | |
996 | fsnotify_modify(file); | |
997 | } | |
998 | return ret; | |
999 | } | |
1000 | ||
1001 | static size_t compat_readv(struct file *file, | |
1002 | const struct compat_iovec __user *vec, | |
1003 | unsigned long vlen, loff_t *pos) | |
1004 | { | |
1005 | ssize_t ret = -EBADF; | |
1006 | ||
1007 | if (!(file->f_mode & FMODE_READ)) | |
1008 | goto out; | |
1009 | ||
1010 | ret = -EINVAL; | |
7f7f25e8 | 1011 | if (!(file->f_mode & FMODE_CAN_READ)) |
72ec3516 AV |
1012 | goto out; |
1013 | ||
1014 | ret = compat_do_readv_writev(READ, file, vec, vlen, pos); | |
1015 | ||
1016 | out: | |
1017 | if (ret > 0) | |
1018 | add_rchar(current, ret); | |
1019 | inc_syscr(current); | |
1020 | return ret; | |
1021 | } | |
1022 | ||
dfd948e3 | 1023 | COMPAT_SYSCALL_DEFINE3(readv, compat_ulong_t, fd, |
72ec3516 | 1024 | const struct compat_iovec __user *,vec, |
dfd948e3 | 1025 | compat_ulong_t, vlen) |
72ec3516 | 1026 | { |
9c225f26 | 1027 | struct fd f = fdget_pos(fd); |
72ec3516 AV |
1028 | ssize_t ret; |
1029 | loff_t pos; | |
1030 | ||
1031 | if (!f.file) | |
1032 | return -EBADF; | |
1033 | pos = f.file->f_pos; | |
1034 | ret = compat_readv(f.file, vec, vlen, &pos); | |
5faf153e AV |
1035 | if (ret >= 0) |
1036 | f.file->f_pos = pos; | |
9c225f26 | 1037 | fdput_pos(f); |
72ec3516 AV |
1038 | return ret; |
1039 | } | |
1040 | ||
378a10f3 HC |
1041 | static long __compat_sys_preadv64(unsigned long fd, |
1042 | const struct compat_iovec __user *vec, | |
1043 | unsigned long vlen, loff_t pos) | |
72ec3516 AV |
1044 | { |
1045 | struct fd f; | |
1046 | ssize_t ret; | |
1047 | ||
1048 | if (pos < 0) | |
1049 | return -EINVAL; | |
1050 | f = fdget(fd); | |
1051 | if (!f.file) | |
1052 | return -EBADF; | |
1053 | ret = -ESPIPE; | |
1054 | if (f.file->f_mode & FMODE_PREAD) | |
1055 | ret = compat_readv(f.file, vec, vlen, &pos); | |
1056 | fdput(f); | |
1057 | return ret; | |
1058 | } | |
1059 | ||
378a10f3 HC |
1060 | #ifdef __ARCH_WANT_COMPAT_SYS_PREADV64 |
1061 | COMPAT_SYSCALL_DEFINE4(preadv64, unsigned long, fd, | |
1062 | const struct compat_iovec __user *,vec, | |
1063 | unsigned long, vlen, loff_t, pos) | |
1064 | { | |
1065 | return __compat_sys_preadv64(fd, vec, vlen, pos); | |
1066 | } | |
1067 | #endif | |
1068 | ||
dfd948e3 | 1069 | COMPAT_SYSCALL_DEFINE5(preadv, compat_ulong_t, fd, |
72ec3516 | 1070 | const struct compat_iovec __user *,vec, |
dfd948e3 | 1071 | compat_ulong_t, vlen, u32, pos_low, u32, pos_high) |
72ec3516 AV |
1072 | { |
1073 | loff_t pos = ((loff_t)pos_high << 32) | pos_low; | |
378a10f3 HC |
1074 | |
1075 | return __compat_sys_preadv64(fd, vec, vlen, pos); | |
72ec3516 AV |
1076 | } |
1077 | ||
1078 | static size_t compat_writev(struct file *file, | |
1079 | const struct compat_iovec __user *vec, | |
1080 | unsigned long vlen, loff_t *pos) | |
1081 | { | |
1082 | ssize_t ret = -EBADF; | |
1083 | ||
1084 | if (!(file->f_mode & FMODE_WRITE)) | |
1085 | goto out; | |
1086 | ||
1087 | ret = -EINVAL; | |
7f7f25e8 | 1088 | if (!(file->f_mode & FMODE_CAN_WRITE)) |
72ec3516 AV |
1089 | goto out; |
1090 | ||
1091 | ret = compat_do_readv_writev(WRITE, file, vec, vlen, pos); | |
1092 | ||
1093 | out: | |
1094 | if (ret > 0) | |
1095 | add_wchar(current, ret); | |
1096 | inc_syscw(current); | |
1097 | return ret; | |
1098 | } | |
1099 | ||
dfd948e3 | 1100 | COMPAT_SYSCALL_DEFINE3(writev, compat_ulong_t, fd, |
72ec3516 | 1101 | const struct compat_iovec __user *, vec, |
dfd948e3 | 1102 | compat_ulong_t, vlen) |
72ec3516 | 1103 | { |
9c225f26 | 1104 | struct fd f = fdget_pos(fd); |
72ec3516 AV |
1105 | ssize_t ret; |
1106 | loff_t pos; | |
1107 | ||
1108 | if (!f.file) | |
1109 | return -EBADF; | |
1110 | pos = f.file->f_pos; | |
1111 | ret = compat_writev(f.file, vec, vlen, &pos); | |
5faf153e AV |
1112 | if (ret >= 0) |
1113 | f.file->f_pos = pos; | |
9c225f26 | 1114 | fdput_pos(f); |
72ec3516 AV |
1115 | return ret; |
1116 | } | |
1117 | ||
378a10f3 HC |
1118 | static long __compat_sys_pwritev64(unsigned long fd, |
1119 | const struct compat_iovec __user *vec, | |
1120 | unsigned long vlen, loff_t pos) | |
72ec3516 AV |
1121 | { |
1122 | struct fd f; | |
1123 | ssize_t ret; | |
1124 | ||
1125 | if (pos < 0) | |
1126 | return -EINVAL; | |
1127 | f = fdget(fd); | |
1128 | if (!f.file) | |
1129 | return -EBADF; | |
1130 | ret = -ESPIPE; | |
1131 | if (f.file->f_mode & FMODE_PWRITE) | |
1132 | ret = compat_writev(f.file, vec, vlen, &pos); | |
1133 | fdput(f); | |
1134 | return ret; | |
1135 | } | |
1136 | ||
378a10f3 HC |
1137 | #ifdef __ARCH_WANT_COMPAT_SYS_PWRITEV64 |
1138 | COMPAT_SYSCALL_DEFINE4(pwritev64, unsigned long, fd, | |
1139 | const struct compat_iovec __user *,vec, | |
1140 | unsigned long, vlen, loff_t, pos) | |
1141 | { | |
1142 | return __compat_sys_pwritev64(fd, vec, vlen, pos); | |
1143 | } | |
1144 | #endif | |
1145 | ||
dfd948e3 | 1146 | COMPAT_SYSCALL_DEFINE5(pwritev, compat_ulong_t, fd, |
72ec3516 | 1147 | const struct compat_iovec __user *,vec, |
dfd948e3 | 1148 | compat_ulong_t, vlen, u32, pos_low, u32, pos_high) |
72ec3516 AV |
1149 | { |
1150 | loff_t pos = ((loff_t)pos_high << 32) | pos_low; | |
378a10f3 HC |
1151 | |
1152 | return __compat_sys_pwritev64(fd, vec, vlen, pos); | |
72ec3516 AV |
1153 | } |
1154 | #endif | |
1155 | ||
19f4fc3a AV |
1156 | static ssize_t do_sendfile(int out_fd, int in_fd, loff_t *ppos, |
1157 | size_t count, loff_t max) | |
1da177e4 | 1158 | { |
2903ff01 AV |
1159 | struct fd in, out; |
1160 | struct inode *in_inode, *out_inode; | |
1da177e4 | 1161 | loff_t pos; |
7995bd28 | 1162 | loff_t out_pos; |
1da177e4 | 1163 | ssize_t retval; |
2903ff01 | 1164 | int fl; |
1da177e4 LT |
1165 | |
1166 | /* | |
1167 | * Get input file, and verify that it is ok.. | |
1168 | */ | |
1169 | retval = -EBADF; | |
2903ff01 AV |
1170 | in = fdget(in_fd); |
1171 | if (!in.file) | |
1da177e4 | 1172 | goto out; |
2903ff01 | 1173 | if (!(in.file->f_mode & FMODE_READ)) |
1da177e4 | 1174 | goto fput_in; |
1da177e4 | 1175 | retval = -ESPIPE; |
7995bd28 AV |
1176 | if (!ppos) { |
1177 | pos = in.file->f_pos; | |
1178 | } else { | |
1179 | pos = *ppos; | |
2903ff01 | 1180 | if (!(in.file->f_mode & FMODE_PREAD)) |
1da177e4 | 1181 | goto fput_in; |
7995bd28 AV |
1182 | } |
1183 | retval = rw_verify_area(READ, in.file, &pos, count); | |
e28cc715 | 1184 | if (retval < 0) |
1da177e4 | 1185 | goto fput_in; |
e28cc715 | 1186 | count = retval; |
1da177e4 | 1187 | |
1da177e4 LT |
1188 | /* |
1189 | * Get output file, and verify that it is ok.. | |
1190 | */ | |
1191 | retval = -EBADF; | |
2903ff01 AV |
1192 | out = fdget(out_fd); |
1193 | if (!out.file) | |
1da177e4 | 1194 | goto fput_in; |
2903ff01 | 1195 | if (!(out.file->f_mode & FMODE_WRITE)) |
1da177e4 LT |
1196 | goto fput_out; |
1197 | retval = -EINVAL; | |
496ad9aa AV |
1198 | in_inode = file_inode(in.file); |
1199 | out_inode = file_inode(out.file); | |
7995bd28 AV |
1200 | out_pos = out.file->f_pos; |
1201 | retval = rw_verify_area(WRITE, out.file, &out_pos, count); | |
e28cc715 | 1202 | if (retval < 0) |
1da177e4 | 1203 | goto fput_out; |
e28cc715 | 1204 | count = retval; |
1da177e4 | 1205 | |
1da177e4 LT |
1206 | if (!max) |
1207 | max = min(in_inode->i_sb->s_maxbytes, out_inode->i_sb->s_maxbytes); | |
1208 | ||
1da177e4 LT |
1209 | if (unlikely(pos + count > max)) { |
1210 | retval = -EOVERFLOW; | |
1211 | if (pos >= max) | |
1212 | goto fput_out; | |
1213 | count = max - pos; | |
1214 | } | |
1215 | ||
d96e6e71 | 1216 | fl = 0; |
534f2aaa | 1217 | #if 0 |
d96e6e71 JA |
1218 | /* |
1219 | * We need to debate whether we can enable this or not. The | |
1220 | * man page documents EAGAIN return for the output at least, | |
1221 | * and the application is arguably buggy if it doesn't expect | |
1222 | * EAGAIN on a non-blocking file descriptor. | |
1223 | */ | |
2903ff01 | 1224 | if (in.file->f_flags & O_NONBLOCK) |
d96e6e71 | 1225 | fl = SPLICE_F_NONBLOCK; |
534f2aaa | 1226 | #endif |
50cd2c57 | 1227 | file_start_write(out.file); |
7995bd28 | 1228 | retval = do_splice_direct(in.file, &pos, out.file, &out_pos, count, fl); |
50cd2c57 | 1229 | file_end_write(out.file); |
1da177e4 LT |
1230 | |
1231 | if (retval > 0) { | |
4b98d11b AD |
1232 | add_rchar(current, retval); |
1233 | add_wchar(current, retval); | |
a68c2f12 SW |
1234 | fsnotify_access(in.file); |
1235 | fsnotify_modify(out.file); | |
7995bd28 AV |
1236 | out.file->f_pos = out_pos; |
1237 | if (ppos) | |
1238 | *ppos = pos; | |
1239 | else | |
1240 | in.file->f_pos = pos; | |
1da177e4 | 1241 | } |
1da177e4 | 1242 | |
4b98d11b AD |
1243 | inc_syscr(current); |
1244 | inc_syscw(current); | |
7995bd28 | 1245 | if (pos > max) |
1da177e4 LT |
1246 | retval = -EOVERFLOW; |
1247 | ||
1248 | fput_out: | |
2903ff01 | 1249 | fdput(out); |
1da177e4 | 1250 | fput_in: |
2903ff01 | 1251 | fdput(in); |
1da177e4 LT |
1252 | out: |
1253 | return retval; | |
1254 | } | |
1255 | ||
002c8976 | 1256 | SYSCALL_DEFINE4(sendfile, int, out_fd, int, in_fd, off_t __user *, offset, size_t, count) |
1da177e4 LT |
1257 | { |
1258 | loff_t pos; | |
1259 | off_t off; | |
1260 | ssize_t ret; | |
1261 | ||
1262 | if (offset) { | |
1263 | if (unlikely(get_user(off, offset))) | |
1264 | return -EFAULT; | |
1265 | pos = off; | |
1266 | ret = do_sendfile(out_fd, in_fd, &pos, count, MAX_NON_LFS); | |
1267 | if (unlikely(put_user(pos, offset))) | |
1268 | return -EFAULT; | |
1269 | return ret; | |
1270 | } | |
1271 | ||
1272 | return do_sendfile(out_fd, in_fd, NULL, count, 0); | |
1273 | } | |
1274 | ||
002c8976 | 1275 | SYSCALL_DEFINE4(sendfile64, int, out_fd, int, in_fd, loff_t __user *, offset, size_t, count) |
1da177e4 LT |
1276 | { |
1277 | loff_t pos; | |
1278 | ssize_t ret; | |
1279 | ||
1280 | if (offset) { | |
1281 | if (unlikely(copy_from_user(&pos, offset, sizeof(loff_t)))) | |
1282 | return -EFAULT; | |
1283 | ret = do_sendfile(out_fd, in_fd, &pos, count, 0); | |
1284 | if (unlikely(put_user(pos, offset))) | |
1285 | return -EFAULT; | |
1286 | return ret; | |
1287 | } | |
1288 | ||
1289 | return do_sendfile(out_fd, in_fd, NULL, count, 0); | |
1290 | } | |
19f4fc3a AV |
1291 | |
1292 | #ifdef CONFIG_COMPAT | |
1293 | COMPAT_SYSCALL_DEFINE4(sendfile, int, out_fd, int, in_fd, | |
1294 | compat_off_t __user *, offset, compat_size_t, count) | |
1295 | { | |
1296 | loff_t pos; | |
1297 | off_t off; | |
1298 | ssize_t ret; | |
1299 | ||
1300 | if (offset) { | |
1301 | if (unlikely(get_user(off, offset))) | |
1302 | return -EFAULT; | |
1303 | pos = off; | |
1304 | ret = do_sendfile(out_fd, in_fd, &pos, count, MAX_NON_LFS); | |
1305 | if (unlikely(put_user(pos, offset))) | |
1306 | return -EFAULT; | |
1307 | return ret; | |
1308 | } | |
1309 | ||
1310 | return do_sendfile(out_fd, in_fd, NULL, count, 0); | |
1311 | } | |
1312 | ||
1313 | COMPAT_SYSCALL_DEFINE4(sendfile64, int, out_fd, int, in_fd, | |
1314 | compat_loff_t __user *, offset, compat_size_t, count) | |
1315 | { | |
1316 | loff_t pos; | |
1317 | ssize_t ret; | |
1318 | ||
1319 | if (offset) { | |
1320 | if (unlikely(copy_from_user(&pos, offset, sizeof(loff_t)))) | |
1321 | return -EFAULT; | |
1322 | ret = do_sendfile(out_fd, in_fd, &pos, count, 0); | |
1323 | if (unlikely(put_user(pos, offset))) | |
1324 | return -EFAULT; | |
1325 | return ret; | |
1326 | } | |
1327 | ||
1328 | return do_sendfile(out_fd, in_fd, NULL, count, 0); | |
1329 | } | |
1330 | #endif | |
29732938 ZB |
1331 | |
1332 | /* | |
1333 | * copy_file_range() differs from regular file read and write in that it | |
1334 | * specifically allows return partial success. When it does so is up to | |
1335 | * the copy_file_range method. | |
1336 | */ | |
1337 | ssize_t vfs_copy_file_range(struct file *file_in, loff_t pos_in, | |
1338 | struct file *file_out, loff_t pos_out, | |
1339 | size_t len, unsigned int flags) | |
1340 | { | |
1341 | struct inode *inode_in = file_inode(file_in); | |
1342 | struct inode *inode_out = file_inode(file_out); | |
1343 | ssize_t ret; | |
1344 | ||
1345 | if (flags != 0) | |
1346 | return -EINVAL; | |
1347 | ||
1348 | /* copy_file_range allows full ssize_t len, ignoring MAX_RW_COUNT */ | |
1349 | ret = rw_verify_area(READ, file_in, &pos_in, len); | |
1350 | if (ret >= 0) | |
1351 | ret = rw_verify_area(WRITE, file_out, &pos_out, len); | |
1352 | if (ret < 0) | |
1353 | return ret; | |
1354 | ||
1355 | if (!(file_in->f_mode & FMODE_READ) || | |
1356 | !(file_out->f_mode & FMODE_WRITE) || | |
eac70053 | 1357 | (file_out->f_flags & O_APPEND)) |
29732938 ZB |
1358 | return -EBADF; |
1359 | ||
1360 | /* this could be relaxed once a method supports cross-fs copies */ | |
1361 | if (inode_in->i_sb != inode_out->i_sb) | |
1362 | return -EXDEV; | |
1363 | ||
1364 | if (len == 0) | |
1365 | return 0; | |
1366 | ||
1367 | ret = mnt_want_write_file(file_out); | |
1368 | if (ret) | |
1369 | return ret; | |
1370 | ||
eac70053 AS |
1371 | ret = -EOPNOTSUPP; |
1372 | if (file_out->f_op->copy_file_range) | |
1373 | ret = file_out->f_op->copy_file_range(file_in, pos_in, file_out, | |
1374 | pos_out, len, flags); | |
1375 | if (ret == -EOPNOTSUPP) | |
1376 | ret = do_splice_direct(file_in, &pos_in, file_out, &pos_out, | |
1377 | len > MAX_RW_COUNT ? MAX_RW_COUNT : len, 0); | |
1378 | ||
29732938 ZB |
1379 | if (ret > 0) { |
1380 | fsnotify_access(file_in); | |
1381 | add_rchar(current, ret); | |
1382 | fsnotify_modify(file_out); | |
1383 | add_wchar(current, ret); | |
1384 | } | |
1385 | inc_syscr(current); | |
1386 | inc_syscw(current); | |
1387 | ||
1388 | mnt_drop_write_file(file_out); | |
1389 | ||
1390 | return ret; | |
1391 | } | |
1392 | EXPORT_SYMBOL(vfs_copy_file_range); | |
1393 | ||
1394 | SYSCALL_DEFINE6(copy_file_range, int, fd_in, loff_t __user *, off_in, | |
1395 | int, fd_out, loff_t __user *, off_out, | |
1396 | size_t, len, unsigned int, flags) | |
1397 | { | |
1398 | loff_t pos_in; | |
1399 | loff_t pos_out; | |
1400 | struct fd f_in; | |
1401 | struct fd f_out; | |
1402 | ssize_t ret = -EBADF; | |
1403 | ||
1404 | f_in = fdget(fd_in); | |
1405 | if (!f_in.file) | |
1406 | goto out2; | |
1407 | ||
1408 | f_out = fdget(fd_out); | |
1409 | if (!f_out.file) | |
1410 | goto out1; | |
1411 | ||
1412 | ret = -EFAULT; | |
1413 | if (off_in) { | |
1414 | if (copy_from_user(&pos_in, off_in, sizeof(loff_t))) | |
1415 | goto out; | |
1416 | } else { | |
1417 | pos_in = f_in.file->f_pos; | |
1418 | } | |
1419 | ||
1420 | if (off_out) { | |
1421 | if (copy_from_user(&pos_out, off_out, sizeof(loff_t))) | |
1422 | goto out; | |
1423 | } else { | |
1424 | pos_out = f_out.file->f_pos; | |
1425 | } | |
1426 | ||
1427 | ret = vfs_copy_file_range(f_in.file, pos_in, f_out.file, pos_out, len, | |
1428 | flags); | |
1429 | if (ret > 0) { | |
1430 | pos_in += ret; | |
1431 | pos_out += ret; | |
1432 | ||
1433 | if (off_in) { | |
1434 | if (copy_to_user(off_in, &pos_in, sizeof(loff_t))) | |
1435 | ret = -EFAULT; | |
1436 | } else { | |
1437 | f_in.file->f_pos = pos_in; | |
1438 | } | |
1439 | ||
1440 | if (off_out) { | |
1441 | if (copy_to_user(off_out, &pos_out, sizeof(loff_t))) | |
1442 | ret = -EFAULT; | |
1443 | } else { | |
1444 | f_out.file->f_pos = pos_out; | |
1445 | } | |
1446 | } | |
1447 | ||
1448 | out: | |
1449 | fdput(f_out); | |
1450 | out1: | |
1451 | fdput(f_in); | |
1452 | out2: | |
1453 | return ret; | |
1454 | } |