]>
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> |
06ae43f3 | 19 | #include "internal.h" |
1da177e4 LT |
20 | |
21 | #include <asm/uaccess.h> | |
22 | #include <asm/unistd.h> | |
23 | ||
c0bd14af AV |
24 | typedef ssize_t (*io_fn_t)(struct file *, char __user *, size_t, loff_t *); |
25 | typedef ssize_t (*iov_fn_t)(struct kiocb *, const struct iovec *, | |
26 | unsigned long, loff_t); | |
27 | ||
4b6f5d20 | 28 | const struct file_operations generic_ro_fops = { |
1da177e4 | 29 | .llseek = generic_file_llseek, |
543ade1f BP |
30 | .read = do_sync_read, |
31 | .aio_read = generic_file_aio_read, | |
1da177e4 | 32 | .mmap = generic_file_readonly_mmap, |
534f2aaa | 33 | .splice_read = generic_file_splice_read, |
1da177e4 LT |
34 | }; |
35 | ||
36 | EXPORT_SYMBOL(generic_ro_fops); | |
37 | ||
cccb5a1e | 38 | static inline int unsigned_offsets(struct file *file) |
4a3956c7 | 39 | { |
cccb5a1e | 40 | return file->f_mode & FMODE_UNSIGNED_OFFSET; |
4a3956c7 KH |
41 | } |
42 | ||
ef3d0fd2 AK |
43 | static loff_t lseek_execute(struct file *file, struct inode *inode, |
44 | loff_t offset, loff_t maxsize) | |
45 | { | |
46 | if (offset < 0 && !unsigned_offsets(file)) | |
47 | return -EINVAL; | |
48 | if (offset > maxsize) | |
49 | return -EINVAL; | |
50 | ||
51 | if (offset != file->f_pos) { | |
52 | file->f_pos = offset; | |
53 | file->f_version = 0; | |
54 | } | |
55 | return offset; | |
56 | } | |
57 | ||
3a8cff4f | 58 | /** |
5760495a | 59 | * generic_file_llseek_size - generic llseek implementation for regular files |
3a8cff4f CH |
60 | * @file: file structure to seek on |
61 | * @offset: file offset to seek to | |
965c8e59 | 62 | * @whence: type of seek |
e8b96eb5 ES |
63 | * @size: max size of this file in file system |
64 | * @eof: offset used for SEEK_END position | |
3a8cff4f | 65 | * |
5760495a | 66 | * This is a variant of generic_file_llseek that allows passing in a custom |
e8b96eb5 | 67 | * maximum file size and a custom EOF position, for e.g. hashed directories |
ef3d0fd2 AK |
68 | * |
69 | * Synchronization: | |
5760495a | 70 | * SEEK_SET and SEEK_END are unsynchronized (but atomic on 64bit platforms) |
ef3d0fd2 AK |
71 | * SEEK_CUR is synchronized against other SEEK_CURs, but not read/writes. |
72 | * read/writes behave like SEEK_SET against seeks. | |
3a8cff4f | 73 | */ |
9465efc9 | 74 | loff_t |
965c8e59 | 75 | generic_file_llseek_size(struct file *file, loff_t offset, int whence, |
e8b96eb5 | 76 | loff_t maxsize, loff_t eof) |
1da177e4 | 77 | { |
1da177e4 LT |
78 | struct inode *inode = file->f_mapping->host; |
79 | ||
965c8e59 | 80 | switch (whence) { |
3a8cff4f | 81 | case SEEK_END: |
e8b96eb5 | 82 | offset += eof; |
3a8cff4f CH |
83 | break; |
84 | case SEEK_CUR: | |
5b6f1eb9 AK |
85 | /* |
86 | * Here we special-case the lseek(fd, 0, SEEK_CUR) | |
87 | * position-querying operation. Avoid rewriting the "same" | |
88 | * f_pos value back to the file because a concurrent read(), | |
89 | * write() or lseek() might have altered it | |
90 | */ | |
91 | if (offset == 0) | |
92 | return file->f_pos; | |
ef3d0fd2 AK |
93 | /* |
94 | * f_lock protects against read/modify/write race with other | |
95 | * SEEK_CURs. Note that parallel writes and reads behave | |
96 | * like SEEK_SET. | |
97 | */ | |
98 | spin_lock(&file->f_lock); | |
99 | offset = lseek_execute(file, inode, file->f_pos + offset, | |
5760495a | 100 | maxsize); |
ef3d0fd2 AK |
101 | spin_unlock(&file->f_lock); |
102 | return offset; | |
982d8165 JB |
103 | case SEEK_DATA: |
104 | /* | |
105 | * In the generic case the entire file is data, so as long as | |
106 | * offset isn't at the end of the file then the offset is data. | |
107 | */ | |
e8b96eb5 | 108 | if (offset >= eof) |
982d8165 JB |
109 | return -ENXIO; |
110 | break; | |
111 | case SEEK_HOLE: | |
112 | /* | |
113 | * There is a virtual hole at the end of the file, so as long as | |
114 | * offset isn't i_size or larger, return i_size. | |
115 | */ | |
e8b96eb5 | 116 | if (offset >= eof) |
982d8165 | 117 | return -ENXIO; |
e8b96eb5 | 118 | offset = eof; |
982d8165 | 119 | break; |
1da177e4 | 120 | } |
3a8cff4f | 121 | |
5760495a AK |
122 | return lseek_execute(file, inode, offset, maxsize); |
123 | } | |
124 | EXPORT_SYMBOL(generic_file_llseek_size); | |
125 | ||
126 | /** | |
127 | * generic_file_llseek - generic llseek implementation for regular files | |
128 | * @file: file structure to seek on | |
129 | * @offset: file offset to seek to | |
965c8e59 | 130 | * @whence: type of seek |
5760495a AK |
131 | * |
132 | * This is a generic implemenation of ->llseek useable for all normal local | |
133 | * filesystems. It just updates the file offset to the value specified by | |
546ae2d2 | 134 | * @offset and @whence. |
5760495a | 135 | */ |
965c8e59 | 136 | loff_t generic_file_llseek(struct file *file, loff_t offset, int whence) |
5760495a AK |
137 | { |
138 | struct inode *inode = file->f_mapping->host; | |
139 | ||
965c8e59 | 140 | return generic_file_llseek_size(file, offset, whence, |
e8b96eb5 ES |
141 | inode->i_sb->s_maxbytes, |
142 | i_size_read(inode)); | |
1da177e4 | 143 | } |
9465efc9 | 144 | EXPORT_SYMBOL(generic_file_llseek); |
1da177e4 | 145 | |
ae6afc3f B |
146 | /** |
147 | * noop_llseek - No Operation Performed llseek implementation | |
148 | * @file: file structure to seek on | |
149 | * @offset: file offset to seek to | |
965c8e59 | 150 | * @whence: type of seek |
ae6afc3f B |
151 | * |
152 | * This is an implementation of ->llseek useable for the rare special case when | |
153 | * userspace expects the seek to succeed but the (device) file is actually not | |
154 | * able to perform the seek. In this case you use noop_llseek() instead of | |
155 | * falling back to the default implementation of ->llseek. | |
156 | */ | |
965c8e59 | 157 | loff_t noop_llseek(struct file *file, loff_t offset, int whence) |
ae6afc3f B |
158 | { |
159 | return file->f_pos; | |
160 | } | |
161 | EXPORT_SYMBOL(noop_llseek); | |
162 | ||
965c8e59 | 163 | loff_t no_llseek(struct file *file, loff_t offset, int whence) |
1da177e4 LT |
164 | { |
165 | return -ESPIPE; | |
166 | } | |
167 | EXPORT_SYMBOL(no_llseek); | |
168 | ||
965c8e59 | 169 | loff_t default_llseek(struct file *file, loff_t offset, int whence) |
1da177e4 | 170 | { |
496ad9aa | 171 | struct inode *inode = file_inode(file); |
16abef0e | 172 | loff_t retval; |
1da177e4 | 173 | |
982d8165 | 174 | mutex_lock(&inode->i_mutex); |
965c8e59 | 175 | switch (whence) { |
7b8e8924 | 176 | case SEEK_END: |
982d8165 | 177 | offset += i_size_read(inode); |
1da177e4 | 178 | break; |
7b8e8924 | 179 | case SEEK_CUR: |
5b6f1eb9 AK |
180 | if (offset == 0) { |
181 | retval = file->f_pos; | |
182 | goto out; | |
183 | } | |
1da177e4 | 184 | offset += file->f_pos; |
982d8165 JB |
185 | break; |
186 | case SEEK_DATA: | |
187 | /* | |
188 | * In the generic case the entire file is data, so as | |
189 | * long as offset isn't at the end of the file then the | |
190 | * offset is data. | |
191 | */ | |
bacb2d81 DC |
192 | if (offset >= inode->i_size) { |
193 | retval = -ENXIO; | |
194 | goto out; | |
195 | } | |
982d8165 JB |
196 | break; |
197 | case SEEK_HOLE: | |
198 | /* | |
199 | * There is a virtual hole at the end of the file, so | |
200 | * as long as offset isn't i_size or larger, return | |
201 | * i_size. | |
202 | */ | |
bacb2d81 DC |
203 | if (offset >= inode->i_size) { |
204 | retval = -ENXIO; | |
205 | goto out; | |
206 | } | |
982d8165 JB |
207 | offset = inode->i_size; |
208 | break; | |
1da177e4 LT |
209 | } |
210 | retval = -EINVAL; | |
cccb5a1e | 211 | if (offset >= 0 || unsigned_offsets(file)) { |
1da177e4 LT |
212 | if (offset != file->f_pos) { |
213 | file->f_pos = offset; | |
214 | file->f_version = 0; | |
215 | } | |
216 | retval = offset; | |
217 | } | |
5b6f1eb9 | 218 | out: |
982d8165 | 219 | mutex_unlock(&inode->i_mutex); |
1da177e4 LT |
220 | return retval; |
221 | } | |
222 | EXPORT_SYMBOL(default_llseek); | |
223 | ||
965c8e59 | 224 | loff_t vfs_llseek(struct file *file, loff_t offset, int whence) |
1da177e4 LT |
225 | { |
226 | loff_t (*fn)(struct file *, loff_t, int); | |
227 | ||
228 | fn = no_llseek; | |
229 | if (file->f_mode & FMODE_LSEEK) { | |
1da177e4 LT |
230 | if (file->f_op && file->f_op->llseek) |
231 | fn = file->f_op->llseek; | |
232 | } | |
965c8e59 | 233 | return fn(file, offset, whence); |
1da177e4 LT |
234 | } |
235 | EXPORT_SYMBOL(vfs_llseek); | |
236 | ||
965c8e59 | 237 | SYSCALL_DEFINE3(lseek, unsigned int, fd, off_t, offset, unsigned int, whence) |
1da177e4 LT |
238 | { |
239 | off_t retval; | |
2903ff01 AV |
240 | struct fd f = fdget(fd); |
241 | if (!f.file) | |
242 | return -EBADF; | |
1da177e4 LT |
243 | |
244 | retval = -EINVAL; | |
965c8e59 AM |
245 | if (whence <= SEEK_MAX) { |
246 | loff_t res = vfs_llseek(f.file, offset, whence); | |
1da177e4 LT |
247 | retval = res; |
248 | if (res != (loff_t)retval) | |
249 | retval = -EOVERFLOW; /* LFS: should only happen on 32 bit platforms */ | |
250 | } | |
2903ff01 | 251 | fdput(f); |
1da177e4 LT |
252 | return retval; |
253 | } | |
254 | ||
561c6731 AV |
255 | #ifdef CONFIG_COMPAT |
256 | COMPAT_SYSCALL_DEFINE3(lseek, unsigned int, fd, compat_off_t, offset, unsigned int, whence) | |
257 | { | |
258 | return sys_lseek(fd, offset, whence); | |
259 | } | |
260 | #endif | |
261 | ||
1da177e4 | 262 | #ifdef __ARCH_WANT_SYS_LLSEEK |
003d7ab4 HC |
263 | SYSCALL_DEFINE5(llseek, unsigned int, fd, unsigned long, offset_high, |
264 | unsigned long, offset_low, loff_t __user *, result, | |
965c8e59 | 265 | unsigned int, whence) |
1da177e4 LT |
266 | { |
267 | int retval; | |
2903ff01 | 268 | struct fd f = fdget(fd); |
1da177e4 | 269 | loff_t offset; |
1da177e4 | 270 | |
2903ff01 AV |
271 | if (!f.file) |
272 | return -EBADF; | |
1da177e4 LT |
273 | |
274 | retval = -EINVAL; | |
965c8e59 | 275 | if (whence > SEEK_MAX) |
1da177e4 LT |
276 | goto out_putf; |
277 | ||
2903ff01 | 278 | offset = vfs_llseek(f.file, ((loff_t) offset_high << 32) | offset_low, |
965c8e59 | 279 | whence); |
1da177e4 LT |
280 | |
281 | retval = (int)offset; | |
282 | if (offset >= 0) { | |
283 | retval = -EFAULT; | |
284 | if (!copy_to_user(result, &offset, sizeof(offset))) | |
285 | retval = 0; | |
286 | } | |
287 | out_putf: | |
2903ff01 | 288 | fdput(f); |
1da177e4 LT |
289 | return retval; |
290 | } | |
291 | #endif | |
292 | ||
e28cc715 LT |
293 | /* |
294 | * rw_verify_area doesn't like huge counts. We limit | |
295 | * them to something that fits in "int" so that others | |
296 | * won't have to do range checks all the time. | |
297 | */ | |
1da177e4 LT |
298 | int rw_verify_area(int read_write, struct file *file, loff_t *ppos, size_t count) |
299 | { | |
300 | struct inode *inode; | |
301 | loff_t pos; | |
c43e259c | 302 | int retval = -EINVAL; |
1da177e4 | 303 | |
496ad9aa | 304 | inode = file_inode(file); |
e28cc715 | 305 | if (unlikely((ssize_t) count < 0)) |
c43e259c | 306 | return retval; |
1da177e4 | 307 | pos = *ppos; |
cccb5a1e AV |
308 | if (unlikely(pos < 0)) { |
309 | if (!unsigned_offsets(file)) | |
310 | return retval; | |
311 | if (count >= -pos) /* both values are in 0..LLONG_MAX */ | |
312 | return -EOVERFLOW; | |
313 | } else if (unlikely((loff_t) (pos + count) < 0)) { | |
314 | if (!unsigned_offsets(file)) | |
4a3956c7 KH |
315 | return retval; |
316 | } | |
1da177e4 | 317 | |
a16877ca | 318 | if (unlikely(inode->i_flock && mandatory_lock(inode))) { |
c43e259c | 319 | retval = locks_mandatory_area( |
e28cc715 LT |
320 | read_write == READ ? FLOCK_VERIFY_READ : FLOCK_VERIFY_WRITE, |
321 | inode, file, pos, count); | |
322 | if (retval < 0) | |
323 | return retval; | |
324 | } | |
c43e259c JM |
325 | retval = security_file_permission(file, |
326 | read_write == READ ? MAY_READ : MAY_WRITE); | |
327 | if (retval) | |
328 | return retval; | |
e28cc715 | 329 | return count > MAX_RW_COUNT ? MAX_RW_COUNT : count; |
1da177e4 LT |
330 | } |
331 | ||
63e68809 BL |
332 | static void wait_on_retry_sync_kiocb(struct kiocb *iocb) |
333 | { | |
334 | set_current_state(TASK_UNINTERRUPTIBLE); | |
335 | if (!kiocbIsKicked(iocb)) | |
336 | schedule(); | |
337 | else | |
338 | kiocbClearKicked(iocb); | |
339 | __set_current_state(TASK_RUNNING); | |
340 | } | |
341 | ||
1da177e4 LT |
342 | ssize_t do_sync_read(struct file *filp, char __user *buf, size_t len, loff_t *ppos) |
343 | { | |
027445c3 | 344 | struct iovec iov = { .iov_base = buf, .iov_len = len }; |
1da177e4 LT |
345 | struct kiocb kiocb; |
346 | ssize_t ret; | |
347 | ||
348 | init_sync_kiocb(&kiocb, filp); | |
349 | kiocb.ki_pos = *ppos; | |
027445c3 | 350 | kiocb.ki_left = len; |
61964eba | 351 | kiocb.ki_nbytes = len; |
027445c3 BP |
352 | |
353 | for (;;) { | |
354 | ret = filp->f_op->aio_read(&kiocb, &iov, 1, kiocb.ki_pos); | |
355 | if (ret != -EIOCBRETRY) | |
356 | break; | |
63e68809 | 357 | wait_on_retry_sync_kiocb(&kiocb); |
027445c3 | 358 | } |
63e68809 | 359 | |
1da177e4 LT |
360 | if (-EIOCBQUEUED == ret) |
361 | ret = wait_on_sync_kiocb(&kiocb); | |
362 | *ppos = kiocb.ki_pos; | |
363 | return ret; | |
364 | } | |
365 | ||
366 | EXPORT_SYMBOL(do_sync_read); | |
367 | ||
368 | ssize_t vfs_read(struct file *file, char __user *buf, size_t count, loff_t *pos) | |
369 | { | |
370 | ssize_t ret; | |
371 | ||
372 | if (!(file->f_mode & FMODE_READ)) | |
373 | return -EBADF; | |
374 | if (!file->f_op || (!file->f_op->read && !file->f_op->aio_read)) | |
375 | return -EINVAL; | |
376 | if (unlikely(!access_ok(VERIFY_WRITE, buf, count))) | |
377 | return -EFAULT; | |
378 | ||
379 | ret = rw_verify_area(READ, file, pos, count); | |
e28cc715 LT |
380 | if (ret >= 0) { |
381 | count = ret; | |
c43e259c JM |
382 | if (file->f_op->read) |
383 | ret = file->f_op->read(file, buf, count, pos); | |
384 | else | |
385 | ret = do_sync_read(file, buf, count, pos); | |
386 | if (ret > 0) { | |
2a12a9d7 | 387 | fsnotify_access(file); |
c43e259c | 388 | add_rchar(current, ret); |
1da177e4 | 389 | } |
c43e259c | 390 | inc_syscr(current); |
1da177e4 LT |
391 | } |
392 | ||
393 | return ret; | |
394 | } | |
395 | ||
396 | EXPORT_SYMBOL(vfs_read); | |
397 | ||
398 | ssize_t do_sync_write(struct file *filp, const char __user *buf, size_t len, loff_t *ppos) | |
399 | { | |
027445c3 | 400 | struct iovec iov = { .iov_base = (void __user *)buf, .iov_len = len }; |
1da177e4 LT |
401 | struct kiocb kiocb; |
402 | ssize_t ret; | |
403 | ||
404 | init_sync_kiocb(&kiocb, filp); | |
405 | kiocb.ki_pos = *ppos; | |
027445c3 | 406 | kiocb.ki_left = len; |
61964eba | 407 | kiocb.ki_nbytes = len; |
027445c3 BP |
408 | |
409 | for (;;) { | |
410 | ret = filp->f_op->aio_write(&kiocb, &iov, 1, kiocb.ki_pos); | |
411 | if (ret != -EIOCBRETRY) | |
412 | break; | |
63e68809 | 413 | wait_on_retry_sync_kiocb(&kiocb); |
027445c3 | 414 | } |
63e68809 | 415 | |
1da177e4 LT |
416 | if (-EIOCBQUEUED == ret) |
417 | ret = wait_on_sync_kiocb(&kiocb); | |
418 | *ppos = kiocb.ki_pos; | |
419 | return ret; | |
420 | } | |
421 | ||
422 | EXPORT_SYMBOL(do_sync_write); | |
423 | ||
06ae43f3 AV |
424 | ssize_t __kernel_write(struct file *file, const char *buf, size_t count, loff_t *pos) |
425 | { | |
426 | mm_segment_t old_fs; | |
427 | const char __user *p; | |
428 | ssize_t ret; | |
429 | ||
3e84f48e AV |
430 | if (!file->f_op || (!file->f_op->write && !file->f_op->aio_write)) |
431 | return -EINVAL; | |
432 | ||
06ae43f3 AV |
433 | old_fs = get_fs(); |
434 | set_fs(get_ds()); | |
435 | p = (__force const char __user *)buf; | |
436 | if (count > MAX_RW_COUNT) | |
437 | count = MAX_RW_COUNT; | |
438 | if (file->f_op->write) | |
439 | ret = file->f_op->write(file, p, count, pos); | |
440 | else | |
441 | ret = do_sync_write(file, p, count, pos); | |
442 | set_fs(old_fs); | |
443 | if (ret > 0) { | |
444 | fsnotify_modify(file); | |
445 | add_wchar(current, ret); | |
446 | } | |
447 | inc_syscw(current); | |
448 | return ret; | |
449 | } | |
450 | ||
1da177e4 LT |
451 | ssize_t vfs_write(struct file *file, const char __user *buf, size_t count, loff_t *pos) |
452 | { | |
453 | ssize_t ret; | |
454 | ||
455 | if (!(file->f_mode & FMODE_WRITE)) | |
456 | return -EBADF; | |
457 | if (!file->f_op || (!file->f_op->write && !file->f_op->aio_write)) | |
458 | return -EINVAL; | |
459 | if (unlikely(!access_ok(VERIFY_READ, buf, count))) | |
460 | return -EFAULT; | |
461 | ||
462 | ret = rw_verify_area(WRITE, file, pos, count); | |
e28cc715 LT |
463 | if (ret >= 0) { |
464 | count = ret; | |
03d95eb2 | 465 | file_start_write(file); |
c43e259c JM |
466 | if (file->f_op->write) |
467 | ret = file->f_op->write(file, buf, count, pos); | |
468 | else | |
469 | ret = do_sync_write(file, buf, count, pos); | |
470 | if (ret > 0) { | |
2a12a9d7 | 471 | fsnotify_modify(file); |
c43e259c | 472 | add_wchar(current, ret); |
1da177e4 | 473 | } |
c43e259c | 474 | inc_syscw(current); |
03d95eb2 | 475 | file_end_write(file); |
1da177e4 LT |
476 | } |
477 | ||
478 | return ret; | |
479 | } | |
480 | ||
481 | EXPORT_SYMBOL(vfs_write); | |
482 | ||
483 | static inline loff_t file_pos_read(struct file *file) | |
484 | { | |
485 | return file->f_pos; | |
486 | } | |
487 | ||
488 | static inline void file_pos_write(struct file *file, loff_t pos) | |
489 | { | |
490 | file->f_pos = pos; | |
491 | } | |
492 | ||
3cdad428 | 493 | SYSCALL_DEFINE3(read, unsigned int, fd, char __user *, buf, size_t, count) |
1da177e4 | 494 | { |
2903ff01 | 495 | struct fd f = fdget(fd); |
1da177e4 | 496 | ssize_t ret = -EBADF; |
1da177e4 | 497 | |
2903ff01 AV |
498 | if (f.file) { |
499 | loff_t pos = file_pos_read(f.file); | |
500 | ret = vfs_read(f.file, buf, count, &pos); | |
501 | file_pos_write(f.file, pos); | |
502 | fdput(f); | |
1da177e4 | 503 | } |
1da177e4 LT |
504 | return ret; |
505 | } | |
1da177e4 | 506 | |
3cdad428 HC |
507 | SYSCALL_DEFINE3(write, unsigned int, fd, const char __user *, buf, |
508 | size_t, count) | |
1da177e4 | 509 | { |
2903ff01 | 510 | struct fd f = fdget(fd); |
1da177e4 | 511 | ssize_t ret = -EBADF; |
1da177e4 | 512 | |
2903ff01 AV |
513 | if (f.file) { |
514 | loff_t pos = file_pos_read(f.file); | |
515 | ret = vfs_write(f.file, buf, count, &pos); | |
516 | file_pos_write(f.file, pos); | |
517 | fdput(f); | |
1da177e4 LT |
518 | } |
519 | ||
520 | return ret; | |
521 | } | |
522 | ||
4a0fd5bf AV |
523 | SYSCALL_DEFINE4(pread64, unsigned int, fd, char __user *, buf, |
524 | size_t, count, loff_t, pos) | |
1da177e4 | 525 | { |
2903ff01 | 526 | struct fd f; |
1da177e4 | 527 | ssize_t ret = -EBADF; |
1da177e4 LT |
528 | |
529 | if (pos < 0) | |
530 | return -EINVAL; | |
531 | ||
2903ff01 AV |
532 | f = fdget(fd); |
533 | if (f.file) { | |
1da177e4 | 534 | ret = -ESPIPE; |
2903ff01 AV |
535 | if (f.file->f_mode & FMODE_PREAD) |
536 | ret = vfs_read(f.file, buf, count, &pos); | |
537 | fdput(f); | |
1da177e4 LT |
538 | } |
539 | ||
540 | return ret; | |
541 | } | |
542 | ||
4a0fd5bf AV |
543 | SYSCALL_DEFINE4(pwrite64, unsigned int, fd, const char __user *, buf, |
544 | size_t, count, loff_t, pos) | |
1da177e4 | 545 | { |
2903ff01 | 546 | struct fd f; |
1da177e4 | 547 | ssize_t ret = -EBADF; |
1da177e4 LT |
548 | |
549 | if (pos < 0) | |
550 | return -EINVAL; | |
551 | ||
2903ff01 AV |
552 | f = fdget(fd); |
553 | if (f.file) { | |
1da177e4 | 554 | ret = -ESPIPE; |
2903ff01 AV |
555 | if (f.file->f_mode & FMODE_PWRITE) |
556 | ret = vfs_write(f.file, buf, count, &pos); | |
557 | fdput(f); | |
1da177e4 LT |
558 | } |
559 | ||
560 | return ret; | |
561 | } | |
562 | ||
563 | /* | |
564 | * Reduce an iovec's length in-place. Return the resulting number of segments | |
565 | */ | |
566 | unsigned long iov_shorten(struct iovec *iov, unsigned long nr_segs, size_t to) | |
567 | { | |
568 | unsigned long seg = 0; | |
569 | size_t len = 0; | |
570 | ||
571 | while (seg < nr_segs) { | |
572 | seg++; | |
573 | if (len + iov->iov_len >= to) { | |
574 | iov->iov_len = to - len; | |
575 | break; | |
576 | } | |
577 | len += iov->iov_len; | |
578 | iov++; | |
579 | } | |
580 | return seg; | |
581 | } | |
19295529 | 582 | EXPORT_SYMBOL(iov_shorten); |
1da177e4 | 583 | |
72ec3516 | 584 | static ssize_t do_sync_readv_writev(struct file *filp, const struct iovec *iov, |
ee0b3e67 BP |
585 | unsigned long nr_segs, size_t len, loff_t *ppos, iov_fn_t fn) |
586 | { | |
587 | struct kiocb kiocb; | |
588 | ssize_t ret; | |
589 | ||
590 | init_sync_kiocb(&kiocb, filp); | |
591 | kiocb.ki_pos = *ppos; | |
592 | kiocb.ki_left = len; | |
593 | kiocb.ki_nbytes = len; | |
594 | ||
595 | for (;;) { | |
596 | ret = fn(&kiocb, iov, nr_segs, kiocb.ki_pos); | |
597 | if (ret != -EIOCBRETRY) | |
598 | break; | |
599 | wait_on_retry_sync_kiocb(&kiocb); | |
600 | } | |
601 | ||
602 | if (ret == -EIOCBQUEUED) | |
603 | ret = wait_on_sync_kiocb(&kiocb); | |
604 | *ppos = kiocb.ki_pos; | |
605 | return ret; | |
606 | } | |
607 | ||
608 | /* Do it by hand, with file-ops */ | |
72ec3516 | 609 | static ssize_t do_loop_readv_writev(struct file *filp, struct iovec *iov, |
ee0b3e67 BP |
610 | unsigned long nr_segs, loff_t *ppos, io_fn_t fn) |
611 | { | |
612 | struct iovec *vector = iov; | |
613 | ssize_t ret = 0; | |
614 | ||
615 | while (nr_segs > 0) { | |
616 | void __user *base; | |
617 | size_t len; | |
618 | ssize_t nr; | |
619 | ||
620 | base = vector->iov_base; | |
621 | len = vector->iov_len; | |
622 | vector++; | |
623 | nr_segs--; | |
624 | ||
625 | nr = fn(filp, base, len, ppos); | |
626 | ||
627 | if (nr < 0) { | |
628 | if (!ret) | |
629 | ret = nr; | |
630 | break; | |
631 | } | |
632 | ret += nr; | |
633 | if (nr != len) | |
634 | break; | |
635 | } | |
636 | ||
637 | return ret; | |
638 | } | |
639 | ||
1da177e4 LT |
640 | /* A write operation does a read from user space and vice versa */ |
641 | #define vrfy_dir(type) ((type) == READ ? VERIFY_WRITE : VERIFY_READ) | |
642 | ||
eed4e51f BP |
643 | ssize_t rw_copy_check_uvector(int type, const struct iovec __user * uvector, |
644 | unsigned long nr_segs, unsigned long fast_segs, | |
645 | struct iovec *fast_pointer, | |
ac34ebb3 | 646 | struct iovec **ret_pointer) |
435f49a5 | 647 | { |
eed4e51f | 648 | unsigned long seg; |
435f49a5 | 649 | ssize_t ret; |
eed4e51f BP |
650 | struct iovec *iov = fast_pointer; |
651 | ||
435f49a5 LT |
652 | /* |
653 | * SuS says "The readv() function *may* fail if the iovcnt argument | |
654 | * was less than or equal to 0, or greater than {IOV_MAX}. Linux has | |
655 | * traditionally returned zero for zero segments, so... | |
656 | */ | |
eed4e51f BP |
657 | if (nr_segs == 0) { |
658 | ret = 0; | |
435f49a5 | 659 | goto out; |
eed4e51f BP |
660 | } |
661 | ||
435f49a5 LT |
662 | /* |
663 | * First get the "struct iovec" from user memory and | |
664 | * verify all the pointers | |
665 | */ | |
eed4e51f BP |
666 | if (nr_segs > UIO_MAXIOV) { |
667 | ret = -EINVAL; | |
435f49a5 | 668 | goto out; |
eed4e51f BP |
669 | } |
670 | if (nr_segs > fast_segs) { | |
435f49a5 | 671 | iov = kmalloc(nr_segs*sizeof(struct iovec), GFP_KERNEL); |
eed4e51f BP |
672 | if (iov == NULL) { |
673 | ret = -ENOMEM; | |
435f49a5 | 674 | goto out; |
eed4e51f | 675 | } |
435f49a5 | 676 | } |
eed4e51f BP |
677 | if (copy_from_user(iov, uvector, nr_segs*sizeof(*uvector))) { |
678 | ret = -EFAULT; | |
435f49a5 | 679 | goto out; |
eed4e51f BP |
680 | } |
681 | ||
435f49a5 | 682 | /* |
eed4e51f BP |
683 | * According to the Single Unix Specification we should return EINVAL |
684 | * if an element length is < 0 when cast to ssize_t or if the | |
685 | * total length would overflow the ssize_t return value of the | |
686 | * system call. | |
435f49a5 LT |
687 | * |
688 | * Linux caps all read/write calls to MAX_RW_COUNT, and avoids the | |
689 | * overflow case. | |
690 | */ | |
eed4e51f | 691 | ret = 0; |
435f49a5 LT |
692 | for (seg = 0; seg < nr_segs; seg++) { |
693 | void __user *buf = iov[seg].iov_base; | |
694 | ssize_t len = (ssize_t)iov[seg].iov_len; | |
eed4e51f BP |
695 | |
696 | /* see if we we're about to use an invalid len or if | |
697 | * it's about to overflow ssize_t */ | |
435f49a5 | 698 | if (len < 0) { |
eed4e51f | 699 | ret = -EINVAL; |
435f49a5 | 700 | goto out; |
eed4e51f | 701 | } |
ac34ebb3 | 702 | if (type >= 0 |
fcf63409 | 703 | && unlikely(!access_ok(vrfy_dir(type), buf, len))) { |
eed4e51f | 704 | ret = -EFAULT; |
435f49a5 LT |
705 | goto out; |
706 | } | |
707 | if (len > MAX_RW_COUNT - ret) { | |
708 | len = MAX_RW_COUNT - ret; | |
709 | iov[seg].iov_len = len; | |
eed4e51f | 710 | } |
eed4e51f | 711 | ret += len; |
435f49a5 | 712 | } |
eed4e51f BP |
713 | out: |
714 | *ret_pointer = iov; | |
715 | return ret; | |
716 | } | |
717 | ||
1da177e4 LT |
718 | static ssize_t do_readv_writev(int type, struct file *file, |
719 | const struct iovec __user * uvector, | |
720 | unsigned long nr_segs, loff_t *pos) | |
721 | { | |
1da177e4 LT |
722 | size_t tot_len; |
723 | struct iovec iovstack[UIO_FASTIOV]; | |
ee0b3e67 | 724 | struct iovec *iov = iovstack; |
1da177e4 | 725 | ssize_t ret; |
1da177e4 LT |
726 | io_fn_t fn; |
727 | iov_fn_t fnv; | |
728 | ||
eed4e51f BP |
729 | if (!file->f_op) { |
730 | ret = -EINVAL; | |
1da177e4 | 731 | goto out; |
1da177e4 | 732 | } |
1da177e4 | 733 | |
eed4e51f | 734 | ret = rw_copy_check_uvector(type, uvector, nr_segs, |
ac34ebb3 | 735 | ARRAY_SIZE(iovstack), iovstack, &iov); |
eed4e51f | 736 | if (ret <= 0) |
1da177e4 | 737 | goto out; |
1da177e4 | 738 | |
eed4e51f | 739 | tot_len = ret; |
1da177e4 | 740 | ret = rw_verify_area(type, file, pos, tot_len); |
e28cc715 | 741 | if (ret < 0) |
411b67b4 | 742 | goto out; |
1da177e4 LT |
743 | |
744 | fnv = NULL; | |
745 | if (type == READ) { | |
746 | fn = file->f_op->read; | |
ee0b3e67 | 747 | fnv = file->f_op->aio_read; |
1da177e4 LT |
748 | } else { |
749 | fn = (io_fn_t)file->f_op->write; | |
ee0b3e67 | 750 | fnv = file->f_op->aio_write; |
03d95eb2 | 751 | file_start_write(file); |
1da177e4 LT |
752 | } |
753 | ||
ee0b3e67 BP |
754 | if (fnv) |
755 | ret = do_sync_readv_writev(file, iov, nr_segs, tot_len, | |
756 | pos, fnv); | |
757 | else | |
758 | ret = do_loop_readv_writev(file, iov, nr_segs, pos, fn); | |
1da177e4 | 759 | |
03d95eb2 AV |
760 | if (type != READ) |
761 | file_end_write(file); | |
762 | ||
1da177e4 LT |
763 | out: |
764 | if (iov != iovstack) | |
765 | kfree(iov); | |
0eeca283 RL |
766 | if ((ret + (type == READ)) > 0) { |
767 | if (type == READ) | |
2a12a9d7 | 768 | fsnotify_access(file); |
0eeca283 | 769 | else |
2a12a9d7 | 770 | fsnotify_modify(file); |
0eeca283 | 771 | } |
1da177e4 | 772 | return ret; |
1da177e4 LT |
773 | } |
774 | ||
775 | ssize_t vfs_readv(struct file *file, const struct iovec __user *vec, | |
776 | unsigned long vlen, loff_t *pos) | |
777 | { | |
778 | if (!(file->f_mode & FMODE_READ)) | |
779 | return -EBADF; | |
ee0b3e67 | 780 | if (!file->f_op || (!file->f_op->aio_read && !file->f_op->read)) |
1da177e4 LT |
781 | return -EINVAL; |
782 | ||
783 | return do_readv_writev(READ, file, vec, vlen, pos); | |
784 | } | |
785 | ||
786 | EXPORT_SYMBOL(vfs_readv); | |
787 | ||
788 | ssize_t vfs_writev(struct file *file, const struct iovec __user *vec, | |
789 | unsigned long vlen, loff_t *pos) | |
790 | { | |
791 | if (!(file->f_mode & FMODE_WRITE)) | |
792 | return -EBADF; | |
ee0b3e67 | 793 | if (!file->f_op || (!file->f_op->aio_write && !file->f_op->write)) |
1da177e4 LT |
794 | return -EINVAL; |
795 | ||
796 | return do_readv_writev(WRITE, file, vec, vlen, pos); | |
797 | } | |
798 | ||
799 | EXPORT_SYMBOL(vfs_writev); | |
800 | ||
3cdad428 HC |
801 | SYSCALL_DEFINE3(readv, unsigned long, fd, const struct iovec __user *, vec, |
802 | unsigned long, vlen) | |
1da177e4 | 803 | { |
2903ff01 | 804 | struct fd f = fdget(fd); |
1da177e4 | 805 | ssize_t ret = -EBADF; |
1da177e4 | 806 | |
2903ff01 AV |
807 | if (f.file) { |
808 | loff_t pos = file_pos_read(f.file); | |
809 | ret = vfs_readv(f.file, vec, vlen, &pos); | |
810 | file_pos_write(f.file, pos); | |
811 | fdput(f); | |
1da177e4 LT |
812 | } |
813 | ||
814 | if (ret > 0) | |
4b98d11b AD |
815 | add_rchar(current, ret); |
816 | inc_syscr(current); | |
1da177e4 LT |
817 | return ret; |
818 | } | |
819 | ||
3cdad428 HC |
820 | SYSCALL_DEFINE3(writev, unsigned long, fd, const struct iovec __user *, vec, |
821 | unsigned long, vlen) | |
1da177e4 | 822 | { |
2903ff01 | 823 | struct fd f = fdget(fd); |
1da177e4 | 824 | ssize_t ret = -EBADF; |
1da177e4 | 825 | |
2903ff01 AV |
826 | if (f.file) { |
827 | loff_t pos = file_pos_read(f.file); | |
828 | ret = vfs_writev(f.file, vec, vlen, &pos); | |
829 | file_pos_write(f.file, pos); | |
830 | fdput(f); | |
1da177e4 LT |
831 | } |
832 | ||
833 | if (ret > 0) | |
4b98d11b AD |
834 | add_wchar(current, ret); |
835 | inc_syscw(current); | |
1da177e4 LT |
836 | return ret; |
837 | } | |
838 | ||
601cc11d LT |
839 | static inline loff_t pos_from_hilo(unsigned long high, unsigned long low) |
840 | { | |
841 | #define HALF_LONG_BITS (BITS_PER_LONG / 2) | |
842 | return (((loff_t)high << HALF_LONG_BITS) << HALF_LONG_BITS) | low; | |
843 | } | |
844 | ||
f3554f4b | 845 | SYSCALL_DEFINE5(preadv, unsigned long, fd, const struct iovec __user *, vec, |
601cc11d | 846 | unsigned long, vlen, unsigned long, pos_l, unsigned long, pos_h) |
f3554f4b | 847 | { |
601cc11d | 848 | loff_t pos = pos_from_hilo(pos_h, pos_l); |
2903ff01 | 849 | struct fd f; |
f3554f4b | 850 | ssize_t ret = -EBADF; |
f3554f4b GH |
851 | |
852 | if (pos < 0) | |
853 | return -EINVAL; | |
854 | ||
2903ff01 AV |
855 | f = fdget(fd); |
856 | if (f.file) { | |
f3554f4b | 857 | ret = -ESPIPE; |
2903ff01 AV |
858 | if (f.file->f_mode & FMODE_PREAD) |
859 | ret = vfs_readv(f.file, vec, vlen, &pos); | |
860 | fdput(f); | |
f3554f4b GH |
861 | } |
862 | ||
863 | if (ret > 0) | |
864 | add_rchar(current, ret); | |
865 | inc_syscr(current); | |
866 | return ret; | |
867 | } | |
868 | ||
869 | SYSCALL_DEFINE5(pwritev, unsigned long, fd, const struct iovec __user *, vec, | |
601cc11d | 870 | unsigned long, vlen, unsigned long, pos_l, unsigned long, pos_h) |
f3554f4b | 871 | { |
601cc11d | 872 | loff_t pos = pos_from_hilo(pos_h, pos_l); |
2903ff01 | 873 | struct fd f; |
f3554f4b | 874 | ssize_t ret = -EBADF; |
f3554f4b GH |
875 | |
876 | if (pos < 0) | |
877 | return -EINVAL; | |
878 | ||
2903ff01 AV |
879 | f = fdget(fd); |
880 | if (f.file) { | |
f3554f4b | 881 | ret = -ESPIPE; |
2903ff01 AV |
882 | if (f.file->f_mode & FMODE_PWRITE) |
883 | ret = vfs_writev(f.file, vec, vlen, &pos); | |
884 | fdput(f); | |
f3554f4b GH |
885 | } |
886 | ||
887 | if (ret > 0) | |
888 | add_wchar(current, ret); | |
889 | inc_syscw(current); | |
890 | return ret; | |
891 | } | |
892 | ||
72ec3516 AV |
893 | #ifdef CONFIG_COMPAT |
894 | ||
895 | static ssize_t compat_do_readv_writev(int type, struct file *file, | |
896 | const struct compat_iovec __user *uvector, | |
897 | unsigned long nr_segs, loff_t *pos) | |
898 | { | |
899 | compat_ssize_t tot_len; | |
900 | struct iovec iovstack[UIO_FASTIOV]; | |
901 | struct iovec *iov = iovstack; | |
902 | ssize_t ret; | |
903 | io_fn_t fn; | |
904 | iov_fn_t fnv; | |
905 | ||
906 | ret = -EINVAL; | |
907 | if (!file->f_op) | |
908 | goto out; | |
909 | ||
910 | ret = -EFAULT; | |
911 | if (!access_ok(VERIFY_READ, uvector, nr_segs*sizeof(*uvector))) | |
912 | goto out; | |
913 | ||
914 | ret = compat_rw_copy_check_uvector(type, uvector, nr_segs, | |
915 | UIO_FASTIOV, iovstack, &iov); | |
916 | if (ret <= 0) | |
917 | goto out; | |
918 | ||
919 | tot_len = ret; | |
920 | ret = rw_verify_area(type, file, pos, tot_len); | |
921 | if (ret < 0) | |
922 | goto out; | |
923 | ||
924 | fnv = NULL; | |
925 | if (type == READ) { | |
926 | fn = file->f_op->read; | |
927 | fnv = file->f_op->aio_read; | |
928 | } else { | |
929 | fn = (io_fn_t)file->f_op->write; | |
930 | fnv = file->f_op->aio_write; | |
03d95eb2 | 931 | file_start_write(file); |
72ec3516 AV |
932 | } |
933 | ||
03d95eb2 | 934 | if (fnv) |
72ec3516 AV |
935 | ret = do_sync_readv_writev(file, iov, nr_segs, tot_len, |
936 | pos, fnv); | |
03d95eb2 | 937 | else |
72ec3516 AV |
938 | ret = do_loop_readv_writev(file, iov, nr_segs, pos, fn); |
939 | ||
03d95eb2 AV |
940 | if (type != READ) |
941 | file_end_write(file); | |
942 | ||
72ec3516 AV |
943 | out: |
944 | if (iov != iovstack) | |
945 | kfree(iov); | |
946 | if ((ret + (type == READ)) > 0) { | |
947 | if (type == READ) | |
948 | fsnotify_access(file); | |
949 | else | |
950 | fsnotify_modify(file); | |
951 | } | |
952 | return ret; | |
953 | } | |
954 | ||
955 | static size_t compat_readv(struct file *file, | |
956 | const struct compat_iovec __user *vec, | |
957 | unsigned long vlen, loff_t *pos) | |
958 | { | |
959 | ssize_t ret = -EBADF; | |
960 | ||
961 | if (!(file->f_mode & FMODE_READ)) | |
962 | goto out; | |
963 | ||
964 | ret = -EINVAL; | |
965 | if (!file->f_op || (!file->f_op->aio_read && !file->f_op->read)) | |
966 | goto out; | |
967 | ||
968 | ret = compat_do_readv_writev(READ, file, vec, vlen, pos); | |
969 | ||
970 | out: | |
971 | if (ret > 0) | |
972 | add_rchar(current, ret); | |
973 | inc_syscr(current); | |
974 | return ret; | |
975 | } | |
976 | ||
977 | COMPAT_SYSCALL_DEFINE3(readv, unsigned long, fd, | |
978 | const struct compat_iovec __user *,vec, | |
979 | unsigned long, vlen) | |
980 | { | |
981 | struct fd f = fdget(fd); | |
982 | ssize_t ret; | |
983 | loff_t pos; | |
984 | ||
985 | if (!f.file) | |
986 | return -EBADF; | |
987 | pos = f.file->f_pos; | |
988 | ret = compat_readv(f.file, vec, vlen, &pos); | |
989 | f.file->f_pos = pos; | |
990 | fdput(f); | |
991 | return ret; | |
992 | } | |
993 | ||
994 | COMPAT_SYSCALL_DEFINE4(preadv64, unsigned long, fd, | |
995 | const struct compat_iovec __user *,vec, | |
996 | unsigned long, vlen, loff_t, pos) | |
997 | { | |
998 | struct fd f; | |
999 | ssize_t ret; | |
1000 | ||
1001 | if (pos < 0) | |
1002 | return -EINVAL; | |
1003 | f = fdget(fd); | |
1004 | if (!f.file) | |
1005 | return -EBADF; | |
1006 | ret = -ESPIPE; | |
1007 | if (f.file->f_mode & FMODE_PREAD) | |
1008 | ret = compat_readv(f.file, vec, vlen, &pos); | |
1009 | fdput(f); | |
1010 | return ret; | |
1011 | } | |
1012 | ||
1013 | COMPAT_SYSCALL_DEFINE5(preadv, unsigned long, fd, | |
1014 | const struct compat_iovec __user *,vec, | |
1015 | unsigned long, vlen, u32, pos_low, u32, pos_high) | |
1016 | { | |
1017 | loff_t pos = ((loff_t)pos_high << 32) | pos_low; | |
1018 | return compat_sys_preadv64(fd, vec, vlen, pos); | |
1019 | } | |
1020 | ||
1021 | static size_t compat_writev(struct file *file, | |
1022 | const struct compat_iovec __user *vec, | |
1023 | unsigned long vlen, loff_t *pos) | |
1024 | { | |
1025 | ssize_t ret = -EBADF; | |
1026 | ||
1027 | if (!(file->f_mode & FMODE_WRITE)) | |
1028 | goto out; | |
1029 | ||
1030 | ret = -EINVAL; | |
1031 | if (!file->f_op || (!file->f_op->aio_write && !file->f_op->write)) | |
1032 | goto out; | |
1033 | ||
1034 | ret = compat_do_readv_writev(WRITE, file, vec, vlen, pos); | |
1035 | ||
1036 | out: | |
1037 | if (ret > 0) | |
1038 | add_wchar(current, ret); | |
1039 | inc_syscw(current); | |
1040 | return ret; | |
1041 | } | |
1042 | ||
1043 | COMPAT_SYSCALL_DEFINE3(writev, unsigned long, fd, | |
1044 | const struct compat_iovec __user *, vec, | |
1045 | unsigned long, vlen) | |
1046 | { | |
1047 | struct fd f = fdget(fd); | |
1048 | ssize_t ret; | |
1049 | loff_t pos; | |
1050 | ||
1051 | if (!f.file) | |
1052 | return -EBADF; | |
1053 | pos = f.file->f_pos; | |
1054 | ret = compat_writev(f.file, vec, vlen, &pos); | |
1055 | f.file->f_pos = pos; | |
1056 | fdput(f); | |
1057 | return ret; | |
1058 | } | |
1059 | ||
1060 | COMPAT_SYSCALL_DEFINE4(pwritev64, unsigned long, fd, | |
1061 | const struct compat_iovec __user *,vec, | |
1062 | unsigned long, vlen, loff_t, pos) | |
1063 | { | |
1064 | struct fd f; | |
1065 | ssize_t ret; | |
1066 | ||
1067 | if (pos < 0) | |
1068 | return -EINVAL; | |
1069 | f = fdget(fd); | |
1070 | if (!f.file) | |
1071 | return -EBADF; | |
1072 | ret = -ESPIPE; | |
1073 | if (f.file->f_mode & FMODE_PWRITE) | |
1074 | ret = compat_writev(f.file, vec, vlen, &pos); | |
1075 | fdput(f); | |
1076 | return ret; | |
1077 | } | |
1078 | ||
1079 | COMPAT_SYSCALL_DEFINE5(pwritev, unsigned long, fd, | |
1080 | const struct compat_iovec __user *,vec, | |
1081 | unsigned long, vlen, u32, pos_low, u32, pos_high) | |
1082 | { | |
1083 | loff_t pos = ((loff_t)pos_high << 32) | pos_low; | |
1084 | return compat_sys_pwritev64(fd, vec, vlen, pos); | |
1085 | } | |
1086 | #endif | |
1087 | ||
19f4fc3a AV |
1088 | static ssize_t do_sendfile(int out_fd, int in_fd, loff_t *ppos, |
1089 | size_t count, loff_t max) | |
1da177e4 | 1090 | { |
2903ff01 AV |
1091 | struct fd in, out; |
1092 | struct inode *in_inode, *out_inode; | |
1da177e4 LT |
1093 | loff_t pos; |
1094 | ssize_t retval; | |
2903ff01 | 1095 | int fl; |
1da177e4 LT |
1096 | |
1097 | /* | |
1098 | * Get input file, and verify that it is ok.. | |
1099 | */ | |
1100 | retval = -EBADF; | |
2903ff01 AV |
1101 | in = fdget(in_fd); |
1102 | if (!in.file) | |
1da177e4 | 1103 | goto out; |
2903ff01 | 1104 | if (!(in.file->f_mode & FMODE_READ)) |
1da177e4 | 1105 | goto fput_in; |
1da177e4 LT |
1106 | retval = -ESPIPE; |
1107 | if (!ppos) | |
2903ff01 | 1108 | ppos = &in.file->f_pos; |
1da177e4 | 1109 | else |
2903ff01 | 1110 | if (!(in.file->f_mode & FMODE_PREAD)) |
1da177e4 | 1111 | goto fput_in; |
2903ff01 | 1112 | retval = rw_verify_area(READ, in.file, ppos, count); |
e28cc715 | 1113 | if (retval < 0) |
1da177e4 | 1114 | goto fput_in; |
e28cc715 | 1115 | count = retval; |
1da177e4 | 1116 | |
1da177e4 LT |
1117 | /* |
1118 | * Get output file, and verify that it is ok.. | |
1119 | */ | |
1120 | retval = -EBADF; | |
2903ff01 AV |
1121 | out = fdget(out_fd); |
1122 | if (!out.file) | |
1da177e4 | 1123 | goto fput_in; |
2903ff01 | 1124 | if (!(out.file->f_mode & FMODE_WRITE)) |
1da177e4 LT |
1125 | goto fput_out; |
1126 | retval = -EINVAL; | |
496ad9aa AV |
1127 | in_inode = file_inode(in.file); |
1128 | out_inode = file_inode(out.file); | |
2903ff01 | 1129 | retval = rw_verify_area(WRITE, out.file, &out.file->f_pos, count); |
e28cc715 | 1130 | if (retval < 0) |
1da177e4 | 1131 | goto fput_out; |
e28cc715 | 1132 | count = retval; |
1da177e4 | 1133 | |
1da177e4 LT |
1134 | if (!max) |
1135 | max = min(in_inode->i_sb->s_maxbytes, out_inode->i_sb->s_maxbytes); | |
1136 | ||
1137 | pos = *ppos; | |
1da177e4 LT |
1138 | if (unlikely(pos + count > max)) { |
1139 | retval = -EOVERFLOW; | |
1140 | if (pos >= max) | |
1141 | goto fput_out; | |
1142 | count = max - pos; | |
1143 | } | |
1144 | ||
d96e6e71 | 1145 | fl = 0; |
534f2aaa | 1146 | #if 0 |
d96e6e71 JA |
1147 | /* |
1148 | * We need to debate whether we can enable this or not. The | |
1149 | * man page documents EAGAIN return for the output at least, | |
1150 | * and the application is arguably buggy if it doesn't expect | |
1151 | * EAGAIN on a non-blocking file descriptor. | |
1152 | */ | |
2903ff01 | 1153 | if (in.file->f_flags & O_NONBLOCK) |
d96e6e71 | 1154 | fl = SPLICE_F_NONBLOCK; |
534f2aaa | 1155 | #endif |
2903ff01 | 1156 | retval = do_splice_direct(in.file, ppos, out.file, count, fl); |
1da177e4 LT |
1157 | |
1158 | if (retval > 0) { | |
4b98d11b AD |
1159 | add_rchar(current, retval); |
1160 | add_wchar(current, retval); | |
a68c2f12 SW |
1161 | fsnotify_access(in.file); |
1162 | fsnotify_modify(out.file); | |
1da177e4 | 1163 | } |
1da177e4 | 1164 | |
4b98d11b AD |
1165 | inc_syscr(current); |
1166 | inc_syscw(current); | |
1da177e4 LT |
1167 | if (*ppos > max) |
1168 | retval = -EOVERFLOW; | |
1169 | ||
1170 | fput_out: | |
2903ff01 | 1171 | fdput(out); |
1da177e4 | 1172 | fput_in: |
2903ff01 | 1173 | fdput(in); |
1da177e4 LT |
1174 | out: |
1175 | return retval; | |
1176 | } | |
1177 | ||
002c8976 | 1178 | SYSCALL_DEFINE4(sendfile, int, out_fd, int, in_fd, off_t __user *, offset, size_t, count) |
1da177e4 LT |
1179 | { |
1180 | loff_t pos; | |
1181 | off_t off; | |
1182 | ssize_t ret; | |
1183 | ||
1184 | if (offset) { | |
1185 | if (unlikely(get_user(off, offset))) | |
1186 | return -EFAULT; | |
1187 | pos = off; | |
1188 | ret = do_sendfile(out_fd, in_fd, &pos, count, MAX_NON_LFS); | |
1189 | if (unlikely(put_user(pos, offset))) | |
1190 | return -EFAULT; | |
1191 | return ret; | |
1192 | } | |
1193 | ||
1194 | return do_sendfile(out_fd, in_fd, NULL, count, 0); | |
1195 | } | |
1196 | ||
002c8976 | 1197 | SYSCALL_DEFINE4(sendfile64, int, out_fd, int, in_fd, loff_t __user *, offset, size_t, count) |
1da177e4 LT |
1198 | { |
1199 | loff_t pos; | |
1200 | ssize_t ret; | |
1201 | ||
1202 | if (offset) { | |
1203 | if (unlikely(copy_from_user(&pos, offset, sizeof(loff_t)))) | |
1204 | return -EFAULT; | |
1205 | ret = do_sendfile(out_fd, in_fd, &pos, count, 0); | |
1206 | if (unlikely(put_user(pos, offset))) | |
1207 | return -EFAULT; | |
1208 | return ret; | |
1209 | } | |
1210 | ||
1211 | return do_sendfile(out_fd, in_fd, NULL, count, 0); | |
1212 | } | |
19f4fc3a AV |
1213 | |
1214 | #ifdef CONFIG_COMPAT | |
1215 | COMPAT_SYSCALL_DEFINE4(sendfile, int, out_fd, int, in_fd, | |
1216 | compat_off_t __user *, offset, compat_size_t, count) | |
1217 | { | |
1218 | loff_t pos; | |
1219 | off_t off; | |
1220 | ssize_t ret; | |
1221 | ||
1222 | if (offset) { | |
1223 | if (unlikely(get_user(off, offset))) | |
1224 | return -EFAULT; | |
1225 | pos = off; | |
1226 | ret = do_sendfile(out_fd, in_fd, &pos, count, MAX_NON_LFS); | |
1227 | if (unlikely(put_user(pos, offset))) | |
1228 | return -EFAULT; | |
1229 | return ret; | |
1230 | } | |
1231 | ||
1232 | return do_sendfile(out_fd, in_fd, NULL, count, 0); | |
1233 | } | |
1234 | ||
1235 | COMPAT_SYSCALL_DEFINE4(sendfile64, int, out_fd, int, in_fd, | |
1236 | compat_loff_t __user *, offset, compat_size_t, count) | |
1237 | { | |
1238 | loff_t pos; | |
1239 | ssize_t ret; | |
1240 | ||
1241 | if (offset) { | |
1242 | if (unlikely(copy_from_user(&pos, offset, sizeof(loff_t)))) | |
1243 | return -EFAULT; | |
1244 | ret = do_sendfile(out_fd, in_fd, &pos, count, 0); | |
1245 | if (unlikely(put_user(pos, offset))) | |
1246 | return -EFAULT; | |
1247 | return ret; | |
1248 | } | |
1249 | ||
1250 | return do_sendfile(out_fd, in_fd, NULL, count, 0); | |
1251 | } | |
1252 | #endif |