]>
git.ipfire.org Git - thirdparty/systemd.git/blob - src/basic/copy.c
1 /* SPDX-License-Identifier: LGPL-2.1+ */
3 Copyright 2014 Lennart Poettering
13 #include <sys/sendfile.h>
15 #include <sys/xattr.h>
19 #include "alloc-util.h"
20 #include "btrfs-util.h"
21 #include "chattr-util.h"
23 #include "dirent-util.h"
30 #include "mount-util.h"
31 #include "string-util.h"
33 #include "time-util.h"
34 #include "umask-util.h"
35 #include "user-util.h"
36 #include "xattr-util.h"
38 #define COPY_BUFFER_SIZE (16U*1024U)
40 /* A safety net for descending recursively into file system trees to copy. On Linux PATH_MAX is 4096, which means the
41 * deepest valid path one can build is around 2048, which we hence use as a safety net here, to not spin endlessly in
42 * case of bind mount cycles and suchlike. */
43 #define COPY_DEPTH_MAX 2048U
45 static ssize_t
try_copy_file_range(
46 int fd_in
, loff_t
*off_in
,
47 int fd_out
, loff_t
*off_out
,
57 r
= copy_file_range(fd_in
, off_in
, fd_out
, off_out
, len
, flags
);
59 have
= r
>= 0 || errno
!= ENOSYS
;
69 FD_IS_NONBLOCKING_PIPE
,
72 static int fd_is_nonblock_pipe(int fd
) {
76 /* Checks whether the specified file descriptor refers to a pipe, and if so if O_NONBLOCK is set. */
78 if (fstat(fd
, &st
) < 0)
81 if (!S_ISFIFO(st
.st_mode
))
84 flags
= fcntl(fd
, F_GETFL
);
88 return FLAGS_SET(flags
, O_NONBLOCK
) ? FD_IS_NONBLOCKING_PIPE
: FD_IS_BLOCKING_PIPE
;
96 size_t *ret_remains_size
) {
98 bool try_cfr
= true, try_sendfile
= true, try_splice
= true;
99 int r
, nonblock_pipe
= -1;
100 size_t m
= SSIZE_MAX
; /* that is the maximum that sendfile and c_f_r accept */
105 /* Tries to copy bytes from the file descriptor 'fdf' to 'fdt' in the smartest possible way. Copies a maximum
106 * of 'max_bytes', which may be specified as UINT64_MAX, in which no maximum is applied. Returns negative on
107 * error, zero if EOF is hit before the bytes limit is hit and positive otherwise. If the copy fails for some
108 * reason but we read but didn't yet write some data an ret_remains/ret_remains_size is not NULL, then it will
109 * be initialized with an allocated buffer containing this "remaining" data. Note that these two parameters are
110 * initialized with a valid buffer only on failure and only if there's actually data already read. Otherwise
111 * these parameters if non-NULL are set to NULL. */
115 if (ret_remains_size
)
116 *ret_remains_size
= 0;
118 /* Try btrfs reflinks first. This only works on regular, seekable files, hence let's check the file offsets of
119 * source and destination first. */
120 if ((copy_flags
& COPY_REFLINK
)) {
123 foffset
= lseek(fdf
, 0, SEEK_CUR
);
127 toffset
= lseek(fdt
, 0, SEEK_CUR
);
130 if (foffset
== 0 && toffset
== 0 && max_bytes
== UINT64_MAX
)
131 r
= btrfs_reflink(fdf
, fdt
); /* full file reflink */
133 r
= btrfs_clone_range(fdf
, foffset
, fdt
, toffset
, max_bytes
== UINT64_MAX
? 0 : max_bytes
); /* partial reflink */
137 /* This worked, yay! Now — to be fully correct — let's adjust the file pointers */
138 if (max_bytes
== UINT64_MAX
) {
140 /* We cloned to the end of the source file, let's position the read
141 * pointer there, and query it at the same time. */
142 t
= lseek(fdf
, 0, SEEK_END
);
148 /* Let's adjust the destination file write pointer by the same number
150 t
= lseek(fdt
, toffset
+ (t
- foffset
), SEEK_SET
);
154 return 0; /* we copied the whole thing, hence hit EOF, return 0 */
156 t
= lseek(fdf
, foffset
+ max_bytes
, SEEK_SET
);
160 t
= lseek(fdt
, toffset
+ max_bytes
, SEEK_SET
);
164 return 1; /* we copied only some number of bytes, which worked, but this means we didn't hit EOF, return 1 */
168 log_debug_errno(r
, "Reflinking didn't work, falling back to non-reflink copying: %m");
177 return 1; /* return > 0 if we hit the max_bytes limit */
179 if (max_bytes
!= UINT64_MAX
&& m
> max_bytes
)
182 /* First try copy_file_range(), unless we already tried */
184 n
= try_copy_file_range(fdf
, NULL
, fdt
, NULL
, m
, 0u);
186 if (!IN_SET(n
, -EINVAL
, -ENOSYS
, -EXDEV
, -EBADF
))
190 /* use fallback below */
191 } else if (n
== 0) /* EOF */
198 /* First try sendfile(), unless we already tried */
200 n
= sendfile(fdt
, fdf
, NULL
, m
);
202 if (!IN_SET(errno
, EINVAL
, ENOSYS
))
205 try_sendfile
= false;
206 /* use fallback below */
207 } else if (n
== 0) /* EOF */
214 /* Then try splice, unless we already tried. */
217 /* splice()'s asynchronous I/O support is a bit weird. When it encounters a pipe file
218 * descriptor, then it will ignore its O_NONBLOCK flag and instead only honour the
219 * SPLICE_F_NONBLOCK flag specified in its flag parameter. Let's hide this behaviour here, and
220 * check if either of the specified fds are a pipe, and if so, let's pass the flag
221 * automatically, depending on O_NONBLOCK being set.
223 * Here's a twist though: when we use it to move data between two pipes of which one has
224 * O_NONBLOCK set and the other has not, then we have no individual control over O_NONBLOCK
225 * behaviour. Hence in that case we can't use splice() and still guarantee systematic
226 * O_NONBLOCK behaviour, hence don't. */
228 if (nonblock_pipe
< 0) {
231 /* Check if either of these fds is a pipe, and if so non-blocking or not */
232 a
= fd_is_nonblock_pipe(fdf
);
236 b
= fd_is_nonblock_pipe(fdt
);
240 if ((a
== FD_IS_NO_PIPE
&& b
== FD_IS_NO_PIPE
) ||
241 (a
== FD_IS_BLOCKING_PIPE
&& b
== FD_IS_NONBLOCKING_PIPE
) ||
242 (a
== FD_IS_NONBLOCKING_PIPE
&& b
== FD_IS_BLOCKING_PIPE
))
244 /* splice() only works if one of the fds is a pipe. If neither is, let's skip
245 * this step right-away. As mentioned above, if one of the two fds refers to a
246 * blocking pipe and the other to a non-blocking pipe, we can't use splice()
247 * either, hence don't try either. This hence means we can only use splice() if
248 * either only one of the two fds is a pipe, or if both are pipes with the same
249 * nonblocking flag setting. */
253 nonblock_pipe
= a
== FD_IS_NONBLOCKING_PIPE
|| b
== FD_IS_NONBLOCKING_PIPE
;
258 n
= splice(fdf
, NULL
, fdt
, NULL
, m
, nonblock_pipe
? SPLICE_F_NONBLOCK
: 0);
260 if (!IN_SET(errno
, EINVAL
, ENOSYS
))
264 /* use fallback below */
265 } else if (n
== 0) /* EOF */
272 /* As a fallback just copy bits by hand */
274 uint8_t buf
[MIN(m
, COPY_BUFFER_SIZE
)], *p
= buf
;
277 n
= read(fdf
, buf
, sizeof buf
);
280 if (n
== 0) /* EOF */
287 k
= write(fdt
, p
, z
);
301 if (ret_remains_size
)
302 *ret_remains_size
= z
;
314 if (max_bytes
!= (uint64_t) -1) {
315 assert(max_bytes
>= (uint64_t) n
);
318 /* sendfile accepts at most SSIZE_MAX-offset bytes to copy,
319 * so reduce our maximum by the amount we already copied,
320 * but don't go below our copy buffer size, unless we are
321 * close the limit of bytes we are allowed to copy. */
322 m
= MAX(MIN(COPY_BUFFER_SIZE
, max_bytes
), m
- n
);
325 return 0; /* return 0 if we hit EOF earlier than the size limit */
328 static int fd_copy_symlink(
331 const struct stat
*st
,
336 CopyFlags copy_flags
) {
338 _cleanup_free_
char *target
= NULL
;
345 r
= readlinkat_malloc(df
, from
, &target
);
349 if (symlinkat(target
, dt
, to
) < 0)
353 uid_is_valid(override_uid
) ? override_uid
: st
->st_uid
,
354 gid_is_valid(override_gid
) ? override_gid
: st
->st_gid
,
355 AT_SYMLINK_NOFOLLOW
) < 0)
361 static int fd_copy_regular(
364 const struct stat
*st
,
369 CopyFlags copy_flags
) {
371 _cleanup_close_
int fdf
= -1, fdt
= -1;
372 struct timespec ts
[2];
379 fdf
= openat(df
, from
, O_RDONLY
|O_CLOEXEC
|O_NOCTTY
|O_NOFOLLOW
);
383 fdt
= openat(dt
, to
, O_WRONLY
|O_CREAT
|O_EXCL
|O_CLOEXEC
|O_NOCTTY
|O_NOFOLLOW
, st
->st_mode
& 07777);
387 r
= copy_bytes(fdf
, fdt
, (uint64_t) -1, copy_flags
);
389 (void) unlinkat(dt
, to
, 0);
394 uid_is_valid(override_uid
) ? override_uid
: st
->st_uid
,
395 gid_is_valid(override_gid
) ? override_gid
: st
->st_gid
) < 0)
398 if (fchmod(fdt
, st
->st_mode
& 07777) < 0)
403 (void) futimens(fdt
, ts
);
404 (void) copy_xattr(fdf
, fdt
);
411 (void) unlinkat(dt
, to
, 0);
417 static int fd_copy_fifo(
420 const struct stat
*st
,
425 CopyFlags copy_flags
) {
432 r
= mkfifoat(dt
, to
, st
->st_mode
& 07777);
437 uid_is_valid(override_uid
) ? override_uid
: st
->st_uid
,
438 gid_is_valid(override_gid
) ? override_gid
: st
->st_gid
,
439 AT_SYMLINK_NOFOLLOW
) < 0)
442 if (fchmodat(dt
, to
, st
->st_mode
& 07777, 0) < 0)
448 static int fd_copy_node(
451 const struct stat
*st
,
456 CopyFlags copy_flags
) {
463 r
= mknodat(dt
, to
, st
->st_mode
, st
->st_rdev
);
468 uid_is_valid(override_uid
) ? override_uid
: st
->st_uid
,
469 gid_is_valid(override_gid
) ? override_gid
: st
->st_gid
,
470 AT_SYMLINK_NOFOLLOW
) < 0)
473 if (fchmodat(dt
, to
, st
->st_mode
& 07777, 0) < 0)
479 static int fd_copy_directory(
482 const struct stat
*st
,
485 dev_t original_device
,
489 CopyFlags copy_flags
) {
491 _cleanup_close_
int fdf
= -1, fdt
= -1;
492 _cleanup_closedir_
DIR *d
= NULL
;
501 return -ENAMETOOLONG
;
504 fdf
= openat(df
, from
, O_RDONLY
|O_DIRECTORY
|O_CLOEXEC
|O_NOCTTY
|O_NOFOLLOW
);
506 fdf
= fcntl(df
, F_DUPFD_CLOEXEC
, 3);
515 r
= mkdirat(dt
, to
, st
->st_mode
& 07777);
518 else if (errno
== EEXIST
&& (copy_flags
& COPY_MERGE
))
523 fdt
= openat(dt
, to
, O_RDONLY
|O_DIRECTORY
|O_CLOEXEC
|O_NOCTTY
|O_NOFOLLOW
);
529 FOREACH_DIRENT_ALL(de
, d
, return -errno
) {
533 if (dot_or_dot_dot(de
->d_name
))
536 if (fstatat(dirfd(d
), de
->d_name
, &buf
, AT_SYMLINK_NOFOLLOW
) < 0) {
541 if (S_ISDIR(buf
.st_mode
)) {
543 * Don't descend into directories on other file systems, if this is requested. We do a simple
544 * .st_dev check here, which basically comes for free. Note that we do this check only on
545 * directories, not other kind of file system objects, for two reason:
547 * • The kernel's overlayfs pseudo file system that overlays multiple real file systems
548 * propagates the .st_dev field of the file system a file originates from all the way up
549 * through the stack to stat(). It doesn't do that for directories however. This means that
550 * comparing .st_dev on non-directories suggests that they all are mount points. To avoid
551 * confusion we hence avoid relying on this check for regular files.
553 * • The main reason we do this check at all is to protect ourselves from bind mount cycles,
554 * where we really want to avoid descending down in all eternity. However the .st_dev check
555 * is usually not sufficient for this protection anyway, as bind mount cycles from the same
556 * file system onto itself can't be detected that way. (Note we also do a recursion depth
557 * check, which is probably the better protection in this regard, which is why
558 * COPY_SAME_MOUNT is optional).
561 if (FLAGS_SET(copy_flags
, COPY_SAME_MOUNT
)) {
562 if (buf
.st_dev
!= original_device
)
565 r
= fd_is_mount_point(dirfd(d
), de
->d_name
, 0);
572 q
= fd_copy_directory(dirfd(d
), de
->d_name
, &buf
, fdt
, de
->d_name
, original_device
, depth_left
-1, override_uid
, override_gid
, copy_flags
);
573 } else if (S_ISREG(buf
.st_mode
))
574 q
= fd_copy_regular(dirfd(d
), de
->d_name
, &buf
, fdt
, de
->d_name
, override_uid
, override_gid
, copy_flags
);
575 else if (S_ISLNK(buf
.st_mode
))
576 q
= fd_copy_symlink(dirfd(d
), de
->d_name
, &buf
, fdt
, de
->d_name
, override_uid
, override_gid
, copy_flags
);
577 else if (S_ISFIFO(buf
.st_mode
))
578 q
= fd_copy_fifo(dirfd(d
), de
->d_name
, &buf
, fdt
, de
->d_name
, override_uid
, override_gid
, copy_flags
);
579 else if (S_ISBLK(buf
.st_mode
) || S_ISCHR(buf
.st_mode
) || S_ISSOCK(buf
.st_mode
))
580 q
= fd_copy_node(dirfd(d
), de
->d_name
, &buf
, fdt
, de
->d_name
, override_uid
, override_gid
, copy_flags
);
584 if (q
== -EEXIST
&& (copy_flags
& COPY_MERGE
))
592 struct timespec ut
[2] = {
598 uid_is_valid(override_uid
) ? override_uid
: st
->st_uid
,
599 gid_is_valid(override_gid
) ? override_gid
: st
->st_gid
) < 0)
602 if (fchmod(fdt
, st
->st_mode
& 07777) < 0)
605 (void) copy_xattr(dirfd(d
), fdt
);
606 (void) futimens(fdt
, ut
);
612 int copy_tree_at(int fdf
, const char *from
, int fdt
, const char *to
, uid_t override_uid
, gid_t override_gid
, CopyFlags copy_flags
) {
618 if (fstatat(fdf
, from
, &st
, AT_SYMLINK_NOFOLLOW
) < 0)
621 if (S_ISREG(st
.st_mode
))
622 return fd_copy_regular(fdf
, from
, &st
, fdt
, to
, override_uid
, override_gid
, copy_flags
);
623 else if (S_ISDIR(st
.st_mode
))
624 return fd_copy_directory(fdf
, from
, &st
, fdt
, to
, st
.st_dev
, COPY_DEPTH_MAX
, override_uid
, override_gid
, copy_flags
);
625 else if (S_ISLNK(st
.st_mode
))
626 return fd_copy_symlink(fdf
, from
, &st
, fdt
, to
, override_uid
, override_gid
, copy_flags
);
627 else if (S_ISFIFO(st
.st_mode
))
628 return fd_copy_fifo(fdf
, from
, &st
, fdt
, to
, override_uid
, override_gid
, copy_flags
);
629 else if (S_ISBLK(st
.st_mode
) || S_ISCHR(st
.st_mode
) || S_ISSOCK(st
.st_mode
))
630 return fd_copy_node(fdf
, from
, &st
, fdt
, to
, override_uid
, override_gid
, copy_flags
);
635 int copy_tree(const char *from
, const char *to
, uid_t override_uid
, gid_t override_gid
, CopyFlags copy_flags
) {
636 return copy_tree_at(AT_FDCWD
, from
, AT_FDCWD
, to
, override_uid
, override_gid
, copy_flags
);
639 int copy_directory_fd(int dirfd
, const char *to
, CopyFlags copy_flags
) {
645 if (fstat(dirfd
, &st
) < 0)
648 if (!S_ISDIR(st
.st_mode
))
651 return fd_copy_directory(dirfd
, NULL
, &st
, AT_FDCWD
, to
, st
.st_dev
, COPY_DEPTH_MAX
, UID_INVALID
, GID_INVALID
, copy_flags
);
654 int copy_directory(const char *from
, const char *to
, CopyFlags copy_flags
) {
660 if (lstat(from
, &st
) < 0)
663 if (!S_ISDIR(st
.st_mode
))
666 return fd_copy_directory(AT_FDCWD
, from
, &st
, AT_FDCWD
, to
, st
.st_dev
, COPY_DEPTH_MAX
, UID_INVALID
, GID_INVALID
, copy_flags
);
669 int copy_file_fd(const char *from
, int fdt
, CopyFlags copy_flags
) {
670 _cleanup_close_
int fdf
= -1;
676 fdf
= open(from
, O_RDONLY
|O_CLOEXEC
|O_NOCTTY
);
680 r
= copy_bytes(fdf
, fdt
, (uint64_t) -1, copy_flags
);
682 (void) copy_times(fdf
, fdt
);
683 (void) copy_xattr(fdf
, fdt
);
688 int copy_file(const char *from
, const char *to
, int flags
, mode_t mode
, unsigned chattr_flags
, CopyFlags copy_flags
) {
694 RUN_WITH_UMASK(0000) {
695 fdt
= open(to
, flags
|O_WRONLY
|O_CREAT
|O_CLOEXEC
|O_NOCTTY
, mode
);
700 if (chattr_flags
!= 0)
701 (void) chattr_fd(fdt
, chattr_flags
, (unsigned) -1);
703 r
= copy_file_fd(from
, fdt
, copy_flags
);
710 if (close(fdt
) < 0) {
718 int copy_file_atomic(const char *from
, const char *to
, mode_t mode
, unsigned chattr_flags
, CopyFlags copy_flags
) {
719 _cleanup_free_
char *t
= NULL
;
725 r
= tempfn_random(to
, NULL
, &t
);
729 r
= copy_file(from
, t
, O_NOFOLLOW
|O_EXCL
, mode
, chattr_flags
, copy_flags
);
733 if (copy_flags
& COPY_REPLACE
) {
734 r
= renameat(AT_FDCWD
, t
, AT_FDCWD
, to
);
738 r
= rename_noreplace(AT_FDCWD
, t
, AT_FDCWD
, to
);
747 int copy_times(int fdf
, int fdt
) {
748 struct timespec ut
[2];
755 if (fstat(fdf
, &st
) < 0)
761 if (futimens(fdt
, ut
) < 0)
764 if (fd_getcrtime(fdf
, &crtime
) >= 0)
765 (void) fd_setcrtime(fdt
, crtime
);
770 int copy_xattr(int fdf
, int fdt
) {
771 _cleanup_free_
char *bufa
= NULL
, *bufb
= NULL
;
772 size_t sza
= 100, szb
= 100;
782 n
= flistxattr(fdf
, bufa
, sza
);
800 assert(l
< (size_t) n
);
802 if (startswith(p
, "user.")) {
811 m
= fgetxattr(fdf
, p
, bufb
, szb
);
813 if (errno
== ERANGE
) {
822 if (fsetxattr(fdt
, p
, bufb
, m
, 0) < 0)