]>
git.ipfire.org Git - thirdparty/systemd.git/blob - src/basic/copy.c
1 /* SPDX-License-Identifier: LGPL-2.1+ */
10 #include <sys/sendfile.h>
12 #include <sys/xattr.h>
16 #include "alloc-util.h"
17 #include "btrfs-util.h"
18 #include "chattr-util.h"
20 #include "dirent-util.h"
27 #include "mount-util.h"
28 #include "string-util.h"
30 #include "time-util.h"
31 #include "umask-util.h"
32 #include "user-util.h"
33 #include "xattr-util.h"
35 #define COPY_BUFFER_SIZE (16U*1024U)
37 /* A safety net for descending recursively into file system trees to copy. On Linux PATH_MAX is 4096, which means the
38 * deepest valid path one can build is around 2048, which we hence use as a safety net here, to not spin endlessly in
39 * case of bind mount cycles and suchlike. */
40 #define COPY_DEPTH_MAX 2048U
42 static ssize_t
try_copy_file_range(
43 int fd_in
, loff_t
*off_in
,
44 int fd_out
, loff_t
*off_out
,
54 r
= copy_file_range(fd_in
, off_in
, fd_out
, off_out
, len
, flags
);
56 have
= r
>= 0 || errno
!= ENOSYS
;
66 FD_IS_NONBLOCKING_PIPE
,
69 static int fd_is_nonblock_pipe(int fd
) {
73 /* Checks whether the specified file descriptor refers to a pipe, and if so if O_NONBLOCK is set. */
75 if (fstat(fd
, &st
) < 0)
78 if (!S_ISFIFO(st
.st_mode
))
81 flags
= fcntl(fd
, F_GETFL
);
85 return FLAGS_SET(flags
, O_NONBLOCK
) ? FD_IS_NONBLOCKING_PIPE
: FD_IS_BLOCKING_PIPE
;
93 size_t *ret_remains_size
) {
95 bool try_cfr
= true, try_sendfile
= true, try_splice
= true;
96 int r
, nonblock_pipe
= -1;
97 size_t m
= SSIZE_MAX
; /* that is the maximum that sendfile and c_f_r accept */
102 /* Tries to copy bytes from the file descriptor 'fdf' to 'fdt' in the smartest possible way. Copies a maximum
103 * of 'max_bytes', which may be specified as UINT64_MAX, in which no maximum is applied. Returns negative on
104 * error, zero if EOF is hit before the bytes limit is hit and positive otherwise. If the copy fails for some
105 * reason but we read but didn't yet write some data an ret_remains/ret_remains_size is not NULL, then it will
106 * be initialized with an allocated buffer containing this "remaining" data. Note that these two parameters are
107 * initialized with a valid buffer only on failure and only if there's actually data already read. Otherwise
108 * these parameters if non-NULL are set to NULL. */
112 if (ret_remains_size
)
113 *ret_remains_size
= 0;
115 /* Try btrfs reflinks first. This only works on regular, seekable files, hence let's check the file offsets of
116 * source and destination first. */
117 if ((copy_flags
& COPY_REFLINK
)) {
120 foffset
= lseek(fdf
, 0, SEEK_CUR
);
124 toffset
= lseek(fdt
, 0, SEEK_CUR
);
127 if (foffset
== 0 && toffset
== 0 && max_bytes
== UINT64_MAX
)
128 r
= btrfs_reflink(fdf
, fdt
); /* full file reflink */
130 r
= btrfs_clone_range(fdf
, foffset
, fdt
, toffset
, max_bytes
== UINT64_MAX
? 0 : max_bytes
); /* partial reflink */
134 /* This worked, yay! Now — to be fully correct — let's adjust the file pointers */
135 if (max_bytes
== UINT64_MAX
) {
137 /* We cloned to the end of the source file, let's position the read
138 * pointer there, and query it at the same time. */
139 t
= lseek(fdf
, 0, SEEK_END
);
145 /* Let's adjust the destination file write pointer by the same number
147 t
= lseek(fdt
, toffset
+ (t
- foffset
), SEEK_SET
);
151 return 0; /* we copied the whole thing, hence hit EOF, return 0 */
153 t
= lseek(fdf
, foffset
+ max_bytes
, SEEK_SET
);
157 t
= lseek(fdt
, toffset
+ max_bytes
, SEEK_SET
);
161 return 1; /* we copied only some number of bytes, which worked, but this means we didn't hit EOF, return 1 */
165 log_debug_errno(r
, "Reflinking didn't work, falling back to non-reflink copying: %m");
174 return 1; /* return > 0 if we hit the max_bytes limit */
176 if (max_bytes
!= UINT64_MAX
&& m
> max_bytes
)
179 /* First try copy_file_range(), unless we already tried */
181 n
= try_copy_file_range(fdf
, NULL
, fdt
, NULL
, m
, 0u);
183 if (!IN_SET(n
, -EINVAL
, -ENOSYS
, -EXDEV
, -EBADF
))
187 /* use fallback below */
188 } else if (n
== 0) /* EOF */
195 /* First try sendfile(), unless we already tried */
197 n
= sendfile(fdt
, fdf
, NULL
, m
);
199 if (!IN_SET(errno
, EINVAL
, ENOSYS
))
202 try_sendfile
= false;
203 /* use fallback below */
204 } else if (n
== 0) /* EOF */
211 /* Then try splice, unless we already tried. */
214 /* splice()'s asynchronous I/O support is a bit weird. When it encounters a pipe file
215 * descriptor, then it will ignore its O_NONBLOCK flag and instead only honour the
216 * SPLICE_F_NONBLOCK flag specified in its flag parameter. Let's hide this behaviour here, and
217 * check if either of the specified fds are a pipe, and if so, let's pass the flag
218 * automatically, depending on O_NONBLOCK being set.
220 * Here's a twist though: when we use it to move data between two pipes of which one has
221 * O_NONBLOCK set and the other has not, then we have no individual control over O_NONBLOCK
222 * behaviour. Hence in that case we can't use splice() and still guarantee systematic
223 * O_NONBLOCK behaviour, hence don't. */
225 if (nonblock_pipe
< 0) {
228 /* Check if either of these fds is a pipe, and if so non-blocking or not */
229 a
= fd_is_nonblock_pipe(fdf
);
233 b
= fd_is_nonblock_pipe(fdt
);
237 if ((a
== FD_IS_NO_PIPE
&& b
== FD_IS_NO_PIPE
) ||
238 (a
== FD_IS_BLOCKING_PIPE
&& b
== FD_IS_NONBLOCKING_PIPE
) ||
239 (a
== FD_IS_NONBLOCKING_PIPE
&& b
== FD_IS_BLOCKING_PIPE
))
241 /* splice() only works if one of the fds is a pipe. If neither is, let's skip
242 * this step right-away. As mentioned above, if one of the two fds refers to a
243 * blocking pipe and the other to a non-blocking pipe, we can't use splice()
244 * either, hence don't try either. This hence means we can only use splice() if
245 * either only one of the two fds is a pipe, or if both are pipes with the same
246 * nonblocking flag setting. */
250 nonblock_pipe
= a
== FD_IS_NONBLOCKING_PIPE
|| b
== FD_IS_NONBLOCKING_PIPE
;
255 n
= splice(fdf
, NULL
, fdt
, NULL
, m
, nonblock_pipe
? SPLICE_F_NONBLOCK
: 0);
257 if (!IN_SET(errno
, EINVAL
, ENOSYS
))
261 /* use fallback below */
262 } else if (n
== 0) /* EOF */
269 /* As a fallback just copy bits by hand */
271 uint8_t buf
[MIN(m
, COPY_BUFFER_SIZE
)], *p
= buf
;
274 n
= read(fdf
, buf
, sizeof buf
);
277 if (n
== 0) /* EOF */
284 k
= write(fdt
, p
, z
);
298 if (ret_remains_size
)
299 *ret_remains_size
= z
;
311 if (max_bytes
!= (uint64_t) -1) {
312 assert(max_bytes
>= (uint64_t) n
);
315 /* sendfile accepts at most SSIZE_MAX-offset bytes to copy,
316 * so reduce our maximum by the amount we already copied,
317 * but don't go below our copy buffer size, unless we are
318 * close the limit of bytes we are allowed to copy. */
319 m
= MAX(MIN(COPY_BUFFER_SIZE
, max_bytes
), m
- n
);
322 return 0; /* return 0 if we hit EOF earlier than the size limit */
325 static int fd_copy_symlink(
328 const struct stat
*st
,
333 CopyFlags copy_flags
) {
335 _cleanup_free_
char *target
= NULL
;
342 r
= readlinkat_malloc(df
, from
, &target
);
346 if (symlinkat(target
, dt
, to
) < 0)
350 uid_is_valid(override_uid
) ? override_uid
: st
->st_uid
,
351 gid_is_valid(override_gid
) ? override_gid
: st
->st_gid
,
352 AT_SYMLINK_NOFOLLOW
) < 0)
358 static int fd_copy_regular(
361 const struct stat
*st
,
366 CopyFlags copy_flags
) {
368 _cleanup_close_
int fdf
= -1, fdt
= -1;
369 struct timespec ts
[2];
376 fdf
= openat(df
, from
, O_RDONLY
|O_CLOEXEC
|O_NOCTTY
|O_NOFOLLOW
);
380 fdt
= openat(dt
, to
, O_WRONLY
|O_CREAT
|O_EXCL
|O_CLOEXEC
|O_NOCTTY
|O_NOFOLLOW
, st
->st_mode
& 07777);
384 r
= copy_bytes(fdf
, fdt
, (uint64_t) -1, copy_flags
);
386 (void) unlinkat(dt
, to
, 0);
391 uid_is_valid(override_uid
) ? override_uid
: st
->st_uid
,
392 gid_is_valid(override_gid
) ? override_gid
: st
->st_gid
) < 0)
395 if (fchmod(fdt
, st
->st_mode
& 07777) < 0)
400 (void) futimens(fdt
, ts
);
401 (void) copy_xattr(fdf
, fdt
);
408 (void) unlinkat(dt
, to
, 0);
414 static int fd_copy_fifo(
417 const struct stat
*st
,
422 CopyFlags copy_flags
) {
429 r
= mkfifoat(dt
, to
, st
->st_mode
& 07777);
434 uid_is_valid(override_uid
) ? override_uid
: st
->st_uid
,
435 gid_is_valid(override_gid
) ? override_gid
: st
->st_gid
,
436 AT_SYMLINK_NOFOLLOW
) < 0)
439 if (fchmodat(dt
, to
, st
->st_mode
& 07777, 0) < 0)
445 static int fd_copy_node(
448 const struct stat
*st
,
453 CopyFlags copy_flags
) {
460 r
= mknodat(dt
, to
, st
->st_mode
, st
->st_rdev
);
465 uid_is_valid(override_uid
) ? override_uid
: st
->st_uid
,
466 gid_is_valid(override_gid
) ? override_gid
: st
->st_gid
,
467 AT_SYMLINK_NOFOLLOW
) < 0)
470 if (fchmodat(dt
, to
, st
->st_mode
& 07777, 0) < 0)
476 static int fd_copy_directory(
479 const struct stat
*st
,
482 dev_t original_device
,
486 CopyFlags copy_flags
) {
488 _cleanup_close_
int fdf
= -1, fdt
= -1;
489 _cleanup_closedir_
DIR *d
= NULL
;
498 return -ENAMETOOLONG
;
501 fdf
= openat(df
, from
, O_RDONLY
|O_DIRECTORY
|O_CLOEXEC
|O_NOCTTY
|O_NOFOLLOW
);
503 fdf
= fcntl(df
, F_DUPFD_CLOEXEC
, 3);
512 r
= mkdirat(dt
, to
, st
->st_mode
& 07777);
515 else if (errno
== EEXIST
&& (copy_flags
& COPY_MERGE
))
520 fdt
= openat(dt
, to
, O_RDONLY
|O_DIRECTORY
|O_CLOEXEC
|O_NOCTTY
|O_NOFOLLOW
);
526 FOREACH_DIRENT_ALL(de
, d
, return -errno
) {
530 if (dot_or_dot_dot(de
->d_name
))
533 if (fstatat(dirfd(d
), de
->d_name
, &buf
, AT_SYMLINK_NOFOLLOW
) < 0) {
538 if (S_ISDIR(buf
.st_mode
)) {
540 * Don't descend into directories on other file systems, if this is requested. We do a simple
541 * .st_dev check here, which basically comes for free. Note that we do this check only on
542 * directories, not other kind of file system objects, for two reason:
544 * • The kernel's overlayfs pseudo file system that overlays multiple real file systems
545 * propagates the .st_dev field of the file system a file originates from all the way up
546 * through the stack to stat(). It doesn't do that for directories however. This means that
547 * comparing .st_dev on non-directories suggests that they all are mount points. To avoid
548 * confusion we hence avoid relying on this check for regular files.
550 * • The main reason we do this check at all is to protect ourselves from bind mount cycles,
551 * where we really want to avoid descending down in all eternity. However the .st_dev check
552 * is usually not sufficient for this protection anyway, as bind mount cycles from the same
553 * file system onto itself can't be detected that way. (Note we also do a recursion depth
554 * check, which is probably the better protection in this regard, which is why
555 * COPY_SAME_MOUNT is optional).
558 if (FLAGS_SET(copy_flags
, COPY_SAME_MOUNT
)) {
559 if (buf
.st_dev
!= original_device
)
562 r
= fd_is_mount_point(dirfd(d
), de
->d_name
, 0);
569 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
);
570 } else if (S_ISREG(buf
.st_mode
))
571 q
= fd_copy_regular(dirfd(d
), de
->d_name
, &buf
, fdt
, de
->d_name
, override_uid
, override_gid
, copy_flags
);
572 else if (S_ISLNK(buf
.st_mode
))
573 q
= fd_copy_symlink(dirfd(d
), de
->d_name
, &buf
, fdt
, de
->d_name
, override_uid
, override_gid
, copy_flags
);
574 else if (S_ISFIFO(buf
.st_mode
))
575 q
= fd_copy_fifo(dirfd(d
), de
->d_name
, &buf
, fdt
, de
->d_name
, override_uid
, override_gid
, copy_flags
);
576 else if (S_ISBLK(buf
.st_mode
) || S_ISCHR(buf
.st_mode
) || S_ISSOCK(buf
.st_mode
))
577 q
= fd_copy_node(dirfd(d
), de
->d_name
, &buf
, fdt
, de
->d_name
, override_uid
, override_gid
, copy_flags
);
581 if (q
== -EEXIST
&& (copy_flags
& COPY_MERGE
))
589 struct timespec ut
[2] = {
595 uid_is_valid(override_uid
) ? override_uid
: st
->st_uid
,
596 gid_is_valid(override_gid
) ? override_gid
: st
->st_gid
) < 0)
599 if (fchmod(fdt
, st
->st_mode
& 07777) < 0)
602 (void) copy_xattr(dirfd(d
), fdt
);
603 (void) futimens(fdt
, ut
);
609 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
) {
615 if (fstatat(fdf
, from
, &st
, AT_SYMLINK_NOFOLLOW
) < 0)
618 if (S_ISREG(st
.st_mode
))
619 return fd_copy_regular(fdf
, from
, &st
, fdt
, to
, override_uid
, override_gid
, copy_flags
);
620 else if (S_ISDIR(st
.st_mode
))
621 return fd_copy_directory(fdf
, from
, &st
, fdt
, to
, st
.st_dev
, COPY_DEPTH_MAX
, override_uid
, override_gid
, copy_flags
);
622 else if (S_ISLNK(st
.st_mode
))
623 return fd_copy_symlink(fdf
, from
, &st
, fdt
, to
, override_uid
, override_gid
, copy_flags
);
624 else if (S_ISFIFO(st
.st_mode
))
625 return fd_copy_fifo(fdf
, from
, &st
, fdt
, to
, override_uid
, override_gid
, copy_flags
);
626 else if (S_ISBLK(st
.st_mode
) || S_ISCHR(st
.st_mode
) || S_ISSOCK(st
.st_mode
))
627 return fd_copy_node(fdf
, from
, &st
, fdt
, to
, override_uid
, override_gid
, copy_flags
);
632 int copy_tree(const char *from
, const char *to
, uid_t override_uid
, gid_t override_gid
, CopyFlags copy_flags
) {
633 return copy_tree_at(AT_FDCWD
, from
, AT_FDCWD
, to
, override_uid
, override_gid
, copy_flags
);
636 int copy_directory_fd(int dirfd
, const char *to
, CopyFlags copy_flags
) {
642 if (fstat(dirfd
, &st
) < 0)
645 if (!S_ISDIR(st
.st_mode
))
648 return fd_copy_directory(dirfd
, NULL
, &st
, AT_FDCWD
, to
, st
.st_dev
, COPY_DEPTH_MAX
, UID_INVALID
, GID_INVALID
, copy_flags
);
651 int copy_directory(const char *from
, const char *to
, CopyFlags copy_flags
) {
657 if (lstat(from
, &st
) < 0)
660 if (!S_ISDIR(st
.st_mode
))
663 return fd_copy_directory(AT_FDCWD
, from
, &st
, AT_FDCWD
, to
, st
.st_dev
, COPY_DEPTH_MAX
, UID_INVALID
, GID_INVALID
, copy_flags
);
666 int copy_file_fd(const char *from
, int fdt
, CopyFlags copy_flags
) {
667 _cleanup_close_
int fdf
= -1;
673 fdf
= open(from
, O_RDONLY
|O_CLOEXEC
|O_NOCTTY
);
677 r
= copy_bytes(fdf
, fdt
, (uint64_t) -1, copy_flags
);
679 (void) copy_times(fdf
, fdt
);
680 (void) copy_xattr(fdf
, fdt
);
685 int copy_file(const char *from
, const char *to
, int flags
, mode_t mode
, unsigned chattr_flags
, CopyFlags copy_flags
) {
691 RUN_WITH_UMASK(0000) {
692 fdt
= open(to
, flags
|O_WRONLY
|O_CREAT
|O_CLOEXEC
|O_NOCTTY
, mode
);
697 if (chattr_flags
!= 0)
698 (void) chattr_fd(fdt
, chattr_flags
, (unsigned) -1, NULL
);
700 r
= copy_file_fd(from
, fdt
, copy_flags
);
707 if (close(fdt
) < 0) {
715 int copy_file_atomic(const char *from
, const char *to
, mode_t mode
, unsigned chattr_flags
, CopyFlags copy_flags
) {
716 _cleanup_(unlink_and_freep
) char *t
= NULL
;
717 _cleanup_close_
int fdt
= -1;
723 /* We try to use O_TMPFILE here to create the file if we can. Note that that only works if COPY_REPLACE is not
724 * set though as we need to use linkat() for linking the O_TMPFILE file into the file system but that system
725 * call can't replace existing files. Hence, if COPY_REPLACE is set we create a temporary name in the file
726 * system right-away and unconditionally which we then can renameat() to the right name after we completed
729 if (copy_flags
& COPY_REPLACE
) {
730 r
= tempfn_random(to
, NULL
, &t
);
734 fdt
= open(t
, O_CREAT
|O_EXCL
|O_NOFOLLOW
|O_NOCTTY
|O_WRONLY
|O_CLOEXEC
, 0600);
740 fdt
= open_tmpfile_linkable(to
, O_WRONLY
|O_CLOEXEC
, &t
);
745 if (chattr_flags
!= 0)
746 (void) chattr_fd(fdt
, chattr_flags
, (unsigned) -1, NULL
);
748 r
= copy_file_fd(from
, fdt
, copy_flags
);
752 if (fchmod(fdt
, mode
) < 0)
755 if (copy_flags
& COPY_REPLACE
) {
756 if (renameat(AT_FDCWD
, t
, AT_FDCWD
, to
) < 0)
759 r
= link_tmpfile(fdt
, t
, to
);
768 int copy_times(int fdf
, int fdt
) {
769 struct timespec ut
[2];
776 if (fstat(fdf
, &st
) < 0)
782 if (futimens(fdt
, ut
) < 0)
785 if (fd_getcrtime(fdf
, &crtime
) >= 0)
786 (void) fd_setcrtime(fdt
, crtime
);
791 int copy_xattr(int fdf
, int fdt
) {
792 _cleanup_free_
char *bufa
= NULL
, *bufb
= NULL
;
793 size_t sza
= 100, szb
= 100;
803 n
= flistxattr(fdf
, bufa
, sza
);
821 assert(l
< (size_t) n
);
823 if (startswith(p
, "user.")) {
832 m
= fgetxattr(fdf
, p
, bufb
, szb
);
834 if (errno
== ERANGE
) {
843 if (fsetxattr(fdt
, p
, bufb
, m
, 0) < 0)