]>
git.ipfire.org Git - thirdparty/systemd.git/blob - src/basic/fs-util.c
1 /* SPDX-License-Identifier: LGPL-2.1+ */
9 #include <linux/magic.h>
13 #include "alloc-util.h"
14 #include "dirent-util.h"
17 #include "locale-util.h"
22 #include "parse-util.h"
23 #include "path-util.h"
24 #include "process-util.h"
25 #include "stat-util.h"
26 #include "stdio-util.h"
27 #include "string-util.h"
29 #include "time-util.h"
30 #include "tmpfile-util.h"
31 #include "user-util.h"
34 int unlink_noerrno ( const char * path
) {
45 int rmdir_parents ( const char * path
, const char * stop
) {
54 /* Skip trailing slashes */
55 while ( l
> 0 && path
[ l
- 1 ] == '/' )
61 /* Skip last component */
62 while ( l
> 0 && path
[ l
- 1 ] != '/' )
65 /* Skip trailing slashes */
66 while ( l
> 0 && path
[ l
- 1 ] == '/' )
76 if ( path_startswith ( stop
, t
)) {
92 int rename_noreplace ( int olddirfd
, const char * oldpath
, int newdirfd
, const char * newpath
) {
95 /* Try the ideal approach first */
96 if ( renameat2 ( olddirfd
, oldpath
, newdirfd
, newpath
, RENAME_NOREPLACE
) >= 0 )
99 /* renameat2() exists since Linux 3.15, btrfs and FAT added support for it later. If it is not implemented,
100 * fall back to a different method. */
101 if (! IN_SET ( errno
, EINVAL
, ENOSYS
, ENOTTY
))
104 /* Let's try to use linkat()+unlinkat() as fallback. This doesn't work on directories and on some file systems
105 * that do not support hard links (such as FAT, most prominently), but for files it's pretty close to what we
106 * want â though not atomic (i.e. for a short period both the new and the old filename will exist). */
107 if ( linkat ( olddirfd
, oldpath
, newdirfd
, newpath
, 0 ) >= 0 ) {
109 if ( unlinkat ( olddirfd
, oldpath
, 0 ) < 0 ) {
110 r
= - errno
; /* Backup errno before the following unlinkat() alters it */
111 ( void ) unlinkat ( newdirfd
, newpath
, 0 );
118 if (! IN_SET ( errno
, EINVAL
, ENOSYS
, ENOTTY
, EPERM
)) /* FAT returns EPERM on link()âĻ */
121 /* OK, neither RENAME_NOREPLACE nor linkat()+unlinkat() worked. Let's then fallback to the racy TOCTOU
122 * vulnerable accessat(F_OK) check followed by classic, replacing renameat(), we have nothing better. */
124 if ( faccessat ( newdirfd
, newpath
, F_OK
, AT_SYMLINK_NOFOLLOW
) >= 0 )
129 if ( renameat ( olddirfd
, oldpath
, newdirfd
, newpath
) < 0 )
135 int readlinkat_malloc ( int fd
, const char * p
, char ** ret
) {
136 size_t l
= FILENAME_MAX
+ 1 ;
150 n
= readlinkat ( fd
, p
, c
, l
- 1 );
157 if (( size_t ) n
< l
- 1 ) {
168 int readlink_malloc ( const char * p
, char ** ret
) {
169 return readlinkat_malloc ( AT_FDCWD
, p
, ret
);
172 int readlink_value ( const char * p
, char ** ret
) {
173 _cleanup_free_
char * link
= NULL
;
177 r
= readlink_malloc ( p
, & link
);
181 value
= basename ( link
);
185 value
= strdup ( value
);
194 int readlink_and_make_absolute ( const char * p
, char ** r
) {
195 _cleanup_free_
char * target
= NULL
;
202 j
= readlink_malloc ( p
, & target
);
206 k
= file_in_same_dir ( p
, target
);
214 int chmod_and_chown ( const char * path
, mode_t mode
, uid_t uid
, gid_t gid
) {
215 char fd_path
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int ) + 1 ];
216 _cleanup_close_
int fd
= - 1 ;
219 /* Under the assumption that we are running privileged we first change the access mode and only then hand out
220 * ownership to avoid a window where access is too open. */
222 fd
= open ( path
, O_PATH
| O_CLOEXEC
| O_NOFOLLOW
); /* Let's acquire an O_PATH fd, as precaution to change mode/owner
223 * on the same file */
227 xsprintf ( fd_path
, "/proc/self/fd/%i" , fd
);
229 if ( mode
!= MODE_INVALID
) {
231 if (( mode
& S_IFMT
) != 0 ) {
234 if ( stat ( fd_path
, & st
) < 0 )
237 if (( mode
& S_IFMT
) != ( st
. st_mode
& S_IFMT
))
241 if ( chmod ( fd_path
, mode
& 07777 ) < 0 )
245 if ( uid
!= UID_INVALID
|| gid
!= GID_INVALID
)
246 if ( chown ( fd_path
, uid
, gid
) < 0 )
252 int fchmod_and_chown ( int fd
, mode_t mode
, uid_t uid
, gid_t gid
) {
253 /* Under the assumption that we are running privileged we first change the access mode and only then hand out
254 * ownership to avoid a window where access is too open. */
256 if ( mode
!= MODE_INVALID
) {
258 if (( mode
& S_IFMT
) != 0 ) {
261 if ( fstat ( fd
, & st
) < 0 )
264 if (( mode
& S_IFMT
) != ( st
. st_mode
& S_IFMT
))
268 if ( fchmod ( fd
, mode
& 0777 ) < 0 )
272 if ( uid
!= UID_INVALID
|| gid
!= GID_INVALID
)
273 if ( fchown ( fd
, uid
, gid
) < 0 )
279 int fchmod_umask ( int fd
, mode_t m
) {
284 r
= fchmod ( fd
, m
& (~ u
)) < 0 ? - errno
: 0 ;
290 int fchmod_opath ( int fd
, mode_t m
) {
291 char procfs_path
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int )];
293 /* This function operates also on fd that might have been opened with
294 * O_PATH. Indeed fchmodat() doesn't have the AT_EMPTY_PATH flag like
295 * fchownat() does. */
297 xsprintf ( procfs_path
, "/proc/self/fd/%i" , fd
);
298 if ( chmod ( procfs_path
, m
) < 0 )
304 int fd_warn_permissions ( const char * path
, int fd
) {
307 if ( fstat ( fd
, & st
) < 0 )
310 if ( st
. st_mode
& 0111 )
311 log_warning ( "Configuration file %s is marked executable. Please remove executable permission bits. Proceeding anyway." , path
);
313 if ( st
. st_mode
& 0002 )
314 log_warning ( "Configuration file %s is marked world-writable. Please remove world writability permission bits. Proceeding anyway." , path
);
316 if ( getpid_cached () == 1 && ( st
. st_mode
& 0044 ) != 0044 )
317 log_warning ( "Configuration file %s is marked world-inaccessible. This has no effect as configuration data is accessible via APIs without restrictions. Proceeding anyway." , path
);
322 int touch_file ( const char * path
, bool parents
, usec_t stamp
, uid_t uid
, gid_t gid
, mode_t mode
) {
323 char fdpath
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int )];
324 _cleanup_close_
int fd
= - 1 ;
329 /* Note that touch_file() does not follow symlinks: if invoked on an existing symlink, then it is the symlink
330 * itself which is updated, not its target
332 * Returns the first error we encounter, but tries to apply as much as possible. */
335 ( void ) mkdir_parents ( path
, 0755 );
337 /* Initially, we try to open the node with O_PATH, so that we get a reference to the node. This is useful in
338 * case the path refers to an existing device or socket node, as we can open it successfully in all cases, and
339 * won't trigger any driver magic or so. */
340 fd
= open ( path
, O_PATH
| O_CLOEXEC
| O_NOFOLLOW
);
345 /* if the node doesn't exist yet, we create it, but with O_EXCL, so that we only create a regular file
346 * here, and nothing else */
347 fd
= open ( path
, O_WRONLY
| O_CREAT
| O_EXCL
| O_CLOEXEC
, IN_SET ( mode
, 0 , MODE_INVALID
) ? 0644 : mode
);
352 /* Let's make a path from the fd, and operate on that. With this logic, we can adjust the access mode,
353 * ownership and time of the file node in all cases, even if the fd refers to an O_PATH object â which is
354 * something fchown(), fchmod(), futimensat() don't allow. */
355 xsprintf ( fdpath
, "/proc/self/fd/%i" , fd
);
357 if ( mode
!= MODE_INVALID
)
358 if ( chmod ( fdpath
, mode
) < 0 )
361 if ( uid_is_valid ( uid
) || gid_is_valid ( gid
))
362 if ( chown ( fdpath
, uid
, gid
) < 0 && ret
>= 0 )
365 if ( stamp
!= USEC_INFINITY
) {
366 struct timespec ts
[ 2 ];
368 timespec_store (& ts
[ 0 ], stamp
);
370 r
= utimensat ( AT_FDCWD
, fdpath
, ts
, 0 );
372 r
= utimensat ( AT_FDCWD
, fdpath
, NULL
, 0 );
373 if ( r
< 0 && ret
>= 0 )
379 int touch ( const char * path
) {
380 return touch_file ( path
, false , USEC_INFINITY
, UID_INVALID
, GID_INVALID
, MODE_INVALID
);
383 int symlink_idempotent ( const char * from
, const char * to
, bool make_relative
) {
384 _cleanup_free_
char * relpath
= NULL
;
391 _cleanup_free_
char * parent
= NULL
;
393 parent
= dirname_malloc ( to
);
397 r
= path_make_relative ( parent
, from
, & relpath
);
404 if ( symlink ( from
, to
) < 0 ) {
405 _cleanup_free_
char * p
= NULL
;
410 r
= readlink_malloc ( to
, & p
);
411 if ( r
== - EINVAL
) /* Not a symlink? In that case return the original error we encountered: -EEXIST */
413 if ( r
< 0 ) /* Any other error? In that case propagate it as is */
416 if (! streq ( p
, from
)) /* Not the symlink we want it to be? In that case, propagate the original -EEXIST */
423 int symlink_atomic ( const char * from
, const char * to
) {
424 _cleanup_free_
char * t
= NULL
;
430 r
= tempfn_random ( to
, NULL
, & t
);
434 if ( symlink ( from
, t
) < 0 )
437 if ( rename ( t
, to
) < 0 ) {
445 int mknod_atomic ( const char * path
, mode_t mode
, dev_t dev
) {
446 _cleanup_free_
char * t
= NULL
;
451 r
= tempfn_random ( path
, NULL
, & t
);
455 if ( mknod ( t
, mode
, dev
) < 0 )
458 if ( rename ( t
, path
) < 0 ) {
466 int mkfifo_atomic ( const char * path
, mode_t mode
) {
467 _cleanup_free_
char * t
= NULL
;
472 r
= tempfn_random ( path
, NULL
, & t
);
476 if ( mkfifo ( t
, mode
) < 0 )
479 if ( rename ( t
, path
) < 0 ) {
487 int mkfifoat_atomic ( int dirfd
, const char * path
, mode_t mode
) {
488 _cleanup_free_
char * t
= NULL
;
493 if ( path_is_absolute ( path
))
494 return mkfifo_atomic ( path
, mode
);
496 /* We're only interested in the (random) filename. */
497 r
= tempfn_random_child ( "" , NULL
, & t
);
501 if ( mkfifoat ( dirfd
, t
, mode
) < 0 )
504 if ( renameat ( dirfd
, t
, dirfd
, path
) < 0 ) {
512 int get_files_in_directory ( const char * path
, char *** list
) {
513 _cleanup_closedir_
DIR * d
= NULL
;
515 size_t bufsize
= 0 , n
= 0 ;
516 _cleanup_strv_free_
char ** l
= NULL
;
520 /* Returns all files in a directory in *list, and the number
521 * of files as return value. If list is NULL returns only the
528 FOREACH_DIRENT_ALL ( de
, d
, return - errno
) {
529 dirent_ensure_type ( d
, de
);
531 if (! dirent_is_file ( de
))
535 /* one extra slot is needed for the terminating NULL */
536 if (! GREEDY_REALLOC ( l
, bufsize
, n
+ 2 ))
539 l
[ n
] = strdup ( de
-> d_name
);
554 static int getenv_tmp_dir ( const char ** ret_path
) {
560 /* We use the same order of environment variables python uses in tempfile.gettempdir():
561 * https://docs.python.org/3/library/tempfile.html#tempfile.gettempdir */
562 FOREACH_STRING ( n
, "TMPDIR" , "TEMP" , "TMP" ) {
565 e
= secure_getenv ( n
);
568 if (! path_is_absolute ( e
)) {
572 if (! path_is_normalized ( e
)) {
589 /* Remember first error, to make this more debuggable */
601 static int tmp_dir_internal ( const char * def
, const char ** ret
) {
608 r
= getenv_tmp_dir (& e
);
614 k
= is_dir ( def
, true );
618 return r
< 0 ? r
: k
;
624 int var_tmp_dir ( const char ** ret
) {
626 /* Returns the location for "larger" temporary files, that is backed by physical storage if available, and thus
627 * even might survive a boot: /var/tmp. If $TMPDIR (or related environment variables) are set, its value is
628 * returned preferably however. Note that both this function and tmp_dir() below are affected by $TMPDIR,
629 * making it a variable that overrides all temporary file storage locations. */
631 return tmp_dir_internal ( "/var/tmp" , ret
);
634 int tmp_dir ( const char ** ret
) {
636 /* Similar to var_tmp_dir() above, but returns the location for "smaller" temporary files, which is usually
637 * backed by an in-memory file system: /tmp. */
639 return tmp_dir_internal ( "/tmp" , ret
);
642 int unlink_or_warn ( const char * filename
) {
643 if ( unlink ( filename
) < 0 && errno
!= ENOENT
)
644 /* If the file doesn't exist and the fs simply was read-only (in which
645 * case unlink() returns EROFS even if the file doesn't exist), don't
647 if ( errno
!= EROFS
|| access ( filename
, F_OK
) >= 0 )
648 return log_error_errno ( errno
, "Failed to remove \" %s \" : %m" , filename
);
653 int inotify_add_watch_fd ( int fd
, int what
, uint32_t mask
) {
654 char path
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int ) + 1 ];
657 /* This is like inotify_add_watch(), except that the file to watch is not referenced by a path, but by an fd */
658 xsprintf ( path
, "/proc/self/fd/%i" , what
);
660 r
= inotify_add_watch ( fd
, path
, mask
);
667 static bool unsafe_transition ( const struct stat
* a
, const struct stat
* b
) {
668 /* Returns true if the transition from a to b is safe, i.e. that we never transition from unprivileged to
669 * privileged files or directories. Why bother? So that unprivileged code can't symlink to privileged files
670 * making us believe we read something safe even though it isn't safe in the specific context we open it in. */
672 if ( a
-> st_uid
== 0 ) /* Transitioning from privileged to unprivileged is always fine */
675 return a
-> st_uid
!= b
-> st_uid
; /* Otherwise we need to stay within the same UID */
678 static int log_unsafe_transition ( int a
, int b
, const char * path
, unsigned flags
) {
679 _cleanup_free_
char * n1
= NULL
, * n2
= NULL
;
681 if (! FLAGS_SET ( flags
, CHASE_WARN
))
684 ( void ) fd_get_path ( a
, & n1
);
685 ( void ) fd_get_path ( b
, & n2
);
687 return log_warning_errno ( SYNTHETIC_ERRNO ( ENOLINK
),
688 "Detected unsafe path transition %s %s %s during canonicalization of %s." ,
689 n1
, special_glyph ( SPECIAL_GLYPH_ARROW
), n2
, path
);
692 static int log_autofs_mount_point ( int fd
, const char * path
, unsigned flags
) {
693 _cleanup_free_
char * n1
= NULL
;
695 if (! FLAGS_SET ( flags
, CHASE_WARN
))
698 ( void ) fd_get_path ( fd
, & n1
);
700 return log_warning_errno ( SYNTHETIC_ERRNO ( EREMOTE
),
701 "Detected autofs mount point %s during canonicalization of %s." ,
705 int chase_symlinks ( const char * path
, const char * original_root
, unsigned flags
, char ** ret
) {
706 _cleanup_free_
char * buffer
= NULL
, * done
= NULL
, * root
= NULL
;
707 _cleanup_close_
int fd
= - 1 ;
708 unsigned max_follow
= CHASE_SYMLINKS_MAX
; /* how many symlinks to follow before giving up and returning ELOOP */
709 struct stat previous_stat
;
716 /* Either the file may be missing, or we return an fd to the final object, but both make no sense */
717 if ( FLAGS_SET ( flags
, CHASE_NONEXISTENT
| CHASE_OPEN
))
720 if ( FLAGS_SET ( flags
, CHASE_STEP
| CHASE_OPEN
))
726 /* This is a lot like canonicalize_file_name(), but takes an additional "root" parameter, that allows following
727 * symlinks relative to a root directory, instead of the root of the host.
729 * Note that "root" primarily matters if we encounter an absolute symlink. It is also used when following
730 * relative symlinks to ensure they cannot be used to "escape" the root directory. The path parameter passed is
731 * assumed to be already prefixed by it, except if the CHASE_PREFIX_ROOT flag is set, in which case it is first
732 * prefixed accordingly.
734 * Algorithmically this operates on two path buffers: "done" are the components of the path we already
735 * processed and resolved symlinks, "." and ".." of. "todo" are the components of the path we still need to
736 * process. On each iteration, we move one component from "todo" to "done", processing it's special meaning
737 * each time. The "todo" path always starts with at least one slash, the "done" path always ends in no
738 * slash. We always keep an O_PATH fd to the component we are currently processing, thus keeping lookup races
741 * Suggested usage: whenever you want to canonicalize a path, use this function. Pass the absolute path you got
742 * as-is: fully qualified and relative to your host's root. Optionally, specify the root parameter to tell this
743 * function what to do when encountering a symlink with an absolute path as directory: prefix it by the
746 * There are three ways to invoke this function:
748 * 1. Without CHASE_STEP or CHASE_OPEN: in this case the path is resolved and the normalized path is returned
749 * in `ret`. The return value is < 0 on error. If CHASE_NONEXISTENT is also set 0 is returned if the file
750 * doesn't exist, > 0 otherwise. If CHASE_NONEXISTENT is not set >= 0 is returned if the destination was
751 * found, -ENOENT if it doesn't.
753 * 2. With CHASE_OPEN: in this case the destination is opened after chasing it as O_PATH and this file
754 * descriptor is returned as return value. This is useful to open files relative to some root
755 * directory. Note that the returned O_PATH file descriptors must be converted into a regular one (using
756 * fd_reopen() or such) before it can be used for reading/writing. CHASE_OPEN may not be combined with
759 * 3. With CHASE_STEP: in this case only a single step of the normalization is executed, i.e. only the first
760 * symlink or ".." component of the path is resolved, and the resulting path is returned. This is useful if
761 * a caller wants to trace the a path through the file system verbosely. Returns < 0 on error, > 0 if the
762 * path is fully normalized, and == 0 for each normalization step. This may be combined with
763 * CHASE_NONEXISTENT, in which case 1 is returned when a component is not found.
765 * 4. With CHASE_SAFE: in this case the path must not contain unsafe transitions, i.e. transitions from
766 * unprivileged to privileged files or directories. In such cases the return value is -ENOLINK. If
767 * CHASE_WARN is also set a warning describing the unsafe transition is emitted.
769 * 5. With CHASE_NO_AUTOFS: in this case if an autofs mount point is encountered, the path normalization is
770 * aborted and -EREMOTE is returned. If CHASE_WARN is also set a warning showing the path of the mount point
775 /* A root directory of "/" or "" is identical to none */
776 if ( empty_or_root ( original_root
))
777 original_root
= NULL
;
779 if (! original_root
&& ! ret
&& ( flags
& ( CHASE_NONEXISTENT
| CHASE_NO_AUTOFS
| CHASE_SAFE
| CHASE_OPEN
| CHASE_STEP
)) == CHASE_OPEN
) {
780 /* Shortcut the CHASE_OPEN case if the caller isn't interested in the actual path and has no root set
781 * and doesn't care about any of the other special features we provide either. */
782 r
= open ( path
, O_PATH
| O_CLOEXEC
|(( flags
& CHASE_NOFOLLOW
) ? O_NOFOLLOW
: 0 ));
790 r
= path_make_absolute_cwd ( original_root
, & root
);
794 if ( flags
& CHASE_PREFIX_ROOT
) {
796 /* We don't support relative paths in combination with a root directory */
797 if (! path_is_absolute ( path
))
800 path
= prefix_roota ( root
, path
);
804 r
= path_make_absolute_cwd ( path
, & buffer
);
808 fd
= open ( "/" , O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
812 if ( flags
& CHASE_SAFE
) {
813 if ( fstat ( fd
, & previous_stat
) < 0 )
819 _cleanup_free_
char * first
= NULL
;
820 _cleanup_close_
int child
= - 1 ;
824 /* Determine length of first component in the path */
825 n
= strspn ( todo
, "/" ); /* The slashes */
826 m
= n
+ strcspn ( todo
+ n
, "/" ); /* The entire length of the component */
828 /* Extract the first component. */
829 first
= strndup ( todo
, m
);
835 /* Empty? Then we reached the end. */
839 /* Just a single slash? Then we reached the end. */
840 if ( path_equal ( first
, "/" )) {
841 /* Preserve the trailing slash */
843 if ( flags
& CHASE_TRAIL_SLASH
)
844 if (! strextend (& done
, "/" , NULL
))
850 /* Just a dot? Then let's eat this up. */
851 if ( path_equal ( first
, "/." ))
854 /* Two dots? Then chop off the last bit of what we already found out. */
855 if ( path_equal ( first
, "/.." )) {
856 _cleanup_free_
char * parent
= NULL
;
857 _cleanup_close_
int fd_parent
= - 1 ;
859 /* If we already are at the top, then going up will not change anything. This is in-line with
860 * how the kernel handles this. */
861 if ( empty_or_root ( done
))
864 parent
= dirname_malloc ( done
);
868 /* Don't allow this to leave the root dir. */
870 path_startswith ( done
, root
) &&
871 ! path_startswith ( parent
, root
))
874 free_and_replace ( done
, parent
);
876 if ( flags
& CHASE_STEP
)
879 fd_parent
= openat ( fd
, ".." , O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
883 if ( flags
& CHASE_SAFE
) {
884 if ( fstat ( fd_parent
, & st
) < 0 )
887 if ( unsafe_transition (& previous_stat
, & st
))
888 return log_unsafe_transition ( fd
, fd_parent
, path
, flags
);
894 fd
= TAKE_FD ( fd_parent
);
899 /* Otherwise let's see what this is. */
900 child
= openat ( fd
, first
+ n
, O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
903 if ( errno
== ENOENT
&&
904 ( flags
& CHASE_NONEXISTENT
) &&
905 ( isempty ( todo
) || path_is_normalized ( todo
))) {
907 /* If CHASE_NONEXISTENT is set, and the path does not exist, then that's OK, return
908 * what we got so far. But don't allow this if the remaining path contains "../ or "./"
909 * or something else weird. */
911 /* If done is "/", as first also contains slash at the head, then remove this redundant slash. */
912 if ( streq_ptr ( done
, "/" ))
915 if (! strextend (& done
, first
, todo
, NULL
))
925 if ( fstat ( child
, & st
) < 0 )
927 if (( flags
& CHASE_SAFE
) &&
928 unsafe_transition (& previous_stat
, & st
))
929 return log_unsafe_transition ( fd
, child
, path
, flags
);
933 if (( flags
& CHASE_NO_AUTOFS
) &&
934 fd_is_fs_type ( child
, AUTOFS_SUPER_MAGIC
) > 0 )
935 return log_autofs_mount_point ( child
, path
, flags
);
937 if ( S_ISLNK ( st
. st_mode
) && !(( flags
& CHASE_NOFOLLOW
) && isempty ( todo
))) {
940 _cleanup_free_
char * destination
= NULL
;
942 /* This is a symlink, in this case read the destination. But let's make sure we don't follow
943 * symlinks without bounds. */
944 if (-- max_follow
<= 0 )
947 r
= readlinkat_malloc ( fd
, first
+ n
, & destination
);
950 if ( isempty ( destination
))
953 if ( path_is_absolute ( destination
)) {
955 /* An absolute destination. Start the loop from the beginning, but use the root
956 * directory as base. */
959 fd
= open ( root
?: "/" , O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
963 if ( flags
& CHASE_SAFE
) {
964 if ( fstat ( fd
, & st
) < 0 )
967 if ( unsafe_transition (& previous_stat
, & st
))
968 return log_unsafe_transition ( child
, fd
, path
, flags
);
975 /* Note that we do not revalidate the root, we take it as is. */
984 /* Prefix what's left to do with what we just read, and start the loop again, but
985 * remain in the current directory. */
986 joined
= strjoin ( destination
, todo
);
988 joined
= strjoin ( "/" , destination
, todo
);
993 todo
= buffer
= joined
;
995 if ( flags
& CHASE_STEP
)
1001 /* If this is not a symlink, then let's just add the name we read to what we already verified. */
1003 done
= TAKE_PTR ( first
);
1005 /* If done is "/", as first also contains slash at the head, then remove this redundant slash. */
1006 if ( streq ( done
, "/" ))
1009 if (! strextend (& done
, first
, NULL
))
1013 /* And iterate again, but go one directory further down. */
1015 fd
= TAKE_FD ( child
);
1019 /* Special case, turn the empty string into "/", to indicate the root directory. */
1026 * ret
= TAKE_PTR ( done
);
1028 if ( flags
& CHASE_OPEN
) {
1029 /* Return the O_PATH fd we currently are looking to the caller. It can translate it to a proper fd by
1030 * opening /proc/self/fd/xyz. */
1036 if ( flags
& CHASE_STEP
)
1045 c
= strjoin ( strempty ( done
), todo
);
1055 int chase_symlinks_and_open (
1058 unsigned chase_flags
,
1062 _cleanup_close_
int path_fd
= - 1 ;
1063 _cleanup_free_
char * p
= NULL
;
1066 if ( chase_flags
& CHASE_NONEXISTENT
)
1069 if ( empty_or_root ( root
) && ! ret_path
&& ( chase_flags
& ( CHASE_NO_AUTOFS
| CHASE_SAFE
)) == 0 ) {
1070 /* Shortcut this call if none of the special features of this call are requested */
1071 r
= open ( path
, open_flags
);
1078 path_fd
= chase_symlinks ( path
, root
, chase_flags
| CHASE_OPEN
, ret_path
? & p
: NULL
);
1082 r
= fd_reopen ( path_fd
, open_flags
);
1087 * ret_path
= TAKE_PTR ( p
);
1092 int chase_symlinks_and_opendir (
1095 unsigned chase_flags
,
1099 char procfs_path
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int )];
1100 _cleanup_close_
int path_fd
= - 1 ;
1101 _cleanup_free_
char * p
= NULL
;
1106 if ( chase_flags
& CHASE_NONEXISTENT
)
1109 if ( empty_or_root ( root
) && ! ret_path
&& ( chase_flags
& ( CHASE_NO_AUTOFS
| CHASE_SAFE
)) == 0 ) {
1110 /* Shortcut this call if none of the special features of this call are requested */
1119 path_fd
= chase_symlinks ( path
, root
, chase_flags
| CHASE_OPEN
, ret_path
? & p
: NULL
);
1123 xsprintf ( procfs_path
, "/proc/self/fd/%i" , path_fd
);
1124 d
= opendir ( procfs_path
);
1129 * ret_path
= TAKE_PTR ( p
);
1135 int chase_symlinks_and_stat (
1138 unsigned chase_flags
,
1140 struct stat
* ret_stat
) {
1142 _cleanup_close_
int path_fd
= - 1 ;
1143 _cleanup_free_
char * p
= NULL
;
1148 if ( chase_flags
& CHASE_NONEXISTENT
)
1151 if ( empty_or_root ( root
) && ! ret_path
&& ( chase_flags
& ( CHASE_NO_AUTOFS
| CHASE_SAFE
)) == 0 ) {
1152 /* Shortcut this call if none of the special features of this call are requested */
1153 if ( stat ( path
, ret_stat
) < 0 )
1159 path_fd
= chase_symlinks ( path
, root
, chase_flags
| CHASE_OPEN
, ret_path
? & p
: NULL
);
1163 if ( fstat ( path_fd
, ret_stat
) < 0 )
1167 * ret_path
= TAKE_PTR ( p
);
1169 if ( chase_flags
& CHASE_OPEN
)
1170 return TAKE_FD ( path_fd
);
1175 int access_fd ( int fd
, int mode
) {
1176 char p
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( fd
) + 1 ];
1179 /* Like access() but operates on an already open fd */
1181 xsprintf ( p
, "/proc/self/fd/%i" , fd
);
1182 r
= access ( p
, mode
);
1189 void unlink_tempfilep ( char (* p
)[]) {
1190 /* If the file is created with mkstemp(), it will (almost always)
1191 * change the suffix. Treat this as a sign that the file was
1192 * successfully created. We ignore both the rare case where the
1193 * original suffix is used and unlink failures. */
1194 if (! endswith (* p
, ".XXXXXX" ))
1195 ( void ) unlink_noerrno (* p
);
1198 int unlinkat_deallocate ( int fd
, const char * name
, int flags
) {
1199 _cleanup_close_
int truncate_fd
= - 1 ;
1203 /* Operates like unlinkat() but also deallocates the file contents if it is a regular file and there's no other
1204 * link to it. This is useful to ensure that other processes that might have the file open for reading won't be
1205 * able to keep the data pinned on disk forever. This call is particular useful whenever we execute clean-up
1206 * jobs ("vacuuming"), where we want to make sure the data is really gone and the disk space released and
1207 * returned to the free pool.
1209 * Deallocation is preferably done by FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE (đ) if supported, which means
1210 * the file won't change size. That's a good thing since we shouldn't needlessly trigger SIGBUS in other
1211 * programs that have mmap()ed the file. (The assumption here is that changing file contents to all zeroes
1212 * underneath those programs is the better choice than simply triggering SIGBUS in them which truncation does.)
1213 * However if hole punching is not implemented in the kernel or file system we'll fall back to normal file
1214 * truncation (đĒ), as our goal of deallocating the data space trumps our goal of being nice to readers (đ).
1216 * Note that we attempt deallocation, but failure to succeed with that is not considered fatal, as long as the
1217 * primary job â to delete the file â is accomplished. */
1219 if (( flags
& AT_REMOVEDIR
) == 0 ) {
1220 truncate_fd
= openat ( fd
, name
, O_WRONLY
| O_CLOEXEC
| O_NOCTTY
| O_NOFOLLOW
| O_NONBLOCK
);
1221 if ( truncate_fd
< 0 ) {
1223 /* If this failed because the file doesn't exist propagate the error right-away. Also,
1224 * AT_REMOVEDIR wasn't set, and we tried to open the file for writing, which means EISDIR is
1225 * returned when this is a directory but we are not supposed to delete those, hence propagate
1226 * the error right-away too. */
1227 if ( IN_SET ( errno
, ENOENT
, EISDIR
))
1230 if ( errno
!= ELOOP
) /* don't complain if this is a symlink */
1231 log_debug_errno ( errno
, "Failed to open file '%s' for deallocation, ignoring: %m" , name
);
1235 if ( unlinkat ( fd
, name
, flags
) < 0 )
1238 if ( truncate_fd
< 0 ) /* Don't have a file handle, can't do more âšī¸ */
1241 if ( fstat ( truncate_fd
, & st
) < 0 ) {
1242 log_debug_errno ( errno
, "Failed to stat file '%s' for deallocation, ignoring: %m" , name
);
1246 if (! S_ISREG ( st
. st_mode
) || st
. st_blocks
== 0 || st
. st_nlink
> 0 )
1249 /* If this is a regular file, it actually took up space on disk and there are no other links it's time to
1250 * punch-hole/truncate this to release the disk space. */
1252 bs
= MAX ( st
. st_blksize
, 512 );
1253 l
= DIV_ROUND_UP ( st
. st_size
, bs
) * bs
; /* Round up to next block size */
1255 if ( fallocate ( truncate_fd
, FALLOC_FL_PUNCH_HOLE
| FALLOC_FL_KEEP_SIZE
, 0 , l
) >= 0 )
1256 return 0 ; /* Successfully punched a hole! đ */
1258 /* Fall back to truncation */
1259 if ( ftruncate ( truncate_fd
, 0 ) < 0 ) {
1260 log_debug_errno ( errno
, "Failed to truncate file to 0, ignoring: %m" );
1267 int fsync_directory_of_file ( int fd
) {
1268 _cleanup_free_
char * path
= NULL
;
1269 _cleanup_close_
int dfd
= - 1 ;
1272 r
= fd_verify_regular ( fd
);
1276 r
= fd_get_path ( fd
, & path
);
1278 log_debug_errno ( r
, "Failed to query /proc/self/fd/%d%s: %m" ,
1280 r
== - EOPNOTSUPP
? ", ignoring" : "" );
1282 if ( r
== - EOPNOTSUPP
)
1283 /* If /proc is not available, we're most likely running in some
1284 * chroot environment, and syncing the directory is not very
1285 * important in that case. Let's just silently do nothing. */
1291 if (! path_is_absolute ( path
))
1294 dfd
= open_parent ( path
, O_CLOEXEC
, 0 );
1304 int fsync_path_at ( int at_fd
, const char * path
) {
1305 _cleanup_close_
int opened_fd
= - 1 ;
1308 if ( isempty ( path
)) {
1309 if ( at_fd
== AT_FDCWD
) {
1310 opened_fd
= open ( "." , O_RDONLY
| O_DIRECTORY
| O_CLOEXEC
);
1319 opened_fd
= openat ( at_fd
, path
, O_RDONLY
| O_CLOEXEC
);
1332 int open_parent ( const char * path
, int flags
, mode_t mode
) {
1333 _cleanup_free_
char * parent
= NULL
;
1338 if ( path_equal ( path
, "/" )) /* requesting the parent of the root dir is fishy, let's prohibit that */
1341 parent
= dirname_malloc ( path
);
1345 /* Let's insist on O_DIRECTORY since the parent of a file or directory is a directory. Except if we open an
1346 * O_TMPFILE file, because in that case we are actually create a regular file below the parent directory. */
1348 if (( flags
& O_PATH
) == O_PATH
)
1349 flags
|= O_DIRECTORY
;
1350 else if (( flags
& O_TMPFILE
) != O_TMPFILE
)
1351 flags
|= O_DIRECTORY
| O_RDONLY
;
1353 fd
= open ( parent
, flags
, mode
);