]>
git.ipfire.org Git - thirdparty/systemd.git/blob - src/basic/fs-util.c
1 /* SPDX-License-Identifier: LGPL-2.1+ */
8 #include <linux/falloc.h>
9 #include <linux/magic.h>
12 #include "alloc-util.h"
13 #include "dirent-util.h"
16 #include "locale-util.h"
19 #include "missing_fs.h"
20 #include "missing_syscall.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 _cleanup_close_
int fd
= - 1 ;
219 fd
= open ( path
, O_PATH
| O_CLOEXEC
| O_NOFOLLOW
); /* Let's acquire an O_PATH fd, as precaution to change
220 * mode/owner on the same file */
224 return fchmod_and_chown ( fd
, mode
, uid
, gid
);
227 int fchmod_and_chown ( int fd
, mode_t mode
, uid_t uid
, gid_t gid
) {
228 bool do_chown
, do_chmod
;
231 /* Change ownership and access mode of the specified fd. Tries to do so safely, ensuring that at no
232 * point in time the access mode is above the old access mode under the old ownership or the new
233 * access mode under the new ownership. Note: this call tries hard to leave the access mode
234 * unaffected if the uid/gid is changed, i.e. it undoes implicit suid/sgid dropping the kernel does
237 * This call is happy with O_PATH fds. */
239 if ( fstat ( fd
, & st
) < 0 )
243 ( uid
!= UID_INVALID
&& st
. st_uid
!= uid
) ||
244 ( gid
!= GID_INVALID
&& st
. st_gid
!= gid
);
247 ! S_ISLNK ( st
. st_mode
) && /* chmod is not defined on symlinks */
248 (( mode
!= MODE_INVALID
&& (( st
. st_mode
^ mode
) & 07777 ) != 0 ) ||
249 do_chown
); /* If we change ownership, make sure we reset the mode afterwards, since chown()
250 * modifies the access mode too */
252 if ( mode
== MODE_INVALID
)
253 mode
= st
. st_mode
; /* If we only shall do a chown(), save original mode, since chown() might break it. */
254 else if (( mode
& S_IFMT
) != 0 && (( mode
^ st
. st_mode
) & S_IFMT
) != 0 )
255 return - EINVAL
; /* insist on the right file type if it was specified */
257 if ( do_chown
&& do_chmod
) {
258 mode_t minimal
= st
. st_mode
& mode
; /* the subset of the old and the new mask */
260 if ((( minimal
^ st
. st_mode
) & 07777 ) != 0 )
261 if ( fchmod_opath ( fd
, minimal
& 07777 ) < 0 )
266 if ( fchownat ( fd
, "" , uid
, gid
, AT_EMPTY_PATH
) < 0 )
270 if ( fchmod_opath ( fd
, mode
& 07777 ) < 0 )
273 return do_chown
|| do_chmod
;
276 int fchmod_umask ( int fd
, mode_t m
) {
281 r
= fchmod ( fd
, m
& (~ u
)) < 0 ? - errno
: 0 ;
287 int fchmod_opath ( int fd
, mode_t m
) {
288 char procfs_path
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int )];
290 /* This function operates also on fd that might have been opened with
291 * O_PATH. Indeed fchmodat() doesn't have the AT_EMPTY_PATH flag like
292 * fchownat() does. */
294 xsprintf ( procfs_path
, "/proc/self/fd/%i" , fd
);
295 if ( chmod ( procfs_path
, m
) < 0 )
301 int fd_warn_permissions ( const char * path
, int fd
) {
304 if ( fstat ( fd
, & st
) < 0 )
307 /* Don't complain if we are reading something that is not a file, for example /dev/null */
308 if (! S_ISREG ( st
. st_mode
))
311 if ( st
. st_mode
& 0111 )
312 log_warning ( "Configuration file %s is marked executable. Please remove executable permission bits. Proceeding anyway." , path
);
314 if ( st
. st_mode
& 0002 )
315 log_warning ( "Configuration file %s is marked world-writable. Please remove world writability permission bits. Proceeding anyway." , path
);
317 if ( getpid_cached () == 1 && ( st
. st_mode
& 0044 ) != 0044 )
318 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
);
323 int touch_file ( const char * path
, bool parents
, usec_t stamp
, uid_t uid
, gid_t gid
, mode_t mode
) {
324 char fdpath
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int )];
325 _cleanup_close_
int fd
= - 1 ;
330 /* Note that touch_file() does not follow symlinks: if invoked on an existing symlink, then it is the symlink
331 * itself which is updated, not its target
333 * Returns the first error we encounter, but tries to apply as much as possible. */
336 ( void ) mkdir_parents ( path
, 0755 );
338 /* Initially, we try to open the node with O_PATH, so that we get a reference to the node. This is useful in
339 * case the path refers to an existing device or socket node, as we can open it successfully in all cases, and
340 * won't trigger any driver magic or so. */
341 fd
= open ( path
, O_PATH
| O_CLOEXEC
| O_NOFOLLOW
);
346 /* if the node doesn't exist yet, we create it, but with O_EXCL, so that we only create a regular file
347 * here, and nothing else */
348 fd
= open ( path
, O_WRONLY
| O_CREAT
| O_EXCL
| O_CLOEXEC
, IN_SET ( mode
, 0 , MODE_INVALID
) ? 0644 : mode
);
353 /* Let's make a path from the fd, and operate on that. With this logic, we can adjust the access mode,
354 * ownership and time of the file node in all cases, even if the fd refers to an O_PATH object â which is
355 * something fchown(), fchmod(), futimensat() don't allow. */
356 xsprintf ( fdpath
, "/proc/self/fd/%i" , fd
);
358 ret
= fchmod_and_chown ( fd
, mode
, uid
, gid
);
360 if ( stamp
!= USEC_INFINITY
) {
361 struct timespec ts
[ 2 ];
363 timespec_store (& ts
[ 0 ], stamp
);
365 r
= utimensat ( AT_FDCWD
, fdpath
, ts
, 0 );
367 r
= utimensat ( AT_FDCWD
, fdpath
, NULL
, 0 );
368 if ( r
< 0 && ret
>= 0 )
374 int touch ( const char * path
) {
375 return touch_file ( path
, false , USEC_INFINITY
, UID_INVALID
, GID_INVALID
, MODE_INVALID
);
378 int symlink_idempotent ( const char * from
, const char * to
, bool make_relative
) {
379 _cleanup_free_
char * relpath
= NULL
;
386 _cleanup_free_
char * parent
= NULL
;
388 parent
= dirname_malloc ( to
);
392 r
= path_make_relative ( parent
, from
, & relpath
);
399 if ( symlink ( from
, to
) < 0 ) {
400 _cleanup_free_
char * p
= NULL
;
405 r
= readlink_malloc ( to
, & p
);
406 if ( r
== - EINVAL
) /* Not a symlink? In that case return the original error we encountered: -EEXIST */
408 if ( r
< 0 ) /* Any other error? In that case propagate it as is */
411 if (! streq ( p
, from
)) /* Not the symlink we want it to be? In that case, propagate the original -EEXIST */
418 int symlink_atomic ( const char * from
, const char * to
) {
419 _cleanup_free_
char * t
= NULL
;
425 r
= tempfn_random ( to
, NULL
, & t
);
429 if ( symlink ( from
, t
) < 0 )
432 if ( rename ( t
, to
) < 0 ) {
440 int mknod_atomic ( const char * path
, mode_t mode
, dev_t dev
) {
441 _cleanup_free_
char * t
= NULL
;
446 r
= tempfn_random ( path
, NULL
, & t
);
450 if ( mknod ( t
, mode
, dev
) < 0 )
453 if ( rename ( t
, path
) < 0 ) {
461 int mkfifo_atomic ( const char * path
, mode_t mode
) {
462 _cleanup_free_
char * t
= NULL
;
467 r
= tempfn_random ( path
, NULL
, & t
);
471 if ( mkfifo ( t
, mode
) < 0 )
474 if ( rename ( t
, path
) < 0 ) {
482 int mkfifoat_atomic ( int dirfd
, const char * path
, mode_t mode
) {
483 _cleanup_free_
char * t
= NULL
;
488 if ( path_is_absolute ( path
))
489 return mkfifo_atomic ( path
, mode
);
491 /* We're only interested in the (random) filename. */
492 r
= tempfn_random_child ( "" , NULL
, & t
);
496 if ( mkfifoat ( dirfd
, t
, mode
) < 0 )
499 if ( renameat ( dirfd
, t
, dirfd
, path
) < 0 ) {
507 int get_files_in_directory ( const char * path
, char *** list
) {
508 _cleanup_closedir_
DIR * d
= NULL
;
510 size_t bufsize
= 0 , n
= 0 ;
511 _cleanup_strv_free_
char ** l
= NULL
;
515 /* Returns all files in a directory in *list, and the number
516 * of files as return value. If list is NULL returns only the
523 FOREACH_DIRENT_ALL ( de
, d
, return - errno
) {
524 dirent_ensure_type ( d
, de
);
526 if (! dirent_is_file ( de
))
530 /* one extra slot is needed for the terminating NULL */
531 if (! GREEDY_REALLOC ( l
, bufsize
, n
+ 2 ))
534 l
[ n
] = strdup ( de
-> d_name
);
549 static int getenv_tmp_dir ( const char ** ret_path
) {
555 /* We use the same order of environment variables python uses in tempfile.gettempdir():
556 * https://docs.python.org/3/library/tempfile.html#tempfile.gettempdir */
557 FOREACH_STRING ( n
, "TMPDIR" , "TEMP" , "TMP" ) {
560 e
= secure_getenv ( n
);
563 if (! path_is_absolute ( e
)) {
567 if (! path_is_normalized ( e
)) {
584 /* Remember first error, to make this more debuggable */
596 static int tmp_dir_internal ( const char * def
, const char ** ret
) {
603 r
= getenv_tmp_dir (& e
);
609 k
= is_dir ( def
, true );
613 return r
< 0 ? r
: k
;
619 int var_tmp_dir ( const char ** ret
) {
621 /* Returns the location for "larger" temporary files, that is backed by physical storage if available, and thus
622 * even might survive a boot: /var/tmp. If $TMPDIR (or related environment variables) are set, its value is
623 * returned preferably however. Note that both this function and tmp_dir() below are affected by $TMPDIR,
624 * making it a variable that overrides all temporary file storage locations. */
626 return tmp_dir_internal ( "/var/tmp" , ret
);
629 int tmp_dir ( const char ** ret
) {
631 /* Similar to var_tmp_dir() above, but returns the location for "smaller" temporary files, which is usually
632 * backed by an in-memory file system: /tmp. */
634 return tmp_dir_internal ( "/tmp" , ret
);
637 int unlink_or_warn ( const char * filename
) {
638 if ( unlink ( filename
) < 0 && errno
!= ENOENT
)
639 /* If the file doesn't exist and the fs simply was read-only (in which
640 * case unlink() returns EROFS even if the file doesn't exist), don't
642 if ( errno
!= EROFS
|| access ( filename
, F_OK
) >= 0 )
643 return log_error_errno ( errno
, "Failed to remove \" %s \" : %m" , filename
);
648 int inotify_add_watch_fd ( int fd
, int what
, uint32_t mask
) {
649 char path
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int ) + 1 ];
652 /* This is like inotify_add_watch(), except that the file to watch is not referenced by a path, but by an fd */
653 xsprintf ( path
, "/proc/self/fd/%i" , what
);
655 r
= inotify_add_watch ( fd
, path
, mask
);
662 int inotify_add_watch_and_warn ( int fd
, const char * pathname
, uint32_t mask
) {
664 if ( inotify_add_watch ( fd
, pathname
, mask
) < 0 ) {
666 return log_error_errno ( errno
, "Failed to add a watch for %s: inotify watch limit reached" , pathname
);
668 return log_error_errno ( errno
, "Failed to add a watch for %s: %m" , pathname
);
674 static bool unsafe_transition ( const struct stat
* a
, const struct stat
* b
) {
675 /* Returns true if the transition from a to b is safe, i.e. that we never transition from unprivileged to
676 * privileged files or directories. Why bother? So that unprivileged code can't symlink to privileged files
677 * making us believe we read something safe even though it isn't safe in the specific context we open it in. */
679 if ( a
-> st_uid
== 0 ) /* Transitioning from privileged to unprivileged is always fine */
682 return a
-> st_uid
!= b
-> st_uid
; /* Otherwise we need to stay within the same UID */
685 static int log_unsafe_transition ( int a
, int b
, const char * path
, unsigned flags
) {
686 _cleanup_free_
char * n1
= NULL
, * n2
= NULL
;
688 if (! FLAGS_SET ( flags
, CHASE_WARN
))
691 ( void ) fd_get_path ( a
, & n1
);
692 ( void ) fd_get_path ( b
, & n2
);
694 return log_warning_errno ( SYNTHETIC_ERRNO ( ENOLINK
),
695 "Detected unsafe path transition %s %s %s during canonicalization of %s." ,
696 n1
, special_glyph ( SPECIAL_GLYPH_ARROW
), n2
, path
);
699 static int log_autofs_mount_point ( int fd
, const char * path
, unsigned flags
) {
700 _cleanup_free_
char * n1
= NULL
;
702 if (! FLAGS_SET ( flags
, CHASE_WARN
))
705 ( void ) fd_get_path ( fd
, & n1
);
707 return log_warning_errno ( SYNTHETIC_ERRNO ( EREMOTE
),
708 "Detected autofs mount point %s during canonicalization of %s." ,
712 int chase_symlinks ( const char * path
, const char * original_root
, unsigned flags
, char ** ret_path
, int * ret_fd
) {
713 _cleanup_free_
char * buffer
= NULL
, * done
= NULL
, * root
= NULL
;
714 _cleanup_close_
int fd
= - 1 ;
715 unsigned max_follow
= CHASE_SYMLINKS_MAX
; /* how many symlinks to follow before giving up and returning ELOOP */
716 struct stat previous_stat
;
723 /* Either the file may be missing, or we return an fd to the final object, but both make no sense */
724 if (( flags
& CHASE_NONEXISTENT
) && ret_fd
)
727 if (( flags
& CHASE_STEP
) && ret_fd
)
733 /* This is a lot like canonicalize_file_name(), but takes an additional "root" parameter, that allows following
734 * symlinks relative to a root directory, instead of the root of the host.
736 * Note that "root" primarily matters if we encounter an absolute symlink. It is also used when following
737 * relative symlinks to ensure they cannot be used to "escape" the root directory. The path parameter passed is
738 * assumed to be already prefixed by it, except if the CHASE_PREFIX_ROOT flag is set, in which case it is first
739 * prefixed accordingly.
741 * Algorithmically this operates on two path buffers: "done" are the components of the path we already
742 * processed and resolved symlinks, "." and ".." of. "todo" are the components of the path we still need to
743 * process. On each iteration, we move one component from "todo" to "done", processing it's special meaning
744 * each time. The "todo" path always starts with at least one slash, the "done" path always ends in no
745 * slash. We always keep an O_PATH fd to the component we are currently processing, thus keeping lookup races
748 * Suggested usage: whenever you want to canonicalize a path, use this function. Pass the absolute path you got
749 * as-is: fully qualified and relative to your host's root. Optionally, specify the root parameter to tell this
750 * function what to do when encountering a symlink with an absolute path as directory: prefix it by the
753 * There are five ways to invoke this function:
755 * 1. Without CHASE_STEP or ret_fd: in this case the path is resolved and the normalized path is
756 * returned in `ret_path`. The return value is < 0 on error. If CHASE_NONEXISTENT is also set, 0
757 * is returned if the file doesn't exist, > 0 otherwise. If CHASE_NONEXISTENT is not set, >= 0 is
758 * returned if the destination was found, -ENOENT if it wasn't.
760 * 2. With ret_fd: in this case the destination is opened after chasing it as O_PATH and this file
761 * descriptor is returned as return value. This is useful to open files relative to some root
762 * directory. Note that the returned O_PATH file descriptors must be converted into a regular one (using
763 * fd_reopen() or such) before it can be used for reading/writing. ret_fd may not be combined with
766 * 3. With CHASE_STEP: in this case only a single step of the normalization is executed, i.e. only the first
767 * symlink or ".." component of the path is resolved, and the resulting path is returned. This is useful if
768 * a caller wants to trace the a path through the file system verbosely. Returns < 0 on error, > 0 if the
769 * path is fully normalized, and == 0 for each normalization step. This may be combined with
770 * CHASE_NONEXISTENT, in which case 1 is returned when a component is not found.
772 * 4. With CHASE_SAFE: in this case the path must not contain unsafe transitions, i.e. transitions from
773 * unprivileged to privileged files or directories. In such cases the return value is -ENOLINK. If
774 * CHASE_WARN is also set, a warning describing the unsafe transition is emitted.
776 * 5. With CHASE_NO_AUTOFS: in this case if an autofs mount point is encountered, path normalization
777 * is aborted and -EREMOTE is returned. If CHASE_WARN is also set, a warning showing the path of
778 * the mount point is emitted.
781 /* A root directory of "/" or "" is identical to none */
782 if ( empty_or_root ( original_root
))
783 original_root
= NULL
;
785 if (! original_root
&& ! ret_path
&& !( flags
& ( CHASE_NONEXISTENT
| CHASE_NO_AUTOFS
| CHASE_SAFE
| CHASE_STEP
)) && ret_fd
) {
786 /* Shortcut the ret_fd case if the caller isn't interested in the actual path and has no root set
787 * and doesn't care about any of the other special features we provide either. */
788 r
= open ( path
, O_PATH
| O_CLOEXEC
|(( flags
& CHASE_NOFOLLOW
) ? O_NOFOLLOW
: 0 ));
797 r
= path_make_absolute_cwd ( original_root
, & root
);
801 if ( flags
& CHASE_PREFIX_ROOT
) {
802 /* We don't support relative paths in combination with a root directory */
803 if (! path_is_absolute ( path
))
806 path
= prefix_roota ( root
, path
);
810 r
= path_make_absolute_cwd ( path
, & buffer
);
814 fd
= open ( "/" , O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
818 if ( flags
& CHASE_SAFE
) {
819 if ( fstat ( fd
, & previous_stat
) < 0 )
825 _cleanup_free_
char * first
= NULL
;
826 _cleanup_close_
int child
= - 1 ;
830 /* Determine length of first component in the path */
831 n
= strspn ( todo
, "/" ); /* The slashes */
832 m
= n
+ strcspn ( todo
+ n
, "/" ); /* The entire length of the component */
834 /* Extract the first component. */
835 first
= strndup ( todo
, m
);
841 /* Empty? Then we reached the end. */
845 /* Just a single slash? Then we reached the end. */
846 if ( path_equal ( first
, "/" )) {
847 /* Preserve the trailing slash */
849 if ( flags
& CHASE_TRAIL_SLASH
)
850 if (! strextend (& done
, "/" , NULL
))
856 /* Just a dot? Then let's eat this up. */
857 if ( path_equal ( first
, "/." ))
860 /* Two dots? Then chop off the last bit of what we already found out. */
861 if ( path_equal ( first
, "/.." )) {
862 _cleanup_free_
char * parent
= NULL
;
863 _cleanup_close_
int fd_parent
= - 1 ;
865 /* If we already are at the top, then going up will not change anything. This is in-line with
866 * how the kernel handles this. */
867 if ( empty_or_root ( done
))
870 parent
= dirname_malloc ( done
);
874 /* Don't allow this to leave the root dir. */
876 path_startswith ( done
, root
) &&
877 ! path_startswith ( parent
, root
))
880 free_and_replace ( done
, parent
);
882 if ( flags
& CHASE_STEP
)
885 fd_parent
= openat ( fd
, ".." , O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
889 if ( flags
& CHASE_SAFE
) {
890 if ( fstat ( fd_parent
, & st
) < 0 )
893 if ( unsafe_transition (& previous_stat
, & st
))
894 return log_unsafe_transition ( fd
, fd_parent
, path
, flags
);
900 fd
= TAKE_FD ( fd_parent
);
905 /* Otherwise let's see what this is. */
906 child
= openat ( fd
, first
+ n
, O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
909 if ( errno
== ENOENT
&&
910 ( flags
& CHASE_NONEXISTENT
) &&
911 ( isempty ( todo
) || path_is_normalized ( todo
))) {
913 /* If CHASE_NONEXISTENT is set, and the path does not exist, then that's OK, return
914 * what we got so far. But don't allow this if the remaining path contains "../ or "./"
915 * or something else weird. */
917 /* If done is "/", as first also contains slash at the head, then remove this redundant slash. */
918 if ( streq_ptr ( done
, "/" ))
921 if (! strextend (& done
, first
, todo
, NULL
))
931 if ( fstat ( child
, & st
) < 0 )
933 if (( flags
& CHASE_SAFE
) &&
934 ( empty_or_root ( root
) || ( size_t )( todo
- buffer
) > strlen ( root
)) &&
935 unsafe_transition (& previous_stat
, & st
))
936 return log_unsafe_transition ( fd
, child
, path
, flags
);
940 if (( flags
& CHASE_NO_AUTOFS
) &&
941 fd_is_fs_type ( child
, AUTOFS_SUPER_MAGIC
) > 0 )
942 return log_autofs_mount_point ( child
, path
, flags
);
944 if ( S_ISLNK ( st
. st_mode
) && !(( flags
& CHASE_NOFOLLOW
) && isempty ( todo
))) {
946 _cleanup_free_
char * destination
= NULL
;
948 /* This is a symlink, in this case read the destination. But let's make sure we don't follow
949 * symlinks without bounds. */
950 if (-- max_follow
<= 0 )
953 r
= readlinkat_malloc ( fd
, first
+ n
, & destination
);
956 if ( isempty ( destination
))
959 if ( path_is_absolute ( destination
)) {
961 /* An absolute destination. Start the loop from the beginning, but use the root
962 * directory as base. */
965 fd
= open ( root
?: "/" , O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
969 if ( flags
& CHASE_SAFE
) {
970 if ( fstat ( fd
, & st
) < 0 )
973 if ( unsafe_transition (& previous_stat
, & st
))
974 return log_unsafe_transition ( child
, fd
, path
, flags
);
981 /* Note that we do not revalidate the root, we take it as is. */
990 /* Prefix what's left to do with what we just read, and start the loop again, but
991 * remain in the current directory. */
992 joined
= path_join ( destination
, todo
);
994 joined
= path_join ( "/" , destination
, todo
);
999 todo
= buffer
= joined
;
1001 if ( flags
& CHASE_STEP
)
1007 /* If this is not a symlink, then let's just add the name we read to what we already verified. */
1009 done
= TAKE_PTR ( first
);
1011 /* If done is "/", as first also contains slash at the head, then remove this redundant slash. */
1012 if ( streq ( done
, "/" ))
1015 if (! strextend (& done
, first
, NULL
))
1019 /* And iterate again, but go one directory further down. */
1021 fd
= TAKE_FD ( child
);
1025 /* Special case, turn the empty string into "/", to indicate the root directory. */
1032 * ret_path
= TAKE_PTR ( done
);
1035 /* Return the O_PATH fd we currently are looking to the caller. It can translate it to a
1036 * proper fd by opening /proc/self/fd/xyz. */
1039 * ret_fd
= TAKE_FD ( fd
);
1042 if ( flags
& CHASE_STEP
)
1051 c
= strjoin ( strempty ( done
), todo
);
1061 int chase_symlinks_and_open (
1064 unsigned chase_flags
,
1068 _cleanup_close_
int path_fd
= - 1 ;
1069 _cleanup_free_
char * p
= NULL
;
1072 if ( chase_flags
& CHASE_NONEXISTENT
)
1075 if ( empty_or_root ( root
) && ! ret_path
&& ( chase_flags
& ( CHASE_NO_AUTOFS
| CHASE_SAFE
)) == 0 ) {
1076 /* Shortcut this call if none of the special features of this call are requested */
1077 r
= open ( path
, open_flags
);
1084 r
= chase_symlinks ( path
, root
, chase_flags
, ret_path
? & p
: NULL
, & path_fd
);
1087 assert ( path_fd
>= 0 );
1089 r
= fd_reopen ( path_fd
, open_flags
);
1094 * ret_path
= TAKE_PTR ( p
);
1099 int chase_symlinks_and_opendir (
1102 unsigned chase_flags
,
1106 char procfs_path
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int )];
1107 _cleanup_close_
int path_fd
= - 1 ;
1108 _cleanup_free_
char * p
= NULL
;
1114 if ( chase_flags
& CHASE_NONEXISTENT
)
1117 if ( empty_or_root ( root
) && ! ret_path
&& ( chase_flags
& ( CHASE_NO_AUTOFS
| CHASE_SAFE
)) == 0 ) {
1118 /* Shortcut this call if none of the special features of this call are requested */
1127 r
= chase_symlinks ( path
, root
, chase_flags
, ret_path
? & p
: NULL
, & path_fd
);
1130 assert ( path_fd
>= 0 );
1132 xsprintf ( procfs_path
, "/proc/self/fd/%i" , path_fd
);
1133 d
= opendir ( procfs_path
);
1138 * ret_path
= TAKE_PTR ( p
);
1144 int chase_symlinks_and_stat (
1147 unsigned chase_flags
,
1149 struct stat
* ret_stat
,
1152 _cleanup_close_
int path_fd
= - 1 ;
1153 _cleanup_free_
char * p
= NULL
;
1159 if ( chase_flags
& CHASE_NONEXISTENT
)
1162 if ( empty_or_root ( root
) && ! ret_path
&& ( chase_flags
& ( CHASE_NO_AUTOFS
| CHASE_SAFE
)) == 0 ) {
1163 /* Shortcut this call if none of the special features of this call are requested */
1164 if ( stat ( path
, ret_stat
) < 0 )
1170 r
= chase_symlinks ( path
, root
, chase_flags
, ret_path
? & p
: NULL
, & path_fd
);
1173 assert ( path_fd
>= 0 );
1175 if ( fstat ( path_fd
, ret_stat
) < 0 )
1179 * ret_path
= TAKE_PTR ( p
);
1181 * ret_fd
= TAKE_FD ( path_fd
);
1186 int access_fd ( int fd
, int mode
) {
1187 char p
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( fd
) + 1 ];
1190 /* Like access() but operates on an already open fd */
1192 xsprintf ( p
, "/proc/self/fd/%i" , fd
);
1193 r
= access ( p
, mode
);
1200 void unlink_tempfilep ( char (* p
)[]) {
1201 /* If the file is created with mkstemp(), it will (almost always)
1202 * change the suffix. Treat this as a sign that the file was
1203 * successfully created. We ignore both the rare case where the
1204 * original suffix is used and unlink failures. */
1205 if (! endswith (* p
, ".XXXXXX" ))
1206 ( void ) unlink_noerrno (* p
);
1209 int unlinkat_deallocate ( int fd
, const char * name
, int flags
) {
1210 _cleanup_close_
int truncate_fd
= - 1 ;
1214 /* Operates like unlinkat() but also deallocates the file contents if it is a regular file and there's no other
1215 * link to it. This is useful to ensure that other processes that might have the file open for reading won't be
1216 * able to keep the data pinned on disk forever. This call is particular useful whenever we execute clean-up
1217 * jobs ("vacuuming"), where we want to make sure the data is really gone and the disk space released and
1218 * returned to the free pool.
1220 * Deallocation is preferably done by FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE (đ) if supported, which means
1221 * the file won't change size. That's a good thing since we shouldn't needlessly trigger SIGBUS in other
1222 * programs that have mmap()ed the file. (The assumption here is that changing file contents to all zeroes
1223 * underneath those programs is the better choice than simply triggering SIGBUS in them which truncation does.)
1224 * However if hole punching is not implemented in the kernel or file system we'll fall back to normal file
1225 * truncation (đĒ), as our goal of deallocating the data space trumps our goal of being nice to readers (đ).
1227 * Note that we attempt deallocation, but failure to succeed with that is not considered fatal, as long as the
1228 * primary job â to delete the file â is accomplished. */
1230 if (( flags
& AT_REMOVEDIR
) == 0 ) {
1231 truncate_fd
= openat ( fd
, name
, O_WRONLY
| O_CLOEXEC
| O_NOCTTY
| O_NOFOLLOW
| O_NONBLOCK
);
1232 if ( truncate_fd
< 0 ) {
1234 /* If this failed because the file doesn't exist propagate the error right-away. Also,
1235 * AT_REMOVEDIR wasn't set, and we tried to open the file for writing, which means EISDIR is
1236 * returned when this is a directory but we are not supposed to delete those, hence propagate
1237 * the error right-away too. */
1238 if ( IN_SET ( errno
, ENOENT
, EISDIR
))
1241 if ( errno
!= ELOOP
) /* don't complain if this is a symlink */
1242 log_debug_errno ( errno
, "Failed to open file '%s' for deallocation, ignoring: %m" , name
);
1246 if ( unlinkat ( fd
, name
, flags
) < 0 )
1249 if ( truncate_fd
< 0 ) /* Don't have a file handle, can't do more âšī¸ */
1252 if ( fstat ( truncate_fd
, & st
) < 0 ) {
1253 log_debug_errno ( errno
, "Failed to stat file '%s' for deallocation, ignoring: %m" , name
);
1257 if (! S_ISREG ( st
. st_mode
) || st
. st_blocks
== 0 || st
. st_nlink
> 0 )
1260 /* If this is a regular file, it actually took up space on disk and there are no other links it's time to
1261 * punch-hole/truncate this to release the disk space. */
1263 bs
= MAX ( st
. st_blksize
, 512 );
1264 l
= DIV_ROUND_UP ( st
. st_size
, bs
) * bs
; /* Round up to next block size */
1266 if ( fallocate ( truncate_fd
, FALLOC_FL_PUNCH_HOLE
| FALLOC_FL_KEEP_SIZE
, 0 , l
) >= 0 )
1267 return 0 ; /* Successfully punched a hole! đ */
1269 /* Fall back to truncation */
1270 if ( ftruncate ( truncate_fd
, 0 ) < 0 ) {
1271 log_debug_errno ( errno
, "Failed to truncate file to 0, ignoring: %m" );
1278 int fsync_directory_of_file ( int fd
) {
1279 _cleanup_free_
char * path
= NULL
;
1280 _cleanup_close_
int dfd
= - 1 ;
1283 r
= fd_verify_regular ( fd
);
1287 r
= fd_get_path ( fd
, & path
);
1289 log_debug_errno ( r
, "Failed to query /proc/self/fd/%d%s: %m" ,
1291 r
== - EOPNOTSUPP
? ", ignoring" : "" );
1293 if ( r
== - EOPNOTSUPP
)
1294 /* If /proc is not available, we're most likely running in some
1295 * chroot environment, and syncing the directory is not very
1296 * important in that case. Let's just silently do nothing. */
1302 if (! path_is_absolute ( path
))
1305 dfd
= open_parent ( path
, O_CLOEXEC
, 0 );
1315 int fsync_full ( int fd
) {
1318 /* Sync both the file and the directory */
1320 r
= fsync ( fd
) < 0 ? - errno
: 0 ;
1321 q
= fsync_directory_of_file ( fd
);
1323 return r
< 0 ? r
: q
;
1326 int fsync_path_at ( int at_fd
, const char * path
) {
1327 _cleanup_close_
int opened_fd
= - 1 ;
1330 if ( isempty ( path
)) {
1331 if ( at_fd
== AT_FDCWD
) {
1332 opened_fd
= open ( "." , O_RDONLY
| O_DIRECTORY
| O_CLOEXEC
);
1341 opened_fd
= openat ( at_fd
, path
, O_RDONLY
| O_CLOEXEC
);
1354 int syncfs_path ( int atfd
, const char * path
) {
1355 _cleanup_close_
int fd
= - 1 ;
1359 fd
= openat ( atfd
, path
, O_CLOEXEC
| O_RDONLY
| O_NONBLOCK
);
1369 int open_parent ( const char * path
, int flags
, mode_t mode
) {
1370 _cleanup_free_
char * parent
= NULL
;
1375 if ( path_equal ( path
, "/" )) /* requesting the parent of the root dir is fishy, let's prohibit that */
1378 parent
= dirname_malloc ( path
);
1382 /* Let's insist on O_DIRECTORY since the parent of a file or directory is a directory. Except if we open an
1383 * O_TMPFILE file, because in that case we are actually create a regular file below the parent directory. */
1385 if ( FLAGS_SET ( flags
, O_PATH
))
1386 flags
|= O_DIRECTORY
;
1387 else if (! FLAGS_SET ( flags
, O_TMPFILE
))
1388 flags
|= O_DIRECTORY
| O_RDONLY
;
1390 fd
= open ( parent
, flags
, mode
);