]>
git.ipfire.org Git - thirdparty/systemd.git/blob - src/basic/fs-util.c
1 /* SPDX-License-Identifier: LGPL-2.1+ */
9 #include <linux/falloc.h>
10 #include <linux/magic.h>
14 #include "alloc-util.h"
15 #include "dirent-util.h"
18 #include "locale-util.h"
23 #include "parse-util.h"
24 #include "path-util.h"
25 #include "process-util.h"
26 #include "stat-util.h"
27 #include "stdio-util.h"
28 #include "string-util.h"
30 #include "time-util.h"
31 #include "tmpfile-util.h"
32 #include "user-util.h"
35 int unlink_noerrno ( const char * path
) {
46 int rmdir_parents ( const char * path
, const char * stop
) {
55 /* Skip trailing slashes */
56 while ( l
> 0 && path
[ l
- 1 ] == '/' )
62 /* Skip last component */
63 while ( l
> 0 && path
[ l
- 1 ] != '/' )
66 /* Skip trailing slashes */
67 while ( l
> 0 && path
[ l
- 1 ] == '/' )
77 if ( path_startswith ( stop
, t
)) {
93 int rename_noreplace ( int olddirfd
, const char * oldpath
, int newdirfd
, const char * newpath
) {
96 /* Try the ideal approach first */
97 if ( renameat2 ( olddirfd
, oldpath
, newdirfd
, newpath
, RENAME_NOREPLACE
) >= 0 )
100 /* renameat2() exists since Linux 3.15, btrfs and FAT added support for it later. If it is not implemented,
101 * fall back to a different method. */
102 if (! IN_SET ( errno
, EINVAL
, ENOSYS
, ENOTTY
))
105 /* Let's try to use linkat()+unlinkat() as fallback. This doesn't work on directories and on some file systems
106 * that do not support hard links (such as FAT, most prominently), but for files it's pretty close to what we
107 * want â though not atomic (i.e. for a short period both the new and the old filename will exist). */
108 if ( linkat ( olddirfd
, oldpath
, newdirfd
, newpath
, 0 ) >= 0 ) {
110 if ( unlinkat ( olddirfd
, oldpath
, 0 ) < 0 ) {
111 r
= - errno
; /* Backup errno before the following unlinkat() alters it */
112 ( void ) unlinkat ( newdirfd
, newpath
, 0 );
119 if (! IN_SET ( errno
, EINVAL
, ENOSYS
, ENOTTY
, EPERM
)) /* FAT returns EPERM on link()âĻ */
122 /* OK, neither RENAME_NOREPLACE nor linkat()+unlinkat() worked. Let's then fallback to the racy TOCTOU
123 * vulnerable accessat(F_OK) check followed by classic, replacing renameat(), we have nothing better. */
125 if ( faccessat ( newdirfd
, newpath
, F_OK
, AT_SYMLINK_NOFOLLOW
) >= 0 )
130 if ( renameat ( olddirfd
, oldpath
, newdirfd
, newpath
) < 0 )
136 int readlinkat_malloc ( int fd
, const char * p
, char ** ret
) {
137 size_t l
= FILENAME_MAX
+ 1 ;
151 n
= readlinkat ( fd
, p
, c
, l
- 1 );
158 if (( size_t ) n
< l
- 1 ) {
169 int readlink_malloc ( const char * p
, char ** ret
) {
170 return readlinkat_malloc ( AT_FDCWD
, p
, ret
);
173 int readlink_value ( const char * p
, char ** ret
) {
174 _cleanup_free_
char * link
= NULL
;
178 r
= readlink_malloc ( p
, & link
);
182 value
= basename ( link
);
186 value
= strdup ( value
);
195 int readlink_and_make_absolute ( const char * p
, char ** r
) {
196 _cleanup_free_
char * target
= NULL
;
203 j
= readlink_malloc ( p
, & target
);
207 k
= file_in_same_dir ( p
, target
);
215 int chmod_and_chown ( const char * path
, mode_t mode
, uid_t uid
, gid_t gid
) {
216 _cleanup_close_
int fd
= - 1 ;
220 fd
= open ( path
, O_PATH
| O_CLOEXEC
| O_NOFOLLOW
); /* Let's acquire an O_PATH fd, as precaution to change
221 * mode/owner on the same file */
225 return fchmod_and_chown ( fd
, mode
, uid
, gid
);
228 int fchmod_and_chown ( int fd
, mode_t mode
, uid_t uid
, gid_t gid
) {
229 bool do_chown
, do_chmod
;
232 /* Change ownership and access mode of the specified fd. Tries to do so safely, ensuring that at no
233 * point in time the access mode is above the old access mode under the old ownership or the new
234 * access mode under the new ownership. Note: this call tries hard to leave the access mode
235 * unaffected if the uid/gid is changed, i.e. it undoes implicit suid/sgid dropping the kernel does
238 * This call is happy with O_PATH fds. */
240 if ( fstat ( fd
, & st
) < 0 )
244 ( uid
!= UID_INVALID
&& st
. st_uid
!= uid
) ||
245 ( gid
!= GID_INVALID
&& st
. st_gid
!= gid
);
248 ! S_ISLNK ( st
. st_mode
) && /* chmod is not defined on symlinks */
249 (( mode
!= MODE_INVALID
&& (( st
. st_mode
^ mode
) & 07777 ) != 0 ) ||
250 do_chown
); /* If we change ownership, make sure we reset the mode afterwards, since chown()
251 * modifies the access mode too */
253 if ( mode
== MODE_INVALID
)
254 mode
= st
. st_mode
; /* If we only shall do a chown(), save original mode, since chown() might break it. */
255 else if (( mode
& S_IFMT
) != 0 && (( mode
^ st
. st_mode
) & S_IFMT
) != 0 )
256 return - EINVAL
; /* insist on the right file type if it was specified */
258 if ( do_chown
&& do_chmod
) {
259 mode_t minimal
= st
. st_mode
& mode
; /* the subset of the old and the new mask */
261 if ((( minimal
^ st
. st_mode
) & 07777 ) != 0 )
262 if ( fchmod_opath ( fd
, minimal
& 07777 ) < 0 )
267 if ( fchownat ( fd
, "" , uid
, gid
, AT_EMPTY_PATH
) < 0 )
271 if ( fchmod_opath ( fd
, mode
& 07777 ) < 0 )
274 return do_chown
|| do_chmod
;
277 int fchmod_umask ( int fd
, mode_t m
) {
282 r
= fchmod ( fd
, m
& (~ u
)) < 0 ? - errno
: 0 ;
288 int fchmod_opath ( int fd
, mode_t m
) {
289 char procfs_path
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int )];
291 /* This function operates also on fd that might have been opened with
292 * O_PATH. Indeed fchmodat() doesn't have the AT_EMPTY_PATH flag like
293 * fchownat() does. */
295 xsprintf ( procfs_path
, "/proc/self/fd/%i" , fd
);
296 if ( chmod ( procfs_path
, m
) < 0 )
302 int fd_warn_permissions ( const char * path
, int fd
) {
305 if ( fstat ( fd
, & st
) < 0 )
308 /* Don't complain if we are reading something that is not a file, for example /dev/null */
309 if (! S_ISREG ( st
. st_mode
))
312 if ( st
. st_mode
& 0111 )
313 log_warning ( "Configuration file %s is marked executable. Please remove executable permission bits. Proceeding anyway." , path
);
315 if ( st
. st_mode
& 0002 )
316 log_warning ( "Configuration file %s is marked world-writable. Please remove world writability permission bits. Proceeding anyway." , path
);
318 if ( getpid_cached () == 1 && ( st
. st_mode
& 0044 ) != 0044 )
319 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
);
324 int touch_file ( const char * path
, bool parents
, usec_t stamp
, uid_t uid
, gid_t gid
, mode_t mode
) {
325 char fdpath
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int )];
326 _cleanup_close_
int fd
= - 1 ;
331 /* Note that touch_file() does not follow symlinks: if invoked on an existing symlink, then it is the symlink
332 * itself which is updated, not its target
334 * Returns the first error we encounter, but tries to apply as much as possible. */
337 ( void ) mkdir_parents ( path
, 0755 );
339 /* Initially, we try to open the node with O_PATH, so that we get a reference to the node. This is useful in
340 * case the path refers to an existing device or socket node, as we can open it successfully in all cases, and
341 * won't trigger any driver magic or so. */
342 fd
= open ( path
, O_PATH
| O_CLOEXEC
| O_NOFOLLOW
);
347 /* if the node doesn't exist yet, we create it, but with O_EXCL, so that we only create a regular file
348 * here, and nothing else */
349 fd
= open ( path
, O_WRONLY
| O_CREAT
| O_EXCL
| O_CLOEXEC
, IN_SET ( mode
, 0 , MODE_INVALID
) ? 0644 : mode
);
354 /* Let's make a path from the fd, and operate on that. With this logic, we can adjust the access mode,
355 * ownership and time of the file node in all cases, even if the fd refers to an O_PATH object â which is
356 * something fchown(), fchmod(), futimensat() don't allow. */
357 xsprintf ( fdpath
, "/proc/self/fd/%i" , fd
);
359 ret
= fchmod_and_chown ( fd
, mode
, uid
, gid
);
361 if ( stamp
!= USEC_INFINITY
) {
362 struct timespec ts
[ 2 ];
364 timespec_store (& ts
[ 0 ], stamp
);
366 r
= utimensat ( AT_FDCWD
, fdpath
, ts
, 0 );
368 r
= utimensat ( AT_FDCWD
, fdpath
, NULL
, 0 );
369 if ( r
< 0 && ret
>= 0 )
375 int touch ( const char * path
) {
376 return touch_file ( path
, false , USEC_INFINITY
, UID_INVALID
, GID_INVALID
, MODE_INVALID
);
379 int symlink_idempotent ( const char * from
, const char * to
, bool make_relative
) {
380 _cleanup_free_
char * relpath
= NULL
;
387 _cleanup_free_
char * parent
= NULL
;
389 parent
= dirname_malloc ( to
);
393 r
= path_make_relative ( parent
, from
, & relpath
);
400 if ( symlink ( from
, to
) < 0 ) {
401 _cleanup_free_
char * p
= NULL
;
406 r
= readlink_malloc ( to
, & p
);
407 if ( r
== - EINVAL
) /* Not a symlink? In that case return the original error we encountered: -EEXIST */
409 if ( r
< 0 ) /* Any other error? In that case propagate it as is */
412 if (! streq ( p
, from
)) /* Not the symlink we want it to be? In that case, propagate the original -EEXIST */
419 int symlink_atomic ( const char * from
, const char * to
) {
420 _cleanup_free_
char * t
= NULL
;
426 r
= tempfn_random ( to
, NULL
, & t
);
430 if ( symlink ( from
, t
) < 0 )
433 if ( rename ( t
, to
) < 0 ) {
441 int mknod_atomic ( const char * path
, mode_t mode
, dev_t dev
) {
442 _cleanup_free_
char * t
= NULL
;
447 r
= tempfn_random ( path
, NULL
, & t
);
451 if ( mknod ( t
, mode
, dev
) < 0 )
454 if ( rename ( t
, path
) < 0 ) {
462 int mkfifo_atomic ( const char * path
, mode_t mode
) {
463 _cleanup_free_
char * t
= NULL
;
468 r
= tempfn_random ( path
, NULL
, & t
);
472 if ( mkfifo ( t
, mode
) < 0 )
475 if ( rename ( t
, path
) < 0 ) {
483 int mkfifoat_atomic ( int dirfd
, const char * path
, mode_t mode
) {
484 _cleanup_free_
char * t
= NULL
;
489 if ( path_is_absolute ( path
))
490 return mkfifo_atomic ( path
, mode
);
492 /* We're only interested in the (random) filename. */
493 r
= tempfn_random_child ( "" , NULL
, & t
);
497 if ( mkfifoat ( dirfd
, t
, mode
) < 0 )
500 if ( renameat ( dirfd
, t
, dirfd
, path
) < 0 ) {
508 int get_files_in_directory ( const char * path
, char *** list
) {
509 _cleanup_closedir_
DIR * d
= NULL
;
511 size_t bufsize
= 0 , n
= 0 ;
512 _cleanup_strv_free_
char ** l
= NULL
;
516 /* Returns all files in a directory in *list, and the number
517 * of files as return value. If list is NULL returns only the
524 FOREACH_DIRENT_ALL ( de
, d
, return - errno
) {
525 dirent_ensure_type ( d
, de
);
527 if (! dirent_is_file ( de
))
531 /* one extra slot is needed for the terminating NULL */
532 if (! GREEDY_REALLOC ( l
, bufsize
, n
+ 2 ))
535 l
[ n
] = strdup ( de
-> d_name
);
550 static int getenv_tmp_dir ( const char ** ret_path
) {
556 /* We use the same order of environment variables python uses in tempfile.gettempdir():
557 * https://docs.python.org/3/library/tempfile.html#tempfile.gettempdir */
558 FOREACH_STRING ( n
, "TMPDIR" , "TEMP" , "TMP" ) {
561 e
= secure_getenv ( n
);
564 if (! path_is_absolute ( e
)) {
568 if (! path_is_normalized ( e
)) {
585 /* Remember first error, to make this more debuggable */
597 static int tmp_dir_internal ( const char * def
, const char ** ret
) {
604 r
= getenv_tmp_dir (& e
);
610 k
= is_dir ( def
, true );
614 return r
< 0 ? r
: k
;
620 int var_tmp_dir ( const char ** ret
) {
622 /* Returns the location for "larger" temporary files, that is backed by physical storage if available, and thus
623 * even might survive a boot: /var/tmp. If $TMPDIR (or related environment variables) are set, its value is
624 * returned preferably however. Note that both this function and tmp_dir() below are affected by $TMPDIR,
625 * making it a variable that overrides all temporary file storage locations. */
627 return tmp_dir_internal ( "/var/tmp" , ret
);
630 int tmp_dir ( const char ** ret
) {
632 /* Similar to var_tmp_dir() above, but returns the location for "smaller" temporary files, which is usually
633 * backed by an in-memory file system: /tmp. */
635 return tmp_dir_internal ( "/tmp" , ret
);
638 int unlink_or_warn ( const char * filename
) {
639 if ( unlink ( filename
) < 0 && errno
!= ENOENT
)
640 /* If the file doesn't exist and the fs simply was read-only (in which
641 * case unlink() returns EROFS even if the file doesn't exist), don't
643 if ( errno
!= EROFS
|| access ( filename
, F_OK
) >= 0 )
644 return log_error_errno ( errno
, "Failed to remove \" %s \" : %m" , filename
);
649 int inotify_add_watch_fd ( int fd
, int what
, uint32_t mask
) {
650 char path
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int ) + 1 ];
653 /* This is like inotify_add_watch(), except that the file to watch is not referenced by a path, but by an fd */
654 xsprintf ( path
, "/proc/self/fd/%i" , what
);
656 r
= inotify_add_watch ( fd
, path
, mask
);
663 static bool unsafe_transition ( const struct stat
* a
, const struct stat
* b
) {
664 /* Returns true if the transition from a to b is safe, i.e. that we never transition from unprivileged to
665 * privileged files or directories. Why bother? So that unprivileged code can't symlink to privileged files
666 * making us believe we read something safe even though it isn't safe in the specific context we open it in. */
668 if ( a
-> st_uid
== 0 ) /* Transitioning from privileged to unprivileged is always fine */
671 return a
-> st_uid
!= b
-> st_uid
; /* Otherwise we need to stay within the same UID */
674 static int log_unsafe_transition ( int a
, int b
, const char * path
, unsigned flags
) {
675 _cleanup_free_
char * n1
= NULL
, * n2
= NULL
;
677 if (! FLAGS_SET ( flags
, CHASE_WARN
))
680 ( void ) fd_get_path ( a
, & n1
);
681 ( void ) fd_get_path ( b
, & n2
);
683 return log_warning_errno ( SYNTHETIC_ERRNO ( ENOLINK
),
684 "Detected unsafe path transition %s %s %s during canonicalization of %s." ,
685 n1
, special_glyph ( SPECIAL_GLYPH_ARROW
), n2
, path
);
688 static int log_autofs_mount_point ( int fd
, const char * path
, unsigned flags
) {
689 _cleanup_free_
char * n1
= NULL
;
691 if (! FLAGS_SET ( flags
, CHASE_WARN
))
694 ( void ) fd_get_path ( fd
, & n1
);
696 return log_warning_errno ( SYNTHETIC_ERRNO ( EREMOTE
),
697 "Detected autofs mount point %s during canonicalization of %s." ,
701 int chase_symlinks ( const char * path
, const char * original_root
, unsigned flags
, char ** ret
) {
702 _cleanup_free_
char * buffer
= NULL
, * done
= NULL
, * root
= NULL
;
703 _cleanup_close_
int fd
= - 1 ;
704 unsigned max_follow
= CHASE_SYMLINKS_MAX
; /* how many symlinks to follow before giving up and returning ELOOP */
705 struct stat previous_stat
;
712 /* Either the file may be missing, or we return an fd to the final object, but both make no sense */
713 if ( FLAGS_SET ( flags
, CHASE_NONEXISTENT
| CHASE_OPEN
))
716 if ( FLAGS_SET ( flags
, CHASE_STEP
| CHASE_OPEN
))
722 /* This is a lot like canonicalize_file_name(), but takes an additional "root" parameter, that allows following
723 * symlinks relative to a root directory, instead of the root of the host.
725 * Note that "root" primarily matters if we encounter an absolute symlink. It is also used when following
726 * relative symlinks to ensure they cannot be used to "escape" the root directory. The path parameter passed is
727 * assumed to be already prefixed by it, except if the CHASE_PREFIX_ROOT flag is set, in which case it is first
728 * prefixed accordingly.
730 * Algorithmically this operates on two path buffers: "done" are the components of the path we already
731 * processed and resolved symlinks, "." and ".." of. "todo" are the components of the path we still need to
732 * process. On each iteration, we move one component from "todo" to "done", processing it's special meaning
733 * each time. The "todo" path always starts with at least one slash, the "done" path always ends in no
734 * slash. We always keep an O_PATH fd to the component we are currently processing, thus keeping lookup races
737 * Suggested usage: whenever you want to canonicalize a path, use this function. Pass the absolute path you got
738 * as-is: fully qualified and relative to your host's root. Optionally, specify the root parameter to tell this
739 * function what to do when encountering a symlink with an absolute path as directory: prefix it by the
742 * There are three ways to invoke this function:
744 * 1. Without CHASE_STEP or CHASE_OPEN: in this case the path is resolved and the normalized path is returned
745 * in `ret`. The return value is < 0 on error. If CHASE_NONEXISTENT is also set 0 is returned if the file
746 * doesn't exist, > 0 otherwise. If CHASE_NONEXISTENT is not set >= 0 is returned if the destination was
747 * found, -ENOENT if it doesn't.
749 * 2. With CHASE_OPEN: in this case the destination is opened after chasing it as O_PATH and this file
750 * descriptor is returned as return value. This is useful to open files relative to some root
751 * directory. Note that the returned O_PATH file descriptors must be converted into a regular one (using
752 * fd_reopen() or such) before it can be used for reading/writing. CHASE_OPEN may not be combined with
755 * 3. With CHASE_STEP: in this case only a single step of the normalization is executed, i.e. only the first
756 * symlink or ".." component of the path is resolved, and the resulting path is returned. This is useful if
757 * a caller wants to trace the a path through the file system verbosely. Returns < 0 on error, > 0 if the
758 * path is fully normalized, and == 0 for each normalization step. This may be combined with
759 * CHASE_NONEXISTENT, in which case 1 is returned when a component is not found.
761 * 4. With CHASE_SAFE: in this case the path must not contain unsafe transitions, i.e. transitions from
762 * unprivileged to privileged files or directories. In such cases the return value is -ENOLINK. If
763 * CHASE_WARN is also set a warning describing the unsafe transition is emitted.
765 * 5. With CHASE_NO_AUTOFS: in this case if an autofs mount point is encountered, the path normalization is
766 * aborted and -EREMOTE is returned. If CHASE_WARN is also set a warning showing the path of the mount point
771 /* A root directory of "/" or "" is identical to none */
772 if ( empty_or_root ( original_root
))
773 original_root
= NULL
;
775 if (! original_root
&& ! ret
&& ( flags
& ( CHASE_NONEXISTENT
| CHASE_NO_AUTOFS
| CHASE_SAFE
| CHASE_OPEN
| CHASE_STEP
)) == CHASE_OPEN
) {
776 /* Shortcut the CHASE_OPEN case if the caller isn't interested in the actual path and has no root set
777 * and doesn't care about any of the other special features we provide either. */
778 r
= open ( path
, O_PATH
| O_CLOEXEC
|(( flags
& CHASE_NOFOLLOW
) ? O_NOFOLLOW
: 0 ));
786 r
= path_make_absolute_cwd ( original_root
, & root
);
790 if ( flags
& CHASE_PREFIX_ROOT
) {
792 /* We don't support relative paths in combination with a root directory */
793 if (! path_is_absolute ( path
))
796 path
= prefix_roota ( root
, path
);
800 r
= path_make_absolute_cwd ( path
, & buffer
);
804 fd
= open ( "/" , O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
808 if ( flags
& CHASE_SAFE
) {
809 if ( fstat ( fd
, & previous_stat
) < 0 )
815 _cleanup_free_
char * first
= NULL
;
816 _cleanup_close_
int child
= - 1 ;
820 /* Determine length of first component in the path */
821 n
= strspn ( todo
, "/" ); /* The slashes */
822 m
= n
+ strcspn ( todo
+ n
, "/" ); /* The entire length of the component */
824 /* Extract the first component. */
825 first
= strndup ( todo
, m
);
831 /* Empty? Then we reached the end. */
835 /* Just a single slash? Then we reached the end. */
836 if ( path_equal ( first
, "/" )) {
837 /* Preserve the trailing slash */
839 if ( flags
& CHASE_TRAIL_SLASH
)
840 if (! strextend (& done
, "/" , NULL
))
846 /* Just a dot? Then let's eat this up. */
847 if ( path_equal ( first
, "/." ))
850 /* Two dots? Then chop off the last bit of what we already found out. */
851 if ( path_equal ( first
, "/.." )) {
852 _cleanup_free_
char * parent
= NULL
;
853 _cleanup_close_
int fd_parent
= - 1 ;
855 /* If we already are at the top, then going up will not change anything. This is in-line with
856 * how the kernel handles this. */
857 if ( empty_or_root ( done
))
860 parent
= dirname_malloc ( done
);
864 /* Don't allow this to leave the root dir. */
866 path_startswith ( done
, root
) &&
867 ! path_startswith ( parent
, root
))
870 free_and_replace ( done
, parent
);
872 if ( flags
& CHASE_STEP
)
875 fd_parent
= openat ( fd
, ".." , O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
879 if ( flags
& CHASE_SAFE
) {
880 if ( fstat ( fd_parent
, & st
) < 0 )
883 if ( unsafe_transition (& previous_stat
, & st
))
884 return log_unsafe_transition ( fd
, fd_parent
, path
, flags
);
890 fd
= TAKE_FD ( fd_parent
);
895 /* Otherwise let's see what this is. */
896 child
= openat ( fd
, first
+ n
, O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
899 if ( errno
== ENOENT
&&
900 ( flags
& CHASE_NONEXISTENT
) &&
901 ( isempty ( todo
) || path_is_normalized ( todo
))) {
903 /* If CHASE_NONEXISTENT is set, and the path does not exist, then that's OK, return
904 * what we got so far. But don't allow this if the remaining path contains "../ or "./"
905 * or something else weird. */
907 /* If done is "/", as first also contains slash at the head, then remove this redundant slash. */
908 if ( streq_ptr ( done
, "/" ))
911 if (! strextend (& done
, first
, todo
, NULL
))
921 if ( fstat ( child
, & st
) < 0 )
923 if (( flags
& CHASE_SAFE
) &&
924 ( empty_or_root ( root
) || ( size_t )( todo
- buffer
) > strlen ( root
)) &&
925 unsafe_transition (& previous_stat
, & st
))
926 return log_unsafe_transition ( fd
, child
, path
, flags
);
930 if (( flags
& CHASE_NO_AUTOFS
) &&
931 fd_is_fs_type ( child
, AUTOFS_SUPER_MAGIC
) > 0 )
932 return log_autofs_mount_point ( child
, path
, flags
);
934 if ( S_ISLNK ( st
. st_mode
) && !(( flags
& CHASE_NOFOLLOW
) && isempty ( todo
))) {
937 _cleanup_free_
char * destination
= NULL
;
939 /* This is a symlink, in this case read the destination. But let's make sure we don't follow
940 * symlinks without bounds. */
941 if (-- max_follow
<= 0 )
944 r
= readlinkat_malloc ( fd
, first
+ n
, & destination
);
947 if ( isempty ( destination
))
950 if ( path_is_absolute ( destination
)) {
952 /* An absolute destination. Start the loop from the beginning, but use the root
953 * directory as base. */
956 fd
= open ( root
?: "/" , O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
960 if ( flags
& CHASE_SAFE
) {
961 if ( fstat ( fd
, & st
) < 0 )
964 if ( unsafe_transition (& previous_stat
, & st
))
965 return log_unsafe_transition ( child
, fd
, path
, flags
);
972 /* Note that we do not revalidate the root, we take it as is. */
981 /* Prefix what's left to do with what we just read, and start the loop again, but
982 * remain in the current directory. */
983 joined
= strjoin ( destination
, todo
);
985 joined
= strjoin ( "/" , destination
, todo
);
990 todo
= buffer
= joined
;
992 if ( flags
& CHASE_STEP
)
998 /* If this is not a symlink, then let's just add the name we read to what we already verified. */
1000 done
= TAKE_PTR ( first
);
1002 /* If done is "/", as first also contains slash at the head, then remove this redundant slash. */
1003 if ( streq ( done
, "/" ))
1006 if (! strextend (& done
, first
, NULL
))
1010 /* And iterate again, but go one directory further down. */
1012 fd
= TAKE_FD ( child
);
1016 /* Special case, turn the empty string into "/", to indicate the root directory. */
1023 * ret
= TAKE_PTR ( done
);
1025 if ( flags
& CHASE_OPEN
) {
1026 /* Return the O_PATH fd we currently are looking to the caller. It can translate it to a proper fd by
1027 * opening /proc/self/fd/xyz. */
1033 if ( flags
& CHASE_STEP
)
1042 c
= strjoin ( strempty ( done
), todo
);
1052 int chase_symlinks_and_open (
1055 unsigned chase_flags
,
1059 _cleanup_close_
int path_fd
= - 1 ;
1060 _cleanup_free_
char * p
= NULL
;
1063 if ( chase_flags
& CHASE_NONEXISTENT
)
1066 if ( empty_or_root ( root
) && ! ret_path
&& ( chase_flags
& ( CHASE_NO_AUTOFS
| CHASE_SAFE
)) == 0 ) {
1067 /* Shortcut this call if none of the special features of this call are requested */
1068 r
= open ( path
, open_flags
);
1075 path_fd
= chase_symlinks ( path
, root
, chase_flags
| CHASE_OPEN
, ret_path
? & p
: NULL
);
1079 r
= fd_reopen ( path_fd
, open_flags
);
1084 * ret_path
= TAKE_PTR ( p
);
1089 int chase_symlinks_and_opendir (
1092 unsigned chase_flags
,
1096 char procfs_path
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int )];
1097 _cleanup_close_
int path_fd
= - 1 ;
1098 _cleanup_free_
char * p
= NULL
;
1103 if ( chase_flags
& CHASE_NONEXISTENT
)
1106 if ( empty_or_root ( root
) && ! ret_path
&& ( chase_flags
& ( CHASE_NO_AUTOFS
| CHASE_SAFE
)) == 0 ) {
1107 /* Shortcut this call if none of the special features of this call are requested */
1116 path_fd
= chase_symlinks ( path
, root
, chase_flags
| CHASE_OPEN
, ret_path
? & p
: NULL
);
1120 xsprintf ( procfs_path
, "/proc/self/fd/%i" , path_fd
);
1121 d
= opendir ( procfs_path
);
1126 * ret_path
= TAKE_PTR ( p
);
1132 int chase_symlinks_and_stat (
1135 unsigned chase_flags
,
1137 struct stat
* ret_stat
) {
1139 _cleanup_close_
int path_fd
= - 1 ;
1140 _cleanup_free_
char * p
= NULL
;
1145 if ( chase_flags
& CHASE_NONEXISTENT
)
1148 if ( empty_or_root ( root
) && ! ret_path
&& ( chase_flags
& ( CHASE_NO_AUTOFS
| CHASE_SAFE
)) == 0 ) {
1149 /* Shortcut this call if none of the special features of this call are requested */
1150 if ( stat ( path
, ret_stat
) < 0 )
1156 path_fd
= chase_symlinks ( path
, root
, chase_flags
| CHASE_OPEN
, ret_path
? & p
: NULL
);
1160 if ( fstat ( path_fd
, ret_stat
) < 0 )
1164 * ret_path
= TAKE_PTR ( p
);
1166 if ( chase_flags
& CHASE_OPEN
)
1167 return TAKE_FD ( path_fd
);
1172 int access_fd ( int fd
, int mode
) {
1173 char p
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( fd
) + 1 ];
1176 /* Like access() but operates on an already open fd */
1178 xsprintf ( p
, "/proc/self/fd/%i" , fd
);
1179 r
= access ( p
, mode
);
1186 void unlink_tempfilep ( char (* p
)[]) {
1187 /* If the file is created with mkstemp(), it will (almost always)
1188 * change the suffix. Treat this as a sign that the file was
1189 * successfully created. We ignore both the rare case where the
1190 * original suffix is used and unlink failures. */
1191 if (! endswith (* p
, ".XXXXXX" ))
1192 ( void ) unlink_noerrno (* p
);
1195 int unlinkat_deallocate ( int fd
, const char * name
, int flags
) {
1196 _cleanup_close_
int truncate_fd
= - 1 ;
1200 /* Operates like unlinkat() but also deallocates the file contents if it is a regular file and there's no other
1201 * link to it. This is useful to ensure that other processes that might have the file open for reading won't be
1202 * able to keep the data pinned on disk forever. This call is particular useful whenever we execute clean-up
1203 * jobs ("vacuuming"), where we want to make sure the data is really gone and the disk space released and
1204 * returned to the free pool.
1206 * Deallocation is preferably done by FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE (đ) if supported, which means
1207 * the file won't change size. That's a good thing since we shouldn't needlessly trigger SIGBUS in other
1208 * programs that have mmap()ed the file. (The assumption here is that changing file contents to all zeroes
1209 * underneath those programs is the better choice than simply triggering SIGBUS in them which truncation does.)
1210 * However if hole punching is not implemented in the kernel or file system we'll fall back to normal file
1211 * truncation (đĒ), as our goal of deallocating the data space trumps our goal of being nice to readers (đ).
1213 * Note that we attempt deallocation, but failure to succeed with that is not considered fatal, as long as the
1214 * primary job â to delete the file â is accomplished. */
1216 if (( flags
& AT_REMOVEDIR
) == 0 ) {
1217 truncate_fd
= openat ( fd
, name
, O_WRONLY
| O_CLOEXEC
| O_NOCTTY
| O_NOFOLLOW
| O_NONBLOCK
);
1218 if ( truncate_fd
< 0 ) {
1220 /* If this failed because the file doesn't exist propagate the error right-away. Also,
1221 * AT_REMOVEDIR wasn't set, and we tried to open the file for writing, which means EISDIR is
1222 * returned when this is a directory but we are not supposed to delete those, hence propagate
1223 * the error right-away too. */
1224 if ( IN_SET ( errno
, ENOENT
, EISDIR
))
1227 if ( errno
!= ELOOP
) /* don't complain if this is a symlink */
1228 log_debug_errno ( errno
, "Failed to open file '%s' for deallocation, ignoring: %m" , name
);
1232 if ( unlinkat ( fd
, name
, flags
) < 0 )
1235 if ( truncate_fd
< 0 ) /* Don't have a file handle, can't do more âšī¸ */
1238 if ( fstat ( truncate_fd
, & st
) < 0 ) {
1239 log_debug_errno ( errno
, "Failed to stat file '%s' for deallocation, ignoring: %m" , name
);
1243 if (! S_ISREG ( st
. st_mode
) || st
. st_blocks
== 0 || st
. st_nlink
> 0 )
1246 /* If this is a regular file, it actually took up space on disk and there are no other links it's time to
1247 * punch-hole/truncate this to release the disk space. */
1249 bs
= MAX ( st
. st_blksize
, 512 );
1250 l
= DIV_ROUND_UP ( st
. st_size
, bs
) * bs
; /* Round up to next block size */
1252 if ( fallocate ( truncate_fd
, FALLOC_FL_PUNCH_HOLE
| FALLOC_FL_KEEP_SIZE
, 0 , l
) >= 0 )
1253 return 0 ; /* Successfully punched a hole! đ */
1255 /* Fall back to truncation */
1256 if ( ftruncate ( truncate_fd
, 0 ) < 0 ) {
1257 log_debug_errno ( errno
, "Failed to truncate file to 0, ignoring: %m" );
1264 int fsync_directory_of_file ( int fd
) {
1265 _cleanup_free_
char * path
= NULL
;
1266 _cleanup_close_
int dfd
= - 1 ;
1269 r
= fd_verify_regular ( fd
);
1273 r
= fd_get_path ( fd
, & path
);
1275 log_debug_errno ( r
, "Failed to query /proc/self/fd/%d%s: %m" ,
1277 r
== - EOPNOTSUPP
? ", ignoring" : "" );
1279 if ( r
== - EOPNOTSUPP
)
1280 /* If /proc is not available, we're most likely running in some
1281 * chroot environment, and syncing the directory is not very
1282 * important in that case. Let's just silently do nothing. */
1288 if (! path_is_absolute ( path
))
1291 dfd
= open_parent ( path
, O_CLOEXEC
, 0 );
1301 int fsync_path_at ( int at_fd
, const char * path
) {
1302 _cleanup_close_
int opened_fd
= - 1 ;
1305 if ( isempty ( path
)) {
1306 if ( at_fd
== AT_FDCWD
) {
1307 opened_fd
= open ( "." , O_RDONLY
| O_DIRECTORY
| O_CLOEXEC
);
1316 opened_fd
= openat ( at_fd
, path
, O_RDONLY
| O_CLOEXEC
);
1329 int syncfs_path ( int atfd
, const char * path
) {
1330 _cleanup_close_
int fd
= - 1 ;
1334 fd
= openat ( atfd
, path
, O_CLOEXEC
| O_RDONLY
| O_NONBLOCK
);
1344 int open_parent ( const char * path
, int flags
, mode_t mode
) {
1345 _cleanup_free_
char * parent
= NULL
;
1350 if ( path_equal ( path
, "/" )) /* requesting the parent of the root dir is fishy, let's prohibit that */
1353 parent
= dirname_malloc ( path
);
1357 /* Let's insist on O_DIRECTORY since the parent of a file or directory is a directory. Except if we open an
1358 * O_TMPFILE file, because in that case we are actually create a regular file below the parent directory. */
1360 if ( FLAGS_SET ( flags
, O_PATH
))
1361 flags
|= O_DIRECTORY
;
1362 else if (! FLAGS_SET ( flags
, O_TMPFILE
))
1363 flags
|= O_DIRECTORY
| O_RDONLY
;
1365 fd
= open ( parent
, flags
, mode
);