]>
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"
21 #include "parse-util.h"
22 #include "path-util.h"
23 #include "process-util.h"
24 #include "stat-util.h"
25 #include "stdio-util.h"
26 #include "string-util.h"
28 #include "time-util.h"
29 #include "tmpfile-util.h"
30 #include "user-util.h"
33 int unlink_noerrno ( const char * path
) {
44 int rmdir_parents ( const char * path
, const char * stop
) {
53 /* Skip trailing slashes */
54 while ( l
> 0 && path
[ l
- 1 ] == '/' )
60 /* Skip last component */
61 while ( l
> 0 && path
[ l
- 1 ] != '/' )
64 /* Skip trailing slashes */
65 while ( l
> 0 && path
[ l
- 1 ] == '/' )
75 if ( path_startswith ( stop
, t
)) {
91 int rename_noreplace ( int olddirfd
, const char * oldpath
, int newdirfd
, const char * newpath
) {
94 /* Try the ideal approach first */
95 if ( renameat2 ( olddirfd
, oldpath
, newdirfd
, newpath
, RENAME_NOREPLACE
) >= 0 )
98 /* renameat2() exists since Linux 3.15, btrfs and FAT added support for it later. If it is not implemented,
99 * fall back to a different method. */
100 if (! IN_SET ( errno
, EINVAL
, ENOSYS
, ENOTTY
))
103 /* Let's try to use linkat()+unlinkat() as fallback. This doesn't work on directories and on some file systems
104 * that do not support hard links (such as FAT, most prominently), but for files it's pretty close to what we
105 * want â though not atomic (i.e. for a short period both the new and the old filename will exist). */
106 if ( linkat ( olddirfd
, oldpath
, newdirfd
, newpath
, 0 ) >= 0 ) {
108 if ( unlinkat ( olddirfd
, oldpath
, 0 ) < 0 ) {
109 r
= - errno
; /* Backup errno before the following unlinkat() alters it */
110 ( void ) unlinkat ( newdirfd
, newpath
, 0 );
117 if (! IN_SET ( errno
, EINVAL
, ENOSYS
, ENOTTY
, EPERM
)) /* FAT returns EPERM on link()âĻ */
120 /* OK, neither RENAME_NOREPLACE nor linkat()+unlinkat() worked. Let's then fallback to the racy TOCTOU
121 * vulnerable accessat(F_OK) check followed by classic, replacing renameat(), we have nothing better. */
123 if ( faccessat ( newdirfd
, newpath
, F_OK
, AT_SYMLINK_NOFOLLOW
) >= 0 )
128 if ( renameat ( olddirfd
, oldpath
, newdirfd
, newpath
) < 0 )
134 int readlinkat_malloc ( int fd
, const char * p
, char ** ret
) {
135 size_t l
= FILENAME_MAX
+ 1 ;
149 n
= readlinkat ( fd
, p
, c
, l
- 1 );
156 if (( size_t ) n
< l
- 1 ) {
167 int readlink_malloc ( const char * p
, char ** ret
) {
168 return readlinkat_malloc ( AT_FDCWD
, p
, ret
);
171 int readlink_value ( const char * p
, char ** ret
) {
172 _cleanup_free_
char * link
= NULL
;
176 r
= readlink_malloc ( p
, & link
);
180 value
= basename ( link
);
184 value
= strdup ( value
);
193 int readlink_and_make_absolute ( const char * p
, char ** r
) {
194 _cleanup_free_
char * target
= NULL
;
201 j
= readlink_malloc ( p
, & target
);
205 k
= file_in_same_dir ( p
, target
);
213 int chmod_and_chown ( const char * path
, mode_t mode
, uid_t uid
, gid_t gid
) {
214 char fd_path
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int ) + 1 ];
215 _cleanup_close_
int fd
= - 1 ;
218 /* Under the assumption that we are running privileged we first change the access mode and only then hand out
219 * ownership to avoid a window where access is too open. */
221 fd
= open ( path
, O_PATH
| O_CLOEXEC
| O_NOFOLLOW
); /* Let's acquire an O_PATH fd, as precaution to change mode/owner
222 * on the same file */
226 xsprintf ( fd_path
, "/proc/self/fd/%i" , fd
);
228 if ( mode
!= MODE_INVALID
) {
230 if (( mode
& S_IFMT
) != 0 ) {
233 if ( stat ( fd_path
, & st
) < 0 )
236 if (( mode
& S_IFMT
) != ( st
. st_mode
& S_IFMT
))
240 if ( chmod ( fd_path
, mode
& 07777 ) < 0 )
244 if ( uid
!= UID_INVALID
|| gid
!= GID_INVALID
)
245 if ( chown ( fd_path
, uid
, gid
) < 0 )
251 int fchmod_and_chown ( int fd
, mode_t mode
, uid_t uid
, gid_t gid
) {
252 /* Under the assumption that we are running privileged we first change the access mode and only then hand out
253 * ownership to avoid a window where access is too open. */
255 if ( mode
!= MODE_INVALID
) {
257 if (( mode
& S_IFMT
) != 0 ) {
260 if ( fstat ( fd
, & st
) < 0 )
263 if (( mode
& S_IFMT
) != ( st
. st_mode
& S_IFMT
))
267 if ( fchmod ( fd
, mode
& 0777 ) < 0 )
271 if ( uid
!= UID_INVALID
|| gid
!= GID_INVALID
)
272 if ( fchown ( fd
, uid
, gid
) < 0 )
278 int fchmod_umask ( int fd
, mode_t m
) {
283 r
= fchmod ( fd
, m
& (~ u
)) < 0 ? - errno
: 0 ;
289 int fchmod_opath ( int fd
, mode_t m
) {
290 char procfs_path
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int )];
292 /* This function operates also on fd that might have been opened with
293 * O_PATH. Indeed fchmodat() doesn't have the AT_EMPTY_PATH flag like
294 * fchownat() does. */
296 xsprintf ( procfs_path
, "/proc/self/fd/%i" , fd
);
297 if ( chmod ( procfs_path
, m
) < 0 )
303 int fd_warn_permissions ( const char * path
, int fd
) {
306 if ( fstat ( fd
, & st
) < 0 )
309 if ( st
. st_mode
& 0111 )
310 log_warning ( "Configuration file %s is marked executable. Please remove executable permission bits. Proceeding anyway." , path
);
312 if ( st
. st_mode
& 0002 )
313 log_warning ( "Configuration file %s is marked world-writable. Please remove world writability permission bits. Proceeding anyway." , path
);
315 if ( getpid_cached () == 1 && ( st
. st_mode
& 0044 ) != 0044 )
316 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
);
321 int touch_file ( const char * path
, bool parents
, usec_t stamp
, uid_t uid
, gid_t gid
, mode_t mode
) {
322 char fdpath
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int )];
323 _cleanup_close_
int fd
= - 1 ;
328 /* Note that touch_file() does not follow symlinks: if invoked on an existing symlink, then it is the symlink
329 * itself which is updated, not its target
331 * Returns the first error we encounter, but tries to apply as much as possible. */
334 ( void ) mkdir_parents ( path
, 0755 );
336 /* Initially, we try to open the node with O_PATH, so that we get a reference to the node. This is useful in
337 * case the path refers to an existing device or socket node, as we can open it successfully in all cases, and
338 * won't trigger any driver magic or so. */
339 fd
= open ( path
, O_PATH
| O_CLOEXEC
| O_NOFOLLOW
);
344 /* if the node doesn't exist yet, we create it, but with O_EXCL, so that we only create a regular file
345 * here, and nothing else */
346 fd
= open ( path
, O_WRONLY
| O_CREAT
| O_EXCL
| O_CLOEXEC
, IN_SET ( mode
, 0 , MODE_INVALID
) ? 0644 : mode
);
351 /* Let's make a path from the fd, and operate on that. With this logic, we can adjust the access mode,
352 * ownership and time of the file node in all cases, even if the fd refers to an O_PATH object â which is
353 * something fchown(), fchmod(), futimensat() don't allow. */
354 xsprintf ( fdpath
, "/proc/self/fd/%i" , fd
);
356 if ( mode
!= MODE_INVALID
)
357 if ( chmod ( fdpath
, mode
) < 0 )
360 if ( uid_is_valid ( uid
) || gid_is_valid ( gid
))
361 if ( chown ( fdpath
, uid
, gid
) < 0 && ret
>= 0 )
364 if ( stamp
!= USEC_INFINITY
) {
365 struct timespec ts
[ 2 ];
367 timespec_store (& ts
[ 0 ], stamp
);
369 r
= utimensat ( AT_FDCWD
, fdpath
, ts
, 0 );
371 r
= utimensat ( AT_FDCWD
, fdpath
, NULL
, 0 );
372 if ( r
< 0 && ret
>= 0 )
378 int touch ( const char * path
) {
379 return touch_file ( path
, false , USEC_INFINITY
, UID_INVALID
, GID_INVALID
, MODE_INVALID
);
382 int symlink_idempotent ( const char * from
, const char * to
, bool make_relative
) {
383 _cleanup_free_
char * relpath
= NULL
;
390 _cleanup_free_
char * parent
= NULL
;
392 parent
= dirname_malloc ( to
);
396 r
= path_make_relative ( parent
, from
, & relpath
);
403 if ( symlink ( from
, to
) < 0 ) {
404 _cleanup_free_
char * p
= NULL
;
409 r
= readlink_malloc ( to
, & p
);
410 if ( r
== - EINVAL
) /* Not a symlink? In that case return the original error we encountered: -EEXIST */
412 if ( r
< 0 ) /* Any other error? In that case propagate it as is */
415 if (! streq ( p
, from
)) /* Not the symlink we want it to be? In that case, propagate the original -EEXIST */
422 int symlink_atomic ( const char * from
, const char * to
) {
423 _cleanup_free_
char * t
= NULL
;
429 r
= tempfn_random ( to
, NULL
, & t
);
433 if ( symlink ( from
, t
) < 0 )
436 if ( rename ( t
, to
) < 0 ) {
444 int mknod_atomic ( const char * path
, mode_t mode
, dev_t dev
) {
445 _cleanup_free_
char * t
= NULL
;
450 r
= tempfn_random ( path
, NULL
, & t
);
454 if ( mknod ( t
, mode
, dev
) < 0 )
457 if ( rename ( t
, path
) < 0 ) {
465 int mkfifo_atomic ( const char * path
, mode_t mode
) {
466 _cleanup_free_
char * t
= NULL
;
471 r
= tempfn_random ( path
, NULL
, & t
);
475 if ( mkfifo ( t
, mode
) < 0 )
478 if ( rename ( t
, path
) < 0 ) {
486 int mkfifoat_atomic ( int dirfd
, const char * path
, mode_t mode
) {
487 _cleanup_free_
char * t
= NULL
;
492 if ( path_is_absolute ( path
))
493 return mkfifo_atomic ( path
, mode
);
495 /* We're only interested in the (random) filename. */
496 r
= tempfn_random_child ( "" , NULL
, & t
);
500 if ( mkfifoat ( dirfd
, t
, mode
) < 0 )
503 if ( renameat ( dirfd
, t
, dirfd
, path
) < 0 ) {
511 int get_files_in_directory ( const char * path
, char *** list
) {
512 _cleanup_closedir_
DIR * d
= NULL
;
514 size_t bufsize
= 0 , n
= 0 ;
515 _cleanup_strv_free_
char ** l
= NULL
;
519 /* Returns all files in a directory in *list, and the number
520 * of files as return value. If list is NULL returns only the
527 FOREACH_DIRENT_ALL ( de
, d
, return - errno
) {
528 dirent_ensure_type ( d
, de
);
530 if (! dirent_is_file ( de
))
534 /* one extra slot is needed for the terminating NULL */
535 if (! GREEDY_REALLOC ( l
, bufsize
, n
+ 2 ))
538 l
[ n
] = strdup ( de
-> d_name
);
553 static int getenv_tmp_dir ( const char ** ret_path
) {
559 /* We use the same order of environment variables python uses in tempfile.gettempdir():
560 * https://docs.python.org/3/library/tempfile.html#tempfile.gettempdir */
561 FOREACH_STRING ( n
, "TMPDIR" , "TEMP" , "TMP" ) {
564 e
= secure_getenv ( n
);
567 if (! path_is_absolute ( e
)) {
571 if (! path_is_normalized ( e
)) {
588 /* Remember first error, to make this more debuggable */
600 static int tmp_dir_internal ( const char * def
, const char ** ret
) {
607 r
= getenv_tmp_dir (& e
);
613 k
= is_dir ( def
, true );
617 return r
< 0 ? r
: k
;
623 int var_tmp_dir ( const char ** ret
) {
625 /* Returns the location for "larger" temporary files, that is backed by physical storage if available, and thus
626 * even might survive a boot: /var/tmp. If $TMPDIR (or related environment variables) are set, its value is
627 * returned preferably however. Note that both this function and tmp_dir() below are affected by $TMPDIR,
628 * making it a variable that overrides all temporary file storage locations. */
630 return tmp_dir_internal ( "/var/tmp" , ret
);
633 int tmp_dir ( const char ** ret
) {
635 /* Similar to var_tmp_dir() above, but returns the location for "smaller" temporary files, which is usually
636 * backed by an in-memory file system: /tmp. */
638 return tmp_dir_internal ( "/tmp" , ret
);
641 int unlink_or_warn ( const char * filename
) {
642 if ( unlink ( filename
) < 0 && errno
!= ENOENT
)
643 /* If the file doesn't exist and the fs simply was read-only (in which
644 * case unlink() returns EROFS even if the file doesn't exist), don't
646 if ( errno
!= EROFS
|| access ( filename
, F_OK
) >= 0 )
647 return log_error_errno ( errno
, "Failed to remove \" %s \" : %m" , filename
);
652 int inotify_add_watch_fd ( int fd
, int what
, uint32_t mask
) {
653 char path
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int ) + 1 ];
656 /* This is like inotify_add_watch(), except that the file to watch is not referenced by a path, but by an fd */
657 xsprintf ( path
, "/proc/self/fd/%i" , what
);
659 r
= inotify_add_watch ( fd
, path
, mask
);
666 static bool safe_transition ( const struct stat
* a
, const struct stat
* b
) {
667 /* Returns true if the transition from a to b is safe, i.e. that we never transition from unprivileged to
668 * privileged files or directories. Why bother? So that unprivileged code can't symlink to privileged files
669 * making us believe we read something safe even though it isn't safe in the specific context we open it in. */
671 if ( a
-> st_uid
== 0 ) /* Transitioning from privileged to unprivileged is always fine */
674 return a
-> st_uid
== b
-> st_uid
; /* Otherwise we need to stay within the same UID */
677 int chase_symlinks ( const char * path
, const char * original_root
, unsigned flags
, char ** ret
) {
678 _cleanup_free_
char * buffer
= NULL
, * done
= NULL
, * root
= NULL
;
679 _cleanup_close_
int fd
= - 1 ;
680 unsigned max_follow
= CHASE_SYMLINKS_MAX
; /* how many symlinks to follow before giving up and returning ELOOP */
681 struct stat previous_stat
;
688 /* Either the file may be missing, or we return an fd to the final object, but both make no sense */
689 if ( FLAGS_SET ( flags
, CHASE_NONEXISTENT
| CHASE_OPEN
))
692 if ( FLAGS_SET ( flags
, CHASE_STEP
| CHASE_OPEN
))
698 /* This is a lot like canonicalize_file_name(), but takes an additional "root" parameter, that allows following
699 * symlinks relative to a root directory, instead of the root of the host.
701 * Note that "root" primarily matters if we encounter an absolute symlink. It is also used when following
702 * relative symlinks to ensure they cannot be used to "escape" the root directory. The path parameter passed is
703 * assumed to be already prefixed by it, except if the CHASE_PREFIX_ROOT flag is set, in which case it is first
704 * prefixed accordingly.
706 * Algorithmically this operates on two path buffers: "done" are the components of the path we already
707 * processed and resolved symlinks, "." and ".." of. "todo" are the components of the path we still need to
708 * process. On each iteration, we move one component from "todo" to "done", processing it's special meaning
709 * each time. The "todo" path always starts with at least one slash, the "done" path always ends in no
710 * slash. We always keep an O_PATH fd to the component we are currently processing, thus keeping lookup races
713 * Suggested usage: whenever you want to canonicalize a path, use this function. Pass the absolute path you got
714 * as-is: fully qualified and relative to your host's root. Optionally, specify the root parameter to tell this
715 * function what to do when encountering a symlink with an absolute path as directory: prefix it by the
718 * There are three ways to invoke this function:
720 * 1. Without CHASE_STEP or CHASE_OPEN: in this case the path is resolved and the normalized path is returned
721 * in `ret`. The return value is < 0 on error. If CHASE_NONEXISTENT is also set 0 is returned if the file
722 * doesn't exist, > 0 otherwise. If CHASE_NONEXISTENT is not set >= 0 is returned if the destination was
723 * found, -ENOENT if it doesn't.
725 * 2. With CHASE_OPEN: in this case the destination is opened after chasing it as O_PATH and this file
726 * descriptor is returned as return value. This is useful to open files relative to some root
727 * directory. Note that the returned O_PATH file descriptors must be converted into a regular one (using
728 * fd_reopen() or such) before it can be used for reading/writing. CHASE_OPEN may not be combined with
731 * 3. With CHASE_STEP: in this case only a single step of the normalization is executed, i.e. only the first
732 * symlink or ".." component of the path is resolved, and the resulting path is returned. This is useful if
733 * a caller wants to trace the a path through the file system verbosely. Returns < 0 on error, > 0 if the
734 * path is fully normalized, and == 0 for each normalization step. This may be combined with
735 * CHASE_NONEXISTENT, in which case 1 is returned when a component is not found.
739 /* A root directory of "/" or "" is identical to none */
740 if ( empty_or_root ( original_root
))
741 original_root
= NULL
;
743 if (! original_root
&& ! ret
&& ( flags
& ( CHASE_NONEXISTENT
| CHASE_NO_AUTOFS
| CHASE_SAFE
| CHASE_OPEN
| CHASE_STEP
)) == CHASE_OPEN
) {
744 /* Shortcut the CHASE_OPEN case if the caller isn't interested in the actual path and has no root set
745 * and doesn't care about any of the other special features we provide either. */
746 r
= open ( path
, O_PATH
| O_CLOEXEC
|(( flags
& CHASE_NOFOLLOW
) ? O_NOFOLLOW
: 0 ));
754 r
= path_make_absolute_cwd ( original_root
, & root
);
758 if ( flags
& CHASE_PREFIX_ROOT
) {
760 /* We don't support relative paths in combination with a root directory */
761 if (! path_is_absolute ( path
))
764 path
= prefix_roota ( root
, path
);
768 r
= path_make_absolute_cwd ( path
, & buffer
);
772 fd
= open ( "/" , O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
776 if ( flags
& CHASE_SAFE
) {
777 if ( fstat ( fd
, & previous_stat
) < 0 )
783 _cleanup_free_
char * first
= NULL
;
784 _cleanup_close_
int child
= - 1 ;
788 /* Determine length of first component in the path */
789 n
= strspn ( todo
, "/" ); /* The slashes */
790 m
= n
+ strcspn ( todo
+ n
, "/" ); /* The entire length of the component */
792 /* Extract the first component. */
793 first
= strndup ( todo
, m
);
799 /* Empty? Then we reached the end. */
803 /* Just a single slash? Then we reached the end. */
804 if ( path_equal ( first
, "/" )) {
805 /* Preserve the trailing slash */
807 if ( flags
& CHASE_TRAIL_SLASH
)
808 if (! strextend (& done
, "/" , NULL
))
814 /* Just a dot? Then let's eat this up. */
815 if ( path_equal ( first
, "/." ))
818 /* Two dots? Then chop off the last bit of what we already found out. */
819 if ( path_equal ( first
, "/.." )) {
820 _cleanup_free_
char * parent
= NULL
;
821 _cleanup_close_
int fd_parent
= - 1 ;
823 /* If we already are at the top, then going up will not change anything. This is in-line with
824 * how the kernel handles this. */
825 if ( empty_or_root ( done
))
828 parent
= dirname_malloc ( done
);
832 /* Don't allow this to leave the root dir. */
834 path_startswith ( done
, root
) &&
835 ! path_startswith ( parent
, root
))
838 free_and_replace ( done
, parent
);
840 if ( flags
& CHASE_STEP
)
843 fd_parent
= openat ( fd
, ".." , O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
847 if ( flags
& CHASE_SAFE
) {
848 if ( fstat ( fd_parent
, & st
) < 0 )
851 if (! safe_transition (& previous_stat
, & st
))
858 fd
= TAKE_FD ( fd_parent
);
863 /* Otherwise let's see what this is. */
864 child
= openat ( fd
, first
+ n
, O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
867 if ( errno
== ENOENT
&&
868 ( flags
& CHASE_NONEXISTENT
) &&
869 ( isempty ( todo
) || path_is_normalized ( todo
))) {
871 /* If CHASE_NONEXISTENT is set, and the path does not exist, then that's OK, return
872 * what we got so far. But don't allow this if the remaining path contains "../ or "./"
873 * or something else weird. */
875 /* If done is "/", as first also contains slash at the head, then remove this redundant slash. */
876 if ( streq_ptr ( done
, "/" ))
879 if (! strextend (& done
, first
, todo
, NULL
))
889 if ( fstat ( child
, & st
) < 0 )
891 if (( flags
& CHASE_SAFE
) &&
892 ! safe_transition (& previous_stat
, & st
))
897 if (( flags
& CHASE_NO_AUTOFS
) &&
898 fd_is_fs_type ( child
, AUTOFS_SUPER_MAGIC
) > 0 )
901 if ( S_ISLNK ( st
. st_mode
) && !(( flags
& CHASE_NOFOLLOW
) && isempty ( todo
))) {
904 _cleanup_free_
char * destination
= NULL
;
906 /* This is a symlink, in this case read the destination. But let's make sure we don't follow
907 * symlinks without bounds. */
908 if (-- max_follow
<= 0 )
911 r
= readlinkat_malloc ( fd
, first
+ n
, & destination
);
914 if ( isempty ( destination
))
917 if ( path_is_absolute ( destination
)) {
919 /* An absolute destination. Start the loop from the beginning, but use the root
920 * directory as base. */
923 fd
= open ( root
?: "/" , O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
927 if ( flags
& CHASE_SAFE
) {
928 if ( fstat ( fd
, & st
) < 0 )
931 if (! safe_transition (& previous_stat
, & st
))
939 /* Note that we do not revalidate the root, we take it as is. */
948 /* Prefix what's left to do with what we just read, and start the loop again, but
949 * remain in the current directory. */
950 joined
= strjoin ( destination
, todo
);
952 joined
= strjoin ( "/" , destination
, todo
);
957 todo
= buffer
= joined
;
959 if ( flags
& CHASE_STEP
)
965 /* If this is not a symlink, then let's just add the name we read to what we already verified. */
967 done
= TAKE_PTR ( first
);
969 /* If done is "/", as first also contains slash at the head, then remove this redundant slash. */
970 if ( streq ( done
, "/" ))
973 if (! strextend (& done
, first
, NULL
))
977 /* And iterate again, but go one directory further down. */
983 /* Special case, turn the empty string into "/", to indicate the root directory. */
990 * ret
= TAKE_PTR ( done
);
992 if ( flags
& CHASE_OPEN
) {
993 /* Return the O_PATH fd we currently are looking to the caller. It can translate it to a proper fd by
994 * opening /proc/self/fd/xyz. */
1000 if ( flags
& CHASE_STEP
)
1009 c
= strjoin ( strempty ( done
), todo
);
1019 int chase_symlinks_and_open (
1022 unsigned chase_flags
,
1026 _cleanup_close_
int path_fd
= - 1 ;
1027 _cleanup_free_
char * p
= NULL
;
1030 if ( chase_flags
& CHASE_NONEXISTENT
)
1033 if ( empty_or_root ( root
) && ! ret_path
&& ( chase_flags
& ( CHASE_NO_AUTOFS
| CHASE_SAFE
)) == 0 ) {
1034 /* Shortcut this call if none of the special features of this call are requested */
1035 r
= open ( path
, open_flags
);
1042 path_fd
= chase_symlinks ( path
, root
, chase_flags
| CHASE_OPEN
, ret_path
? & p
: NULL
);
1046 r
= fd_reopen ( path_fd
, open_flags
);
1051 * ret_path
= TAKE_PTR ( p
);
1056 int chase_symlinks_and_opendir (
1059 unsigned chase_flags
,
1063 char procfs_path
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int )];
1064 _cleanup_close_
int path_fd
= - 1 ;
1065 _cleanup_free_
char * p
= NULL
;
1070 if ( chase_flags
& CHASE_NONEXISTENT
)
1073 if ( empty_or_root ( root
) && ! ret_path
&& ( chase_flags
& ( CHASE_NO_AUTOFS
| CHASE_SAFE
)) == 0 ) {
1074 /* Shortcut this call if none of the special features of this call are requested */
1083 path_fd
= chase_symlinks ( path
, root
, chase_flags
| CHASE_OPEN
, ret_path
? & p
: NULL
);
1087 xsprintf ( procfs_path
, "/proc/self/fd/%i" , path_fd
);
1088 d
= opendir ( procfs_path
);
1093 * ret_path
= TAKE_PTR ( p
);
1099 int chase_symlinks_and_stat (
1102 unsigned chase_flags
,
1104 struct stat
* ret_stat
) {
1106 _cleanup_close_
int path_fd
= - 1 ;
1107 _cleanup_free_
char * p
= NULL
;
1112 if ( chase_flags
& CHASE_NONEXISTENT
)
1115 if ( empty_or_root ( root
) && ! ret_path
&& ( chase_flags
& ( CHASE_NO_AUTOFS
| CHASE_SAFE
)) == 0 ) {
1116 /* Shortcut this call if none of the special features of this call are requested */
1117 if ( stat ( path
, ret_stat
) < 0 )
1123 path_fd
= chase_symlinks ( path
, root
, chase_flags
| CHASE_OPEN
, ret_path
? & p
: NULL
);
1127 if ( fstat ( path_fd
, ret_stat
) < 0 )
1131 * ret_path
= TAKE_PTR ( p
);
1133 if ( chase_flags
& CHASE_OPEN
)
1134 return TAKE_FD ( path_fd
);
1139 int access_fd ( int fd
, int mode
) {
1140 char p
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( fd
) + 1 ];
1143 /* Like access() but operates on an already open fd */
1145 xsprintf ( p
, "/proc/self/fd/%i" , fd
);
1146 r
= access ( p
, mode
);
1153 void unlink_tempfilep ( char (* p
)[]) {
1154 /* If the file is created with mkstemp(), it will (almost always)
1155 * change the suffix. Treat this as a sign that the file was
1156 * successfully created. We ignore both the rare case where the
1157 * original suffix is used and unlink failures. */
1158 if (! endswith (* p
, ".XXXXXX" ))
1159 ( void ) unlink_noerrno (* p
);
1162 int unlinkat_deallocate ( int fd
, const char * name
, int flags
) {
1163 _cleanup_close_
int truncate_fd
= - 1 ;
1167 /* Operates like unlinkat() but also deallocates the file contents if it is a regular file and there's no other
1168 * link to it. This is useful to ensure that other processes that might have the file open for reading won't be
1169 * able to keep the data pinned on disk forever. This call is particular useful whenever we execute clean-up
1170 * jobs ("vacuuming"), where we want to make sure the data is really gone and the disk space released and
1171 * returned to the free pool.
1173 * Deallocation is preferably done by FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE (đ) if supported, which means
1174 * the file won't change size. That's a good thing since we shouldn't needlessly trigger SIGBUS in other
1175 * programs that have mmap()ed the file. (The assumption here is that changing file contents to all zeroes
1176 * underneath those programs is the better choice than simply triggering SIGBUS in them which truncation does.)
1177 * However if hole punching is not implemented in the kernel or file system we'll fall back to normal file
1178 * truncation (đĒ), as our goal of deallocating the data space trumps our goal of being nice to readers (đ).
1180 * Note that we attempt deallocation, but failure to succeed with that is not considered fatal, as long as the
1181 * primary job â to delete the file â is accomplished. */
1183 if (( flags
& AT_REMOVEDIR
) == 0 ) {
1184 truncate_fd
= openat ( fd
, name
, O_WRONLY
| O_CLOEXEC
| O_NOCTTY
| O_NOFOLLOW
| O_NONBLOCK
);
1185 if ( truncate_fd
< 0 ) {
1187 /* If this failed because the file doesn't exist propagate the error right-away. Also,
1188 * AT_REMOVEDIR wasn't set, and we tried to open the file for writing, which means EISDIR is
1189 * returned when this is a directory but we are not supposed to delete those, hence propagate
1190 * the error right-away too. */
1191 if ( IN_SET ( errno
, ENOENT
, EISDIR
))
1194 if ( errno
!= ELOOP
) /* don't complain if this is a symlink */
1195 log_debug_errno ( errno
, "Failed to open file '%s' for deallocation, ignoring: %m" , name
);
1199 if ( unlinkat ( fd
, name
, flags
) < 0 )
1202 if ( truncate_fd
< 0 ) /* Don't have a file handle, can't do more âšī¸ */
1205 if ( fstat ( truncate_fd
, & st
) < 0 ) {
1206 log_debug_errno ( errno
, "Failed to stat file '%s' for deallocation, ignoring: %m" , name
);
1210 if (! S_ISREG ( st
. st_mode
) || st
. st_blocks
== 0 || st
. st_nlink
> 0 )
1213 /* If this is a regular file, it actually took up space on disk and there are no other links it's time to
1214 * punch-hole/truncate this to release the disk space. */
1216 bs
= MAX ( st
. st_blksize
, 512 );
1217 l
= DIV_ROUND_UP ( st
. st_size
, bs
) * bs
; /* Round up to next block size */
1219 if ( fallocate ( truncate_fd
, FALLOC_FL_PUNCH_HOLE
| FALLOC_FL_KEEP_SIZE
, 0 , l
) >= 0 )
1220 return 0 ; /* Successfully punched a hole! đ */
1222 /* Fall back to truncation */
1223 if ( ftruncate ( truncate_fd
, 0 ) < 0 ) {
1224 log_debug_errno ( errno
, "Failed to truncate file to 0, ignoring: %m" );
1231 int fsync_directory_of_file ( int fd
) {
1232 _cleanup_free_
char * path
= NULL
;
1233 _cleanup_close_
int dfd
= - 1 ;
1236 r
= fd_verify_regular ( fd
);
1240 r
= fd_get_path ( fd
, & path
);
1242 log_debug_errno ( r
, "Failed to query /proc/self/fd/%d%s: %m" ,
1244 r
== - EOPNOTSUPP
? ", ignoring" : "" );
1246 if ( r
== - EOPNOTSUPP
)
1247 /* If /proc is not available, we're most likely running in some
1248 * chroot environment, and syncing the directory is not very
1249 * important in that case. Let's just silently do nothing. */
1255 if (! path_is_absolute ( path
))
1258 dfd
= open_parent ( path
, O_CLOEXEC
, 0 );
1268 int fsync_path_at ( int at_fd
, const char * path
) {
1269 _cleanup_close_
int opened_fd
= - 1 ;
1272 if ( isempty ( path
)) {
1273 if ( at_fd
== AT_FDCWD
) {
1274 opened_fd
= open ( "." , O_RDONLY
| O_DIRECTORY
| O_CLOEXEC
);
1283 opened_fd
= openat ( at_fd
, path
, O_RDONLY
| O_CLOEXEC
);
1296 int open_parent ( const char * path
, int flags
, mode_t mode
) {
1297 _cleanup_free_
char * parent
= NULL
;
1302 if ( path_equal ( path
, "/" )) /* requesting the parent of the root dir is fishy, let's prohibit that */
1305 parent
= dirname_malloc ( path
);
1309 /* Let's insist on O_DIRECTORY since the parent of a file or directory is a directory. Except if we open an
1310 * O_TMPFILE file, because in that case we are actually create a regular file below the parent directory. */
1312 if (( flags
& O_PATH
) == O_PATH
)
1313 flags
|= O_DIRECTORY
;
1314 else if (( flags
& O_TMPFILE
) != O_TMPFILE
)
1315 flags
|= O_DIRECTORY
| O_RDONLY
;
1317 fd
= open ( parent
, flags
, mode
);