]>
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"
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 "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
) {
95 ret
= renameat2 ( olddirfd
, oldpath
, newdirfd
, newpath
, RENAME_NOREPLACE
);
99 /* renameat2() exists since Linux 3.15, btrfs added support for it later.
100 * If it is not implemented, fallback to another method. */
101 if (! IN_SET ( errno
, EINVAL
, ENOSYS
))
104 /* The link()/unlink() fallback does not work on directories. But
105 * renameat() without RENAME_NOREPLACE gives the same semantics on
106 * directories, except when newpath is an *empty* directory. This is
108 ret
= fstatat ( olddirfd
, oldpath
, & buf
, AT_SYMLINK_NOFOLLOW
);
109 if ( ret
>= 0 && S_ISDIR ( buf
. st_mode
)) {
110 ret
= renameat ( olddirfd
, oldpath
, newdirfd
, newpath
);
111 return ret
>= 0 ? 0 : - errno
;
114 /* If it is not a directory, use the link()/unlink() fallback. */
115 ret
= linkat ( olddirfd
, oldpath
, newdirfd
, newpath
, 0 );
119 ret
= unlinkat ( olddirfd
, oldpath
, 0 );
121 /* backup errno before the following unlinkat() alters it */
123 ( void ) unlinkat ( newdirfd
, newpath
, 0 );
131 int readlinkat_malloc ( int fd
, const char * p
, char ** ret
) {
146 n
= readlinkat ( fd
, p
, c
, l
- 1 );
153 if (( size_t ) n
< l
- 1 ) {
164 int readlink_malloc ( const char * p
, char ** ret
) {
165 return readlinkat_malloc ( AT_FDCWD
, p
, ret
);
168 int readlink_value ( const char * p
, char ** ret
) {
169 _cleanup_free_
char * link
= NULL
;
173 r
= readlink_malloc ( p
, & link
);
177 value
= basename ( link
);
181 value
= strdup ( value
);
190 int readlink_and_make_absolute ( const char * p
, char ** r
) {
191 _cleanup_free_
char * target
= NULL
;
198 j
= readlink_malloc ( p
, & target
);
202 k
= file_in_same_dir ( p
, target
);
210 int chmod_and_chown ( const char * path
, mode_t mode
, uid_t uid
, gid_t gid
) {
213 /* Under the assumption that we are running privileged we
214 * first change the access mode and only then hand out
215 * ownership to avoid a window where access is too open. */
217 if ( mode
!= MODE_INVALID
)
218 if ( chmod ( path
, mode
) < 0 )
221 if ( uid
!= UID_INVALID
|| gid
!= GID_INVALID
)
222 if ( chown ( path
, uid
, gid
) < 0 )
228 int fchmod_umask ( int fd
, mode_t m
) {
233 r
= fchmod ( fd
, m
& (~ u
)) < 0 ? - errno
: 0 ;
239 int fchmod_opath ( int fd
, mode_t m
) {
240 char procfs_path
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int )];
242 /* This function operates also on fd that might have been opened with
243 * O_PATH. Indeed fchmodat() doesn't have the AT_EMPTY_PATH flag like
244 * fchownat() does. */
246 xsprintf ( procfs_path
, "/proc/self/fd/%i" , fd
);
248 if ( chmod ( procfs_path
, m
) < 0 )
254 int fd_warn_permissions ( const char * path
, int fd
) {
257 if ( fstat ( fd
, & st
) < 0 )
260 if ( st
. st_mode
& 0111 )
261 log_warning ( "Configuration file %s is marked executable. Please remove executable permission bits. Proceeding anyway." , path
);
263 if ( st
. st_mode
& 0002 )
264 log_warning ( "Configuration file %s is marked world-writable. Please remove world writability permission bits. Proceeding anyway." , path
);
266 if ( getpid_cached () == 1 && ( st
. st_mode
& 0044 ) != 0044 )
267 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
);
272 int touch_file ( const char * path
, bool parents
, usec_t stamp
, uid_t uid
, gid_t gid
, mode_t mode
) {
273 char fdpath
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int )];
274 _cleanup_close_
int fd
= - 1 ;
279 /* Note that touch_file() does not follow symlinks: if invoked on an existing symlink, then it is the symlink
280 * itself which is updated, not its target
282 * Returns the first error we encounter, but tries to apply as much as possible. */
285 ( void ) mkdir_parents ( path
, 0755 );
287 /* Initially, we try to open the node with O_PATH, so that we get a reference to the node. This is useful in
288 * case the path refers to an existing device or socket node, as we can open it successfully in all cases, and
289 * won't trigger any driver magic or so. */
290 fd
= open ( path
, O_PATH
| O_CLOEXEC
| O_NOFOLLOW
);
295 /* if the node doesn't exist yet, we create it, but with O_EXCL, so that we only create a regular file
296 * here, and nothing else */
297 fd
= open ( path
, O_WRONLY
| O_CREAT
| O_EXCL
| O_CLOEXEC
, IN_SET ( mode
, 0 , MODE_INVALID
) ? 0644 : mode
);
302 /* Let's make a path from the fd, and operate on that. With this logic, we can adjust the access mode,
303 * ownership and time of the file node in all cases, even if the fd refers to an O_PATH object â which is
304 * something fchown(), fchmod(), futimensat() don't allow. */
305 xsprintf ( fdpath
, "/proc/self/fd/%i" , fd
);
307 if ( mode
!= MODE_INVALID
)
308 if ( chmod ( fdpath
, mode
) < 0 )
311 if ( uid_is_valid ( uid
) || gid_is_valid ( gid
))
312 if ( chown ( fdpath
, uid
, gid
) < 0 && ret
>= 0 )
315 if ( stamp
!= USEC_INFINITY
) {
316 struct timespec ts
[ 2 ];
318 timespec_store (& ts
[ 0 ], stamp
);
320 r
= utimensat ( AT_FDCWD
, fdpath
, ts
, 0 );
322 r
= utimensat ( AT_FDCWD
, fdpath
, NULL
, 0 );
323 if ( r
< 0 && ret
>= 0 )
329 int touch ( const char * path
) {
330 return touch_file ( path
, false , USEC_INFINITY
, UID_INVALID
, GID_INVALID
, MODE_INVALID
);
333 int symlink_idempotent ( const char * from
, const char * to
) {
339 if ( symlink ( from
, to
) < 0 ) {
340 _cleanup_free_
char * p
= NULL
;
345 r
= readlink_malloc ( to
, & p
);
346 if ( r
== - EINVAL
) /* Not a symlink? In that case return the original error we encountered: -EEXIST */
348 if ( r
< 0 ) /* Any other error? In that case propagate it as is */
351 if (! streq ( p
, from
)) /* Not the symlink we want it to be? In that case, propagate the original -EEXIST */
358 int symlink_atomic ( const char * from
, const char * to
) {
359 _cleanup_free_
char * t
= NULL
;
365 r
= tempfn_random ( to
, NULL
, & t
);
369 if ( symlink ( from
, t
) < 0 )
372 if ( rename ( t
, to
) < 0 ) {
380 int mknod_atomic ( const char * path
, mode_t mode
, dev_t dev
) {
381 _cleanup_free_
char * t
= NULL
;
386 r
= tempfn_random ( path
, NULL
, & t
);
390 if ( mknod ( t
, mode
, dev
) < 0 )
393 if ( rename ( t
, path
) < 0 ) {
401 int mkfifo_atomic ( const char * path
, mode_t mode
) {
402 _cleanup_free_
char * t
= NULL
;
407 r
= tempfn_random ( path
, NULL
, & t
);
411 if ( mkfifo ( t
, mode
) < 0 )
414 if ( rename ( t
, path
) < 0 ) {
422 int get_files_in_directory ( const char * path
, char *** list
) {
423 _cleanup_closedir_
DIR * d
= NULL
;
425 size_t bufsize
= 0 , n
= 0 ;
426 _cleanup_strv_free_
char ** l
= NULL
;
430 /* Returns all files in a directory in *list, and the number
431 * of files as return value. If list is NULL returns only the
438 FOREACH_DIRENT_ALL ( de
, d
, return - errno
) {
439 dirent_ensure_type ( d
, de
);
441 if (! dirent_is_file ( de
))
445 /* one extra slot is needed for the terminating NULL */
446 if (! GREEDY_REALLOC ( l
, bufsize
, n
+ 2 ))
449 l
[ n
] = strdup ( de
-> d_name
);
464 static int getenv_tmp_dir ( const char ** ret_path
) {
470 /* We use the same order of environment variables python uses in tempfile.gettempdir():
471 * https://docs.python.org/3/library/tempfile.html#tempfile.gettempdir */
472 FOREACH_STRING ( n
, "TMPDIR" , "TEMP" , "TMP" ) {
475 e
= secure_getenv ( n
);
478 if (! path_is_absolute ( e
)) {
482 if (! path_is_normalized ( e
)) {
499 /* Remember first error, to make this more debuggable */
511 static int tmp_dir_internal ( const char * def
, const char ** ret
) {
518 r
= getenv_tmp_dir (& e
);
524 k
= is_dir ( def
, true );
528 return r
< 0 ? r
: k
;
534 int var_tmp_dir ( const char ** ret
) {
536 /* Returns the location for "larger" temporary files, that is backed by physical storage if available, and thus
537 * even might survive a boot: /var/tmp. If $TMPDIR (or related environment variables) are set, its value is
538 * returned preferably however. Note that both this function and tmp_dir() below are affected by $TMPDIR,
539 * making it a variable that overrides all temporary file storage locations. */
541 return tmp_dir_internal ( "/var/tmp" , ret
);
544 int tmp_dir ( const char ** ret
) {
546 /* Similar to var_tmp_dir() above, but returns the location for "smaller" temporary files, which is usually
547 * backed by an in-memory file system: /tmp. */
549 return tmp_dir_internal ( "/tmp" , ret
);
552 int unlink_or_warn ( const char * filename
) {
553 if ( unlink ( filename
) < 0 && errno
!= ENOENT
)
554 /* If the file doesn't exist and the fs simply was read-only (in which
555 * case unlink() returns EROFS even if the file doesn't exist), don't
557 if ( errno
!= EROFS
|| access ( filename
, F_OK
) >= 0 )
558 return log_error_errno ( errno
, "Failed to remove \" %s \" : %m" , filename
);
563 int inotify_add_watch_fd ( int fd
, int what
, uint32_t mask
) {
564 char path
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int ) + 1 ];
567 /* This is like inotify_add_watch(), except that the file to watch is not referenced by a path, but by an fd */
568 xsprintf ( path
, "/proc/self/fd/%i" , what
);
570 r
= inotify_add_watch ( fd
, path
, mask
);
577 static bool safe_transition ( const struct stat
* a
, const struct stat
* b
) {
578 /* Returns true if the transition from a to b is safe, i.e. that we never transition from unprivileged to
579 * privileged files or directories. Why bother? So that unprivileged code can't symlink to privileged files
580 * making us believe we read something safe even though it isn't safe in the specific context we open it in. */
582 if ( a
-> st_uid
== 0 ) /* Transitioning from privileged to unprivileged is always fine */
585 return a
-> st_uid
== b
-> st_uid
; /* Otherwise we need to stay within the same UID */
588 int chase_symlinks ( const char * path
, const char * original_root
, unsigned flags
, char ** ret
) {
589 _cleanup_free_
char * buffer
= NULL
, * done
= NULL
, * root
= NULL
;
590 _cleanup_close_
int fd
= - 1 ;
591 unsigned max_follow
= CHASE_SYMLINKS_MAX
; /* how many symlinks to follow before giving up and returning ELOOP */
592 struct stat previous_stat
;
599 /* Either the file may be missing, or we return an fd to the final object, but both make no sense */
600 if ( FLAGS_SET ( flags
, CHASE_NONEXISTENT
| CHASE_OPEN
))
603 if ( FLAGS_SET ( flags
, CHASE_STEP
| CHASE_OPEN
))
609 /* This is a lot like canonicalize_file_name(), but takes an additional "root" parameter, that allows following
610 * symlinks relative to a root directory, instead of the root of the host.
612 * Note that "root" primarily matters if we encounter an absolute symlink. It is also used when following
613 * relative symlinks to ensure they cannot be used to "escape" the root directory. The path parameter passed is
614 * assumed to be already prefixed by it, except if the CHASE_PREFIX_ROOT flag is set, in which case it is first
615 * prefixed accordingly.
617 * Algorithmically this operates on two path buffers: "done" are the components of the path we already
618 * processed and resolved symlinks, "." and ".." of. "todo" are the components of the path we still need to
619 * process. On each iteration, we move one component from "todo" to "done", processing it's special meaning
620 * each time. The "todo" path always starts with at least one slash, the "done" path always ends in no
621 * slash. We always keep an O_PATH fd to the component we are currently processing, thus keeping lookup races
624 * Suggested usage: whenever you want to canonicalize a path, use this function. Pass the absolute path you got
625 * as-is: fully qualified and relative to your host's root. Optionally, specify the root parameter to tell this
626 * function what to do when encountering a symlink with an absolute path as directory: prefix it by the
629 * There are three ways to invoke this function:
631 * 1. Without CHASE_STEP or CHASE_OPEN: in this case the path is resolved and the normalized path is returned
632 * in `ret`. The return value is < 0 on error. If CHASE_NONEXISTENT is also set 0 is returned if the file
633 * doesn't exist, > 0 otherwise. If CHASE_NONEXISTENT is not set >= 0 is returned if the destination was
634 * found, -ENOENT if it doesn't.
636 * 2. With CHASE_OPEN: in this case the destination is opened after chasing it as O_PATH and this file
637 * descriptor is returned as return value. This is useful to open files relative to some root
638 * directory. Note that the returned O_PATH file descriptors must be converted into a regular one (using
639 * fd_reopen() or such) before it can be used for reading/writing. CHASE_OPEN may not be combined with
642 * 3. With CHASE_STEP: in this case only a single step of the normalization is executed, i.e. only the first
643 * symlink or ".." component of the path is resolved, and the resulting path is returned. This is useful if
644 * a caller wants to trace the a path through the file system verbosely. Returns < 0 on error, > 0 if the
645 * path is fully normalized, and == 0 for each normalization step. This may be combined with
646 * CHASE_NONEXISTENT, in which case 1 is returned when a component is not found.
650 /* A root directory of "/" or "" is identical to none */
651 if ( empty_or_root ( original_root
))
652 original_root
= NULL
;
654 if (! original_root
&& ! ret
&& ( flags
& ( CHASE_NONEXISTENT
| CHASE_NO_AUTOFS
| CHASE_SAFE
| CHASE_OPEN
| CHASE_STEP
)) == CHASE_OPEN
) {
655 /* Shortcut the CHASE_OPEN case if the caller isn't interested in the actual path and has no root set
656 * and doesn't care about any of the other special features we provide either. */
657 r
= open ( path
, O_PATH
| O_CLOEXEC
);
665 r
= path_make_absolute_cwd ( original_root
, & root
);
669 if ( flags
& CHASE_PREFIX_ROOT
) {
671 /* We don't support relative paths in combination with a root directory */
672 if (! path_is_absolute ( path
))
675 path
= prefix_roota ( root
, path
);
679 r
= path_make_absolute_cwd ( path
, & buffer
);
683 fd
= open ( "/" , O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
687 if ( flags
& CHASE_SAFE
) {
688 if ( fstat ( fd
, & previous_stat
) < 0 )
694 _cleanup_free_
char * first
= NULL
;
695 _cleanup_close_
int child
= - 1 ;
699 /* Determine length of first component in the path */
700 n
= strspn ( todo
, "/" ); /* The slashes */
701 m
= n
+ strcspn ( todo
+ n
, "/" ); /* The entire length of the component */
703 /* Extract the first component. */
704 first
= strndup ( todo
, m
);
710 /* Empty? Then we reached the end. */
714 /* Just a single slash? Then we reached the end. */
715 if ( path_equal ( first
, "/" )) {
716 /* Preserve the trailing slash */
718 if ( flags
& CHASE_TRAIL_SLASH
)
719 if (! strextend (& done
, "/" , NULL
))
725 /* Just a dot? Then let's eat this up. */
726 if ( path_equal ( first
, "/." ))
729 /* Two dots? Then chop off the last bit of what we already found out. */
730 if ( path_equal ( first
, "/.." )) {
731 _cleanup_free_
char * parent
= NULL
;
732 _cleanup_close_
int fd_parent
= - 1 ;
734 /* If we already are at the top, then going up will not change anything. This is in-line with
735 * how the kernel handles this. */
736 if ( empty_or_root ( done
))
739 parent
= dirname_malloc ( done
);
743 /* Don't allow this to leave the root dir. */
745 path_startswith ( done
, root
) &&
746 ! path_startswith ( parent
, root
))
749 free_and_replace ( done
, parent
);
751 if ( flags
& CHASE_STEP
)
754 fd_parent
= openat ( fd
, ".." , O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
758 if ( flags
& CHASE_SAFE
) {
759 if ( fstat ( fd_parent
, & st
) < 0 )
762 if (! safe_transition (& previous_stat
, & st
))
769 fd
= TAKE_FD ( fd_parent
);
774 /* Otherwise let's see what this is. */
775 child
= openat ( fd
, first
+ n
, O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
778 if ( errno
== ENOENT
&&
779 ( flags
& CHASE_NONEXISTENT
) &&
780 ( isempty ( todo
) || path_is_normalized ( todo
))) {
782 /* If CHASE_NONEXISTENT is set, and the path does not exist, then that's OK, return
783 * what we got so far. But don't allow this if the remaining path contains "../ or "./"
784 * or something else weird. */
786 /* If done is "/", as first also contains slash at the head, then remove this redundant slash. */
787 if ( streq_ptr ( done
, "/" ))
790 if (! strextend (& done
, first
, todo
, NULL
))
800 if ( fstat ( child
, & st
) < 0 )
802 if (( flags
& CHASE_SAFE
) &&
803 ! safe_transition (& previous_stat
, & st
))
808 if (( flags
& CHASE_NO_AUTOFS
) &&
809 fd_is_fs_type ( child
, AUTOFS_SUPER_MAGIC
) > 0 )
812 if ( S_ISLNK ( st
. st_mode
)) {
815 _cleanup_free_
char * destination
= NULL
;
817 /* This is a symlink, in this case read the destination. But let's make sure we don't follow
818 * symlinks without bounds. */
819 if (-- max_follow
<= 0 )
822 r
= readlinkat_malloc ( fd
, first
+ n
, & destination
);
825 if ( isempty ( destination
))
828 if ( path_is_absolute ( destination
)) {
830 /* An absolute destination. Start the loop from the beginning, but use the root
831 * directory as base. */
834 fd
= open ( root
?: "/" , O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
838 if ( flags
& CHASE_SAFE
) {
839 if ( fstat ( fd
, & st
) < 0 )
842 if (! safe_transition (& previous_stat
, & st
))
850 /* Note that we do not revalidate the root, we take it as is. */
859 /* Prefix what's left to do with what we just read, and start the loop again, but
860 * remain in the current directory. */
861 joined
= strjoin ( destination
, todo
);
863 joined
= strjoin ( "/" , destination
, todo
);
868 todo
= buffer
= joined
;
870 if ( flags
& CHASE_STEP
)
876 /* If this is not a symlink, then let's just add the name we read to what we already verified. */
878 done
= TAKE_PTR ( first
);
880 /* If done is "/", as first also contains slash at the head, then remove this redundant slash. */
881 if ( streq ( done
, "/" ))
884 if (! strextend (& done
, first
, NULL
))
888 /* And iterate again, but go one directory further down. */
894 /* Special case, turn the empty string into "/", to indicate the root directory. */
901 * ret
= TAKE_PTR ( done
);
903 if ( flags
& CHASE_OPEN
) {
904 /* Return the O_PATH fd we currently are looking to the caller. It can translate it to a proper fd by
905 * opening /proc/self/fd/xyz. */
911 if ( flags
& CHASE_STEP
)
920 c
= strjoin ( strempty ( done
), todo
);
930 int chase_symlinks_and_open (
933 unsigned chase_flags
,
937 _cleanup_close_
int path_fd
= - 1 ;
938 _cleanup_free_
char * p
= NULL
;
941 if ( chase_flags
& CHASE_NONEXISTENT
)
944 if ( empty_or_root ( root
) && ! ret_path
&& ( chase_flags
& ( CHASE_NO_AUTOFS
| CHASE_SAFE
)) == 0 ) {
945 /* Shortcut this call if none of the special features of this call are requested */
946 r
= open ( path
, open_flags
);
953 path_fd
= chase_symlinks ( path
, root
, chase_flags
| CHASE_OPEN
, ret_path
? & p
: NULL
);
957 r
= fd_reopen ( path_fd
, open_flags
);
962 * ret_path
= TAKE_PTR ( p
);
967 int chase_symlinks_and_opendir (
970 unsigned chase_flags
,
974 char procfs_path
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int )];
975 _cleanup_close_
int path_fd
= - 1 ;
976 _cleanup_free_
char * p
= NULL
;
981 if ( chase_flags
& CHASE_NONEXISTENT
)
984 if ( empty_or_root ( root
) && ! ret_path
&& ( chase_flags
& ( CHASE_NO_AUTOFS
| CHASE_SAFE
)) == 0 ) {
985 /* Shortcut this call if none of the special features of this call are requested */
994 path_fd
= chase_symlinks ( path
, root
, chase_flags
| CHASE_OPEN
, ret_path
? & p
: NULL
);
998 xsprintf ( procfs_path
, "/proc/self/fd/%i" , path_fd
);
999 d
= opendir ( procfs_path
);
1004 * ret_path
= TAKE_PTR ( p
);
1010 int chase_symlinks_and_stat (
1013 unsigned chase_flags
,
1015 struct stat
* ret_stat
) {
1017 _cleanup_close_
int path_fd
= - 1 ;
1018 _cleanup_free_
char * p
= NULL
;
1023 if ( chase_flags
& CHASE_NONEXISTENT
)
1026 if ( empty_or_root ( root
) && ! ret_path
&& ( chase_flags
& ( CHASE_NO_AUTOFS
| CHASE_SAFE
)) == 0 ) {
1027 /* Shortcut this call if none of the special features of this call are requested */
1028 if ( stat ( path
, ret_stat
) < 0 )
1034 path_fd
= chase_symlinks ( path
, root
, chase_flags
| CHASE_OPEN
, ret_path
? & p
: NULL
);
1038 if ( fstat ( path_fd
, ret_stat
) < 0 )
1042 * ret_path
= TAKE_PTR ( p
);
1044 if ( chase_flags
& CHASE_OPEN
)
1045 return TAKE_FD ( path_fd
);
1050 int access_fd ( int fd
, int mode
) {
1051 char p
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( fd
) + 1 ];
1054 /* Like access() but operates on an already open fd */
1056 xsprintf ( p
, "/proc/self/fd/%i" , fd
);
1057 r
= access ( p
, mode
);
1064 void unlink_tempfilep ( char (* p
)[]) {
1065 /* If the file is created with mkstemp(), it will (almost always)
1066 * change the suffix. Treat this as a sign that the file was
1067 * successfully created. We ignore both the rare case where the
1068 * original suffix is used and unlink failures. */
1069 if (! endswith (* p
, ".XXXXXX" ))
1070 ( void ) unlink_noerrno (* p
);
1073 int unlinkat_deallocate ( int fd
, const char * name
, int flags
) {
1074 _cleanup_close_
int truncate_fd
= - 1 ;
1078 /* Operates like unlinkat() but also deallocates the file contents if it is a regular file and there's no other
1079 * link to it. This is useful to ensure that other processes that might have the file open for reading won't be
1080 * able to keep the data pinned on disk forever. This call is particular useful whenever we execute clean-up
1081 * jobs ("vacuuming"), where we want to make sure the data is really gone and the disk space released and
1082 * returned to the free pool.
1084 * Deallocation is preferably done by FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE (đ) if supported, which means
1085 * the file won't change size. That's a good thing since we shouldn't needlessly trigger SIGBUS in other
1086 * programs that have mmap()ed the file. (The assumption here is that changing file contents to all zeroes
1087 * underneath those programs is the better choice than simply triggering SIGBUS in them which truncation does.)
1088 * However if hole punching is not implemented in the kernel or file system we'll fall back to normal file
1089 * truncation (đĒ), as our goal of deallocating the data space trumps our goal of being nice to readers (đ).
1091 * Note that we attempt deallocation, but failure to succeed with that is not considered fatal, as long as the
1092 * primary job â to delete the file â is accomplished. */
1094 if (( flags
& AT_REMOVEDIR
) == 0 ) {
1095 truncate_fd
= openat ( fd
, name
, O_WRONLY
| O_CLOEXEC
| O_NOCTTY
| O_NOFOLLOW
| O_NONBLOCK
);
1096 if ( truncate_fd
< 0 ) {
1098 /* If this failed because the file doesn't exist propagate the error right-away. Also,
1099 * AT_REMOVEDIR wasn't set, and we tried to open the file for writing, which means EISDIR is
1100 * returned when this is a directory but we are not supposed to delete those, hence propagate
1101 * the error right-away too. */
1102 if ( IN_SET ( errno
, ENOENT
, EISDIR
))
1105 if ( errno
!= ELOOP
) /* don't complain if this is a symlink */
1106 log_debug_errno ( errno
, "Failed to open file '%s' for deallocation, ignoring: %m" , name
);
1110 if ( unlinkat ( fd
, name
, flags
) < 0 )
1113 if ( truncate_fd
< 0 ) /* Don't have a file handle, can't do more âšī¸ */
1116 if ( fstat ( truncate_fd
, & st
) < 0 ) {
1117 log_debug_errno ( errno
, "Failed to stat file '%s' for deallocation, ignoring." , name
);
1121 if (! S_ISREG ( st
. st_mode
) || st
. st_blocks
== 0 || st
. st_nlink
> 0 )
1124 /* If this is a regular file, it actually took up space on disk and there are no other links it's time to
1125 * punch-hole/truncate this to release the disk space. */
1127 bs
= MAX ( st
. st_blksize
, 512 );
1128 l
= DIV_ROUND_UP ( st
. st_size
, bs
) * bs
; /* Round up to next block size */
1130 if ( fallocate ( truncate_fd
, FALLOC_FL_PUNCH_HOLE
| FALLOC_FL_KEEP_SIZE
, 0 , l
) >= 0 )
1131 return 0 ; /* Successfully punched a hole! đ */
1133 /* Fall back to truncation */
1134 if ( ftruncate ( truncate_fd
, 0 ) < 0 ) {
1135 log_debug_errno ( errno
, "Failed to truncate file to 0, ignoring: %m" );
1142 int fsync_directory_of_file ( int fd
) {
1143 _cleanup_free_
char * path
= NULL
, * dn
= NULL
;
1144 _cleanup_close_
int dfd
= - 1 ;
1147 r
= fd_verify_regular ( fd
);
1151 r
= fd_get_path ( fd
, & path
);
1153 log_debug_errno ( r
, "Failed to query /proc/self/fd/%d%s: %m" ,
1155 r
== - EOPNOTSUPP
? ", ignoring" : "" );
1157 if ( r
== - EOPNOTSUPP
)
1158 /* If /proc is not available, we're most likely running in some
1159 * chroot environment, and syncing the directory is not very
1160 * important in that case. Let's just silently do nothing. */
1166 if (! path_is_absolute ( path
))
1169 dn
= dirname_malloc ( path
);
1173 dfd
= open ( dn
, O_RDONLY
| O_CLOEXEC
| O_DIRECTORY
);