]>
git.ipfire.org Git - thirdparty/systemd.git/blob - src/basic/fs-util.c
1 /* SPDX-License-Identifier: LGPL-2.1+ */
3 This file is part of systemd.
5 Copyright 2010 Lennart Poettering
14 #include <linux/magic.h>
18 #include "alloc-util.h"
19 #include "dirent-util.h"
27 #include "parse-util.h"
28 #include "path-util.h"
29 #include "process-util.h"
30 #include "stat-util.h"
31 #include "stdio-util.h"
32 #include "string-util.h"
34 #include "time-util.h"
35 #include "user-util.h"
38 int unlink_noerrno ( const char * path
) {
49 int rmdir_parents ( const char * path
, const char * stop
) {
58 /* Skip trailing slashes */
59 while ( l
> 0 && path
[ l
- 1 ] == '/' )
65 /* Skip last component */
66 while ( l
> 0 && path
[ l
- 1 ] != '/' )
69 /* Skip trailing slashes */
70 while ( l
> 0 && path
[ l
- 1 ] == '/' )
80 if ( path_startswith ( stop
, t
)) {
96 int rename_noreplace ( int olddirfd
, const char * oldpath
, int newdirfd
, const char * newpath
) {
100 ret
= renameat2 ( olddirfd
, oldpath
, newdirfd
, newpath
, RENAME_NOREPLACE
);
104 /* renameat2() exists since Linux 3.15, btrfs added support for it later.
105 * If it is not implemented, fallback to another method. */
106 if (! IN_SET ( errno
, EINVAL
, ENOSYS
))
109 /* The link()/unlink() fallback does not work on directories. But
110 * renameat() without RENAME_NOREPLACE gives the same semantics on
111 * directories, except when newpath is an *empty* directory. This is
113 ret
= fstatat ( olddirfd
, oldpath
, & buf
, AT_SYMLINK_NOFOLLOW
);
114 if ( ret
>= 0 && S_ISDIR ( buf
. st_mode
)) {
115 ret
= renameat ( olddirfd
, oldpath
, newdirfd
, newpath
);
116 return ret
>= 0 ? 0 : - errno
;
119 /* If it is not a directory, use the link()/unlink() fallback. */
120 ret
= linkat ( olddirfd
, oldpath
, newdirfd
, newpath
, 0 );
124 ret
= unlinkat ( olddirfd
, oldpath
, 0 );
126 /* backup errno before the following unlinkat() alters it */
128 ( void ) unlinkat ( newdirfd
, newpath
, 0 );
136 int readlinkat_malloc ( int fd
, const char * p
, char ** ret
) {
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
) {
218 /* Under the assumption that we are running privileged we
219 * first change the access mode and only then hand out
220 * ownership to avoid a window where access is too open. */
222 if ( mode
!= MODE_INVALID
)
223 if ( chmod ( path
, mode
) < 0 )
226 if ( uid
!= UID_INVALID
|| gid
!= GID_INVALID
)
227 if ( chown ( path
, uid
, gid
) < 0 )
233 int fchmod_umask ( int fd
, mode_t m
) {
238 r
= fchmod ( fd
, m
& (~ u
)) < 0 ? - errno
: 0 ;
244 int fchmod_opath ( int fd
, mode_t m
) {
245 char procfs_path
[ strlen ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int )];
247 /* This function operates also on fd that might have been opened with
248 * O_PATH. Indeed fchmodat() doesn't have the AT_EMPTY_PATH flag like
249 * fchownat() does. */
251 xsprintf ( procfs_path
, "/proc/self/fd/%i" , fd
);
253 if ( chmod ( procfs_path
, m
) < 0 )
259 int fd_warn_permissions ( const char * path
, int fd
) {
262 if ( fstat ( fd
, & st
) < 0 )
265 if ( st
. st_mode
& 0111 )
266 log_warning ( "Configuration file %s is marked executable. Please remove executable permission bits. Proceeding anyway." , path
);
268 if ( st
. st_mode
& 0002 )
269 log_warning ( "Configuration file %s is marked world-writable. Please remove world writability permission bits. Proceeding anyway." , path
);
271 if ( getpid_cached () == 1 && ( st
. st_mode
& 0044 ) != 0044 )
272 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
);
277 int touch_file ( const char * path
, bool parents
, usec_t stamp
, uid_t uid
, gid_t gid
, mode_t mode
) {
278 char fdpath
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int )];
279 _cleanup_close_
int fd
= - 1 ;
284 /* Note that touch_file() does not follow symlinks: if invoked on an existing symlink, then it is the symlink
285 * itself which is updated, not its target
287 * Returns the first error we encounter, but tries to apply as much as possible. */
290 ( void ) mkdir_parents ( path
, 0755 );
292 /* Initially, we try to open the node with O_PATH, so that we get a reference to the node. This is useful in
293 * case the path refers to an existing device or socket node, as we can open it successfully in all cases, and
294 * won't trigger any driver magic or so. */
295 fd
= open ( path
, O_PATH
| O_CLOEXEC
| O_NOFOLLOW
);
300 /* if the node doesn't exist yet, we create it, but with O_EXCL, so that we only create a regular file
301 * here, and nothing else */
302 fd
= open ( path
, O_WRONLY
| O_CREAT
| O_EXCL
| O_CLOEXEC
, IN_SET ( mode
, 0 , MODE_INVALID
) ? 0644 : mode
);
307 /* Let's make a path from the fd, and operate on that. With this logic, we can adjust the access mode,
308 * ownership and time of the file node in all cases, even if the fd refers to an O_PATH object â which is
309 * something fchown(), fchmod(), futimensat() don't allow. */
310 xsprintf ( fdpath
, "/proc/self/fd/%i" , fd
);
312 if ( mode
!= MODE_INVALID
)
313 if ( chmod ( fdpath
, mode
) < 0 )
316 if ( uid_is_valid ( uid
) || gid_is_valid ( gid
))
317 if ( chown ( fdpath
, uid
, gid
) < 0 && ret
>= 0 )
320 if ( stamp
!= USEC_INFINITY
) {
321 struct timespec ts
[ 2 ];
323 timespec_store (& ts
[ 0 ], stamp
);
325 r
= utimensat ( AT_FDCWD
, fdpath
, ts
, 0 );
327 r
= utimensat ( AT_FDCWD
, fdpath
, NULL
, 0 );
328 if ( r
< 0 && ret
>= 0 )
334 int touch ( const char * path
) {
335 return touch_file ( path
, false , USEC_INFINITY
, UID_INVALID
, GID_INVALID
, MODE_INVALID
);
338 int symlink_idempotent ( const char * from
, const char * to
) {
344 if ( symlink ( from
, to
) < 0 ) {
345 _cleanup_free_
char * p
= NULL
;
350 r
= readlink_malloc ( to
, & p
);
351 if ( r
== - EINVAL
) /* Not a symlink? In that case return the original error we encountered: -EEXIST */
353 if ( r
< 0 ) /* Any other error? In that case propagate it as is */
356 if (! streq ( p
, from
)) /* Not the symlink we want it to be? In that case, propagate the original -EEXIST */
363 int symlink_atomic ( const char * from
, const char * to
) {
364 _cleanup_free_
char * t
= NULL
;
370 r
= tempfn_random ( to
, NULL
, & t
);
374 if ( symlink ( from
, t
) < 0 )
377 if ( rename ( t
, to
) < 0 ) {
385 int mknod_atomic ( const char * path
, mode_t mode
, dev_t dev
) {
386 _cleanup_free_
char * t
= NULL
;
391 r
= tempfn_random ( path
, NULL
, & t
);
395 if ( mknod ( t
, mode
, dev
) < 0 )
398 if ( rename ( t
, path
) < 0 ) {
406 int mkfifo_atomic ( const char * path
, mode_t mode
) {
407 _cleanup_free_
char * t
= NULL
;
412 r
= tempfn_random ( path
, NULL
, & t
);
416 if ( mkfifo ( t
, mode
) < 0 )
419 if ( rename ( t
, path
) < 0 ) {
427 int get_files_in_directory ( const char * path
, char *** list
) {
428 _cleanup_closedir_
DIR * d
= NULL
;
430 size_t bufsize
= 0 , n
= 0 ;
431 _cleanup_strv_free_
char ** l
= NULL
;
435 /* Returns all files in a directory in *list, and the number
436 * of files as return value. If list is NULL returns only the
443 FOREACH_DIRENT_ALL ( de
, d
, return - errno
) {
444 dirent_ensure_type ( d
, de
);
446 if (! dirent_is_file ( de
))
450 /* one extra slot is needed for the terminating NULL */
451 if (! GREEDY_REALLOC ( l
, bufsize
, n
+ 2 ))
454 l
[ n
] = strdup ( de
-> d_name
);
469 static int getenv_tmp_dir ( const char ** ret_path
) {
475 /* We use the same order of environment variables python uses in tempfile.gettempdir():
476 * https://docs.python.org/3/library/tempfile.html#tempfile.gettempdir */
477 FOREACH_STRING ( n
, "TMPDIR" , "TEMP" , "TMP" ) {
480 e
= secure_getenv ( n
);
483 if (! path_is_absolute ( e
)) {
487 if (! path_is_normalized ( e
)) {
504 /* Remember first error, to make this more debuggable */
516 static int tmp_dir_internal ( const char * def
, const char ** ret
) {
523 r
= getenv_tmp_dir (& e
);
529 k
= is_dir ( def
, true );
533 return r
< 0 ? r
: k
;
539 int var_tmp_dir ( const char ** ret
) {
541 /* Returns the location for "larger" temporary files, that is backed by physical storage if available, and thus
542 * even might survive a boot: /var/tmp. If $TMPDIR (or related environment variables) are set, its value is
543 * returned preferably however. Note that both this function and tmp_dir() below are affected by $TMPDIR,
544 * making it a variable that overrides all temporary file storage locations. */
546 return tmp_dir_internal ( "/var/tmp" , ret
);
549 int tmp_dir ( const char ** ret
) {
551 /* Similar to var_tmp_dir() above, but returns the location for "smaller" temporary files, which is usually
552 * backed by an in-memory file system: /tmp. */
554 return tmp_dir_internal ( "/tmp" , ret
);
557 int unlink_or_warn ( const char * filename
) {
558 if ( unlink ( filename
) < 0 && errno
!= ENOENT
)
559 /* If the file doesn't exist and the fs simply was read-only (in which
560 * case unlink() returns EROFS even if the file doesn't exist), don't
562 if ( errno
!= EROFS
|| access ( filename
, F_OK
) >= 0 )
563 return log_error_errno ( errno
, "Failed to remove \" %s \" : %m" , filename
);
568 int inotify_add_watch_fd ( int fd
, int what
, uint32_t mask
) {
569 char path
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int ) + 1 ];
572 /* This is like inotify_add_watch(), except that the file to watch is not referenced by a path, but by an fd */
573 xsprintf ( path
, "/proc/self/fd/%i" , what
);
575 r
= inotify_add_watch ( fd
, path
, mask
);
582 static bool safe_transition ( const struct stat
* a
, const struct stat
* b
) {
583 /* Returns true if the transition from a to b is safe, i.e. that we never transition from unprivileged to
584 * privileged files or directories. Why bother? So that unprivileged code can't symlink to privileged files
585 * making us believe we read something safe even though it isn't safe in the specific context we open it in. */
587 if ( a
-> st_uid
== 0 ) /* Transitioning from privileged to unprivileged is always fine */
590 return a
-> st_uid
== b
-> st_uid
; /* Otherwise we need to stay within the same UID */
593 int chase_symlinks ( const char * path
, const char * original_root
, unsigned flags
, char ** ret
) {
594 _cleanup_free_
char * buffer
= NULL
, * done
= NULL
, * root
= NULL
;
595 _cleanup_close_
int fd
= - 1 ;
596 unsigned max_follow
= CHASE_SYMLINKS_MAX
; /* how many symlinks to follow before giving up and returning ELOOP */
597 struct stat previous_stat
;
604 /* Either the file may be missing, or we return an fd to the final object, but both make no sense */
605 if (( flags
& ( CHASE_NONEXISTENT
| CHASE_OPEN
)) == ( CHASE_NONEXISTENT
| CHASE_OPEN
))
608 if (( flags
& ( CHASE_STEP
| CHASE_OPEN
)) == ( CHASE_STEP
| CHASE_OPEN
))
614 /* This is a lot like canonicalize_file_name(), but takes an additional "root" parameter, that allows following
615 * symlinks relative to a root directory, instead of the root of the host.
617 * Note that "root" primarily matters if we encounter an absolute symlink. It is also used when following
618 * relative symlinks to ensure they cannot be used to "escape" the root directory. The path parameter passed is
619 * assumed to be already prefixed by it, except if the CHASE_PREFIX_ROOT flag is set, in which case it is first
620 * prefixed accordingly.
622 * Algorithmically this operates on two path buffers: "done" are the components of the path we already
623 * processed and resolved symlinks, "." and ".." of. "todo" are the components of the path we still need to
624 * process. On each iteration, we move one component from "todo" to "done", processing it's special meaning
625 * each time. The "todo" path always starts with at least one slash, the "done" path always ends in no
626 * slash. We always keep an O_PATH fd to the component we are currently processing, thus keeping lookup races
629 * Suggested usage: whenever you want to canonicalize a path, use this function. Pass the absolute path you got
630 * as-is: fully qualified and relative to your host's root. Optionally, specify the root parameter to tell this
631 * function what to do when encountering a symlink with an absolute path as directory: prefix it by the
634 * There are three ways to invoke this function:
636 * 1. Without CHASE_STEP or CHASE_OPEN: in this case the path is resolved and the normalized path is returned
637 * in `ret`. The return value is < 0 on error. If CHASE_NONEXISTENT is also set 0 is returned if the file
638 * doesn't exist, > 0 otherwise. If CHASE_NONEXISTENT is not set >= 0 is returned if the destination was
639 * found, -ENOENT if it doesn't.
641 * 2. With CHASE_OPEN: in this case the destination is opened after chasing it as O_PATH and this file
642 * descriptor is returned as return value. This is useful to open files relative to some root
643 * directory. Note that the returned O_PATH file descriptors must be converted into a regular one (using
644 * fd_reopen() or such) before it can be used for reading/writing. CHASE_OPEN may not be combined with
647 * 3. With CHASE_STEP: in this case only a single step of the normalization is executed, i.e. only the first
648 * symlink or ".." component of the path is resolved, and the resulting path is returned. This is useful if
649 * a caller wants to trace the a path through the file system verbosely. Returns < 0 on error, > 0 if the
650 * path is fully normalized, and == 0 for each normalization step. This may be combined with
651 * CHASE_NONEXISTENT, in which case 1 is returned when a component is not found.
655 /* A root directory of "/" or "" is identical to none */
656 if ( empty_or_root ( original_root
))
657 original_root
= NULL
;
659 if (! original_root
&& ! ret
&& ( flags
& ( CHASE_NONEXISTENT
| CHASE_NO_AUTOFS
| CHASE_SAFE
| CHASE_OPEN
| CHASE_STEP
)) == CHASE_OPEN
) {
660 /* Shortcut the CHASE_OPEN case if the caller isn't interested in the actual path and has no root set
661 * and doesn't care about any of the other special features we provide either. */
662 r
= open ( path
, O_PATH
| O_CLOEXEC
);
670 r
= path_make_absolute_cwd ( original_root
, & root
);
674 if ( flags
& CHASE_PREFIX_ROOT
) {
676 /* We don't support relative paths in combination with a root directory */
677 if (! path_is_absolute ( path
))
680 path
= prefix_roota ( root
, path
);
684 r
= path_make_absolute_cwd ( path
, & buffer
);
688 fd
= open ( "/" , O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
692 if ( flags
& CHASE_SAFE
) {
693 if ( fstat ( fd
, & previous_stat
) < 0 )
699 _cleanup_free_
char * first
= NULL
;
700 _cleanup_close_
int child
= - 1 ;
704 /* Determine length of first component in the path */
705 n
= strspn ( todo
, "/" ); /* The slashes */
706 m
= n
+ strcspn ( todo
+ n
, "/" ); /* The entire length of the component */
708 /* Extract the first component. */
709 first
= strndup ( todo
, m
);
715 /* Empty? Then we reached the end. */
719 /* Just a single slash? Then we reached the end. */
720 if ( path_equal ( first
, "/" )) {
721 /* Preserve the trailing slash */
723 if ( flags
& CHASE_TRAIL_SLASH
)
724 if (! strextend (& done
, "/" , NULL
))
730 /* Just a dot? Then let's eat this up. */
731 if ( path_equal ( first
, "/." ))
734 /* Two dots? Then chop off the last bit of what we already found out. */
735 if ( path_equal ( first
, "/.." )) {
736 _cleanup_free_
char * parent
= NULL
;
737 _cleanup_close_
int fd_parent
= - 1 ;
739 /* If we already are at the top, then going up will not change anything. This is in-line with
740 * how the kernel handles this. */
741 if ( empty_or_root ( done
))
744 parent
= dirname_malloc ( done
);
748 /* Don't allow this to leave the root dir. */
750 path_startswith ( done
, root
) &&
751 ! path_startswith ( parent
, root
))
754 free_and_replace ( done
, parent
);
756 if ( flags
& CHASE_STEP
)
759 fd_parent
= openat ( fd
, ".." , O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
763 if ( flags
& CHASE_SAFE
) {
764 if ( fstat ( fd_parent
, & st
) < 0 )
767 if (! safe_transition (& previous_stat
, & st
))
774 fd
= TAKE_FD ( fd_parent
);
779 /* Otherwise let's see what this is. */
780 child
= openat ( fd
, first
+ n
, O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
783 if ( errno
== ENOENT
&&
784 ( flags
& CHASE_NONEXISTENT
) &&
785 ( isempty ( todo
) || path_is_normalized ( todo
))) {
787 /* If CHASE_NONEXISTENT is set, and the path does not exist, then that's OK, return
788 * what we got so far. But don't allow this if the remaining path contains "../ or "./"
789 * or something else weird. */
791 /* If done is "/", as first also contains slash at the head, then remove this redundant slash. */
792 if ( streq_ptr ( done
, "/" ))
795 if (! strextend (& done
, first
, todo
, NULL
))
805 if ( fstat ( child
, & st
) < 0 )
807 if (( flags
& CHASE_SAFE
) &&
808 ! safe_transition (& previous_stat
, & st
))
813 if (( flags
& CHASE_NO_AUTOFS
) &&
814 fd_is_fs_type ( child
, AUTOFS_SUPER_MAGIC
) > 0 )
817 if ( S_ISLNK ( st
. st_mode
)) {
820 _cleanup_free_
char * destination
= NULL
;
822 /* This is a symlink, in this case read the destination. But let's make sure we don't follow
823 * symlinks without bounds. */
824 if (-- max_follow
<= 0 )
827 r
= readlinkat_malloc ( fd
, first
+ n
, & destination
);
830 if ( isempty ( destination
))
833 if ( path_is_absolute ( destination
)) {
835 /* An absolute destination. Start the loop from the beginning, but use the root
836 * directory as base. */
839 fd
= open ( root
?: "/" , O_CLOEXEC
| O_NOFOLLOW
| O_PATH
);
843 if ( flags
& CHASE_SAFE
) {
844 if ( fstat ( fd
, & st
) < 0 )
847 if (! safe_transition (& previous_stat
, & st
))
855 /* Note that we do not revalidate the root, we take it as is. */
864 /* Prefix what's left to do with what we just read, and start the loop again, but
865 * remain in the current directory. */
866 joined
= strjoin ( destination
, todo
);
868 joined
= strjoin ( "/" , destination
, todo
);
873 todo
= buffer
= joined
;
875 if ( flags
& CHASE_STEP
)
881 /* If this is not a symlink, then let's just add the name we read to what we already verified. */
883 done
= TAKE_PTR ( first
);
885 /* If done is "/", as first also contains slash at the head, then remove this redundant slash. */
886 if ( streq ( done
, "/" ))
889 if (! strextend (& done
, first
, NULL
))
893 /* And iterate again, but go one directory further down. */
899 /* Special case, turn the empty string into "/", to indicate the root directory. */
906 * ret
= TAKE_PTR ( done
);
908 if ( flags
& CHASE_OPEN
) {
909 /* Return the O_PATH fd we currently are looking to the caller. It can translate it to a proper fd by
910 * opening /proc/self/fd/xyz. */
916 if ( flags
& CHASE_STEP
)
928 c
= strjoin ( done
, todo
);
948 int chase_symlinks_and_open (
951 unsigned chase_flags
,
955 _cleanup_close_
int path_fd
= - 1 ;
956 _cleanup_free_
char * p
= NULL
;
959 if ( chase_flags
& CHASE_NONEXISTENT
)
962 if ( empty_or_root ( root
) && ! ret_path
&& ( chase_flags
& ( CHASE_NO_AUTOFS
| CHASE_SAFE
)) == 0 ) {
963 /* Shortcut this call if none of the special features of this call are requested */
964 r
= open ( path
, open_flags
);
971 path_fd
= chase_symlinks ( path
, root
, chase_flags
| CHASE_OPEN
, ret_path
? & p
: NULL
);
975 r
= fd_reopen ( path_fd
, open_flags
);
980 * ret_path
= TAKE_PTR ( p
);
985 int chase_symlinks_and_opendir (
988 unsigned chase_flags
,
992 char procfs_path
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( int )];
993 _cleanup_close_
int path_fd
= - 1 ;
994 _cleanup_free_
char * p
= NULL
;
999 if ( chase_flags
& CHASE_NONEXISTENT
)
1002 if ( empty_or_root ( root
) && ! ret_path
&& ( chase_flags
& ( CHASE_NO_AUTOFS
| CHASE_SAFE
)) == 0 ) {
1003 /* Shortcut this call if none of the special features of this call are requested */
1012 path_fd
= chase_symlinks ( path
, root
, chase_flags
| CHASE_OPEN
, ret_path
? & p
: NULL
);
1016 xsprintf ( procfs_path
, "/proc/self/fd/%i" , path_fd
);
1017 d
= opendir ( procfs_path
);
1022 * ret_path
= TAKE_PTR ( p
);
1028 int chase_symlinks_and_stat (
1031 unsigned chase_flags
,
1033 struct stat
* ret_stat
) {
1035 _cleanup_close_
int path_fd
= - 1 ;
1036 _cleanup_free_
char * p
= NULL
;
1041 if ( chase_flags
& CHASE_NONEXISTENT
)
1044 if ( empty_or_root ( root
) && ! ret_path
&& ( chase_flags
& ( CHASE_NO_AUTOFS
| CHASE_SAFE
)) == 0 ) {
1045 /* Shortcut this call if none of the special features of this call are requested */
1046 if ( stat ( path
, ret_stat
) < 0 )
1052 path_fd
= chase_symlinks ( path
, root
, chase_flags
| CHASE_OPEN
, ret_path
? & p
: NULL
);
1056 if ( fstat ( path_fd
, ret_stat
) < 0 )
1060 * ret_path
= TAKE_PTR ( p
);
1062 if ( chase_flags
& CHASE_OPEN
)
1063 return TAKE_FD ( path_fd
);
1068 int access_fd ( int fd
, int mode
) {
1069 char p
[ STRLEN ( "/proc/self/fd/" ) + DECIMAL_STR_MAX ( fd
) + 1 ];
1072 /* Like access() but operates on an already open fd */
1074 xsprintf ( p
, "/proc/self/fd/%i" , fd
);
1075 r
= access ( p
, mode
);
1082 int unlinkat_deallocate ( int fd
, const char * name
, int flags
) {
1083 _cleanup_close_
int truncate_fd
= - 1 ;
1087 /* Operates like unlinkat() but also deallocates the file contents if it is a regular file and there's no other
1088 * link to it. This is useful to ensure that other processes that might have the file open for reading won't be
1089 * able to keep the data pinned on disk forever. This call is particular useful whenever we execute clean-up
1090 * jobs ("vacuuming"), where we want to make sure the data is really gone and the disk space released and
1091 * returned to the free pool.
1093 * Deallocation is preferably done by FALLOC_FL_PUNCH_HOLE|FALLOC_FL_KEEP_SIZE (đ) if supported, which means
1094 * the file won't change size. That's a good thing since we shouldn't needlessly trigger SIGBUS in other
1095 * programs that have mmap()ed the file. (The assumption here is that changing file contents to all zeroes
1096 * underneath those programs is the better choice than simply triggering SIGBUS in them which truncation does.)
1097 * However if hole punching is not implemented in the kernel or file system we'll fall back to normal file
1098 * truncation (đĒ), as our goal of deallocating the data space trumps our goal of being nice to readers (đ).
1100 * Note that we attempt deallocation, but failure to succeed with that is not considered fatal, as long as the
1101 * primary job â to delete the file â is accomplished. */
1103 if (( flags
& AT_REMOVEDIR
) == 0 ) {
1104 truncate_fd
= openat ( fd
, name
, O_WRONLY
| O_CLOEXEC
| O_NOCTTY
| O_NOFOLLOW
| O_NONBLOCK
);
1105 if ( truncate_fd
< 0 ) {
1107 /* If this failed because the file doesn't exist propagate the error right-away. Also,
1108 * AT_REMOVEDIR wasn't set, and we tried to open the file for writing, which means EISDIR is
1109 * returned when this is a directory but we are not supposed to delete those, hence propagate
1110 * the error right-away too. */
1111 if ( IN_SET ( errno
, ENOENT
, EISDIR
))
1114 if ( errno
!= ELOOP
) /* don't complain if this is a symlink */
1115 log_debug_errno ( errno
, "Failed to open file '%s' for deallocation, ignoring: %m" , name
);
1119 if ( unlinkat ( fd
, name
, flags
) < 0 )
1122 if ( truncate_fd
< 0 ) /* Don't have a file handle, can't do more âšī¸ */
1125 if ( fstat ( truncate_fd
, & st
) < 0 ) {
1126 log_debug_errno ( errno
, "Failed to stat file '%s' for deallocation, ignoring." , name
);
1130 if (! S_ISREG ( st
. st_mode
) || st
. st_blocks
== 0 || st
. st_nlink
> 0 )
1133 /* If this is a regular file, it actually took up space on disk and there are no other links it's time to
1134 * punch-hole/truncate this to release the disk space. */
1136 bs
= MAX ( st
. st_blksize
, 512 );
1137 l
= DIV_ROUND_UP ( st
. st_size
, bs
) * bs
; /* Round up to next block size */
1139 if ( fallocate ( truncate_fd
, FALLOC_FL_PUNCH_HOLE
| FALLOC_FL_KEEP_SIZE
, 0 , l
) >= 0 )
1140 return 0 ; /* Successfully punched a hole! đ */
1142 /* Fall back to truncation */
1143 if ( ftruncate ( truncate_fd
, 0 ) < 0 ) {
1144 log_debug_errno ( errno
, "Failed to truncate file to 0, ignoring: %m" );
1151 int fsync_directory_of_file ( int fd
) {
1152 _cleanup_free_
char * path
= NULL
, * dn
= NULL
;
1153 _cleanup_close_
int dfd
= - 1 ;
1156 r
= fd_verify_regular ( fd
);
1160 r
= fd_get_path ( fd
, & path
);
1162 log_debug ( "Failed to query /proc/self/fd/%d%s: %m" ,
1164 r
== - EOPNOTSUPP
? ", ignoring" : "" );
1166 if ( r
== - EOPNOTSUPP
)
1167 /* If /proc is not available, we're most likely running in some
1168 * chroot environment, and syncing the directory is not very
1169 * important in that case. Let's just silently do nothing. */
1175 if (! path_is_absolute ( path
))
1178 dn
= dirname_malloc ( path
);
1182 dfd
= open ( dn
, O_RDONLY
| O_CLOEXEC
| O_DIRECTORY
);