Skip to contentSkip to conversation

Implement AT_SYMLINK_FOLLOW flag for linkat

PATCH In review Ulrich Drepper 3 participants 6 messages Jun 17, 2006 – Jun 18, 2006

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Implement AT_SYMLINK_FOLLOW flag for linkat

When the linkat() syscall was added the flag parameter was added in the last minute but it wasn't used so far. The following patch should change that. My tests show that this is all that's needed.

If OLDNAME is a symlink setting the flag causes linkat to follow the symlink and create a hardlink with the target. This is actually the behavior POSIX demands for link() as well but Linux wisely does not do this. With this flag (which will most likely be in the next POSIX revision) the programmer can choose the behavior, defaulting to the safe variant. As a side effect it is now possible to implement a POSIX-compliant link(2) function for those who are interested.

touch file
ln -s file symlink

linkat(fd, "symlink", fd, "newlink", 0)

    -> newlink is hardlink of symlink

linkat(fd, "symlink", fd, "newlink", AT_SYMLINK_FOLLOW)

    -> newlink is hardlink of file

The value of AT_SYMLINK_FOLLOW is determined by the definition we already use in glibc.

 fs/namei.c            |    6 ++++--
 include/linux/fcntl.h |    1 +
 2 files changed, 5 insertions(+), 2 deletions(-)
fs/namei.c renamed from fs/namei.c-follow +4 -2
@@ -2243,14 +2243,16 @@
 	int error;
 	char * to;
 
-	if (flags != 0)
+	if ((flags & ~AT_SYMLINK_FOLLOW) != 0)
 		return -EINVAL;
 
 	to = getname(newname);
 	if (IS_ERR(to))
 		return PTR_ERR(to);
 
-	error = __user_walk_fd(olddfd, oldname, 0, &old_nd);
+	error = __user_walk_fd(olddfd, oldname,
+			       flags & AT_SYMLINK_FOLLOW ? LOOKUP_FOLLOW : 0,
+			       &old_nd);
 	if (error)
 		goto exit;
 	error = do_path_lookup(newdfd, to, LOOKUP_PARENT, &nd);
include/linux/fcntl.h renamed from include/linux/fcntl.h-follow +1 -0
@@ -29,6 +29,7 @@
 #define AT_SYMLINK_NOFOLLOW	0x100   /* Do not follow symbolic links.  */
 #define AT_REMOVEDIR		0x200   /* Remove directory instead of
                                            unlinking file.  */
+#define AT_SYMLINK_FOLLOW	0x400   /* Follow symbolic links.  */
 
 #ifdef __KERNEL__
 
#1 lore raw
Quoted 11 lines · Ulrich Drepper

When the linkat() syscall was added the flag parameter was added in the last minute but it wasn't used so far. The following patch should change that. My tests show that this is all that's needed.

If OLDNAME is a symlink setting the flag causes linkat to follow the symlink and create a hardlink with the target. This is actually the behavior POSIX demands for link() as well but Linux wisely does not do this. With this flag (which will most likely be in the next POSIX revision) the programmer can choose the behavior, defaulting to the safe variant. As a side effect it is now possible to implement a POSIX-compliant link(2) function for those who are interested.

Where does POSIX require that? IIRC, it was along the lines of "application can't rely on kernel doing the right thing", not "kernel must do the wrong thing"...

#2 lore raw
Quoted 4 lines · Al Viro

Where does POSIX require that? IIRC, it was along the lines of "application can't rely on kernel doing the right thing", not "kernel must do the wrong thing"...

Well, the patch as sent in does seem sane, as long as glibc doesn't start defaulting to the insane behaviour. Giving users the _ability_ to link to the symlink target is certainly not wrong, regardless of any standard. Doing it by default is another matter.

Linus

#3 lore raw
Quoted 4 lines · Linus Torvalds
Quoted 1 line · Al Viro
Quoted 8 lines
Quoted 3 lines

Where does POSIX require that? IIRC, it was along the lines of "application can't rely on kernel doing the right thing", not "kernel must do the wrong thing"...

Well, the patch as sent in does seem sane, as long as glibc doesn't start defaulting to the insane behaviour. Giving users the _ability_ to link to the symlink target is certainly not wrong, regardless of any standard. Doing it by default is another matter.

As long as it doesn't get tied to overloaded environment variable that might be needed for other purposes...

#4 lore raw
In reply to Linus Torvalds
Quoted 4 lines · Linus Torvalds

Well, the patch as sent in does seem sane, as long as glibc doesn't start defaulting to the insane behaviour. Giving users the _ability_ to link to the symlink target is certainly not wrong, regardless of any standard. Doing it by default is another matter.

I do not intend to change the link implementation in glibc. That would be majorly stupid, it'd break the ABI.

The AT_SYMLINK_FOLLOW flag to linkat was the result of the discussion how to resolve the issue of the conflict between POSIX and the Linux implementation of link (BTW: the Solaris link syscall behaves the same as Linux's). This is an easy an non-intrusive way to help people who depend on the questionable POSIx-mandated behavior to work around the incompatiblity. Nothing more. Don't change the link syscall, don't assume the glibc will be changed. This is only one little extra bit of new functionality.

Signature · 1 line
➧ Ulrich Drepper ➧ Red Hat, Inc. ➧ 444 Castro St ➧ Mountain View, CA ❖
Acked-by
#5 lore raw
Quoted 6 lines · Ulrich Drepper
Quoted 4 lines · Linus Torvalds

Well, the patch as sent in does seem sane, as long as glibc doesn't start defaulting to the insane behaviour. Giving users the _ability_ to link to the symlink target is certainly not wrong, regardless of any standard. Doing it by default is another matter.

Quoted 7 lines

I do not intend to change the link implementation in glibc. That would be majorly stupid, it'd break the ABI.

The AT_SYMLINK_FOLLOW flag to linkat was the result of the discussion how to resolve the issue of the conflict between POSIX and the Linux implementation of link (BTW: the Solaris link syscall behaves the same as Linux's).

... while FreeBSD still doesn't have that 4.2BSD bug fixed, the suckers.

Quoted 5 lines

This is an easy an non-intrusive way to help people who
depend on the questionable POSIx-mandated behavior to work around the incompatiblity. Nothing more. Don't change the link syscall, don't assume the glibc will be changed. This is only one little extra bit of new functionality.

*shrug*

Fine by me; it's not really useful, but it's not a serious bloat either.