Re: [RFC][PATCH] getname_maybe_null() - the third variant of pathname copy-in

[Al Viro <viro@zeniv.linux.org.uk>]

On Tue, Oct 15, 2024 at 04:05:15PM +0200, Christian Brauner wrote:
> Looks good.
> 
> Fyi, I'm using your #base.getname as a base for some other work that I'm
> currently doing. So please don't rebase #base.getname anymore. ;)
> 
> Since you have your #work.xattr and #work.stat using it as base it seems
> pretty unlikely anyway but I just thought I mention explicitly that I'm
> relying on that #base.getname branch.

FWIW, I see a problem with that sucker.  The trouble is in the
combination AT_FDCWD, "", AT_EMPTY_PATH.  vfs_empty_path() returns
false on that, so fstatat() in mainline falls back to vfs_statx() and
does the right thing.  This variant does _not_.

Note that your variant of xattr series also ended up failing on e.g.
getxattrat() with such combination:
        if (at_flags & AT_EMPTY_PATH && vfs_empty_path(dfd, pathname)) {
                CLASS(fd, f)(dfd);
                if (!f.file)
                        return -EBADF;
                audit_file(f.file);
                return getxattr(file_mnt_idmap(f.file), file_dentry(f.file),
                                name, value, size);
        }

        lookup_flags = (at_flags & AT_SYMLINK_NOFOLLOW) ? 0 : LOOKUP_FOLLOW;

retry:
        error = user_path_at(dfd, pathname, lookup_flags, &path);

ended up calling user_path_at() with empty pathname and nothing like LOOKUP_EMPTY
in lookup_flags.  Which bails out with -ENOENT, since getname() in there does
so.  My variant bails out with -EBADF and I'd argue that neither is correct.

Not sure what's the sane solution here, need to think for a while...