Re: [Linux-kernel-mentees] [PATCH] fs: fix KMSAN uninit-value bug by initializing nd in do_file_open_root

From: Al Viro
Date: Sat Sep 19 2020 - 12:17:51 EST

On Sat, Sep 19, 2020 at 04:44:51PM +0200, Greg KH wrote:

> > ->dir_uid and ->dir_mode are set when link_path_walk() resolves the pathname
> > to directory + final component. They are used when deciding whether to reject
> > a trailing symlink (on fs.protected_symlinks setups) and whether to allow
> > creation in sticky directories (on fs.protected_regular and fs.protected_fifos
> > setups). Both operations really need the results of successful link_path_walk().
> >
> > I don't see how that could be not a false positive. If we hit the use in
> > may_create_in_sticky(), we'd need the combination of
> > * pathname that consists only of slashes (or it will be initialized)
> > * LAST_NORM in nd->last_type, which is flat-out impossible, since
> > we are left with LAST_ROOT for such pathnames. The same goes for
> > may_follow_link() use - we need WALK_TRAILING in flags to hit it in the
> > first place, which can come from two sources -
> > return walk_component(nd, WALK_TRAILING);
> > in lookup_last() (and walk_component() won't go anywhere near the
> > call chain leading to may_follow_link() without LAST_NORM in nd->last_type)
> > and
> > res = step_into(nd, WALK_TRAILING, dentry, inode, seq);
> > in open_last_lookups(), which also won't go anywhere near that line without
> > LAST_NORM in the nd->last_type.
> >
> > IOW, unless we manage to call that without having called link_path_walk()
> > at all or after link_path_walk() returning an error, we shouldn't hit
> > that. And if we *do* go there without link_path_walk() or with an error
> > from link_path_walk(), we have a much worse problem.
> >
> > I want to see the details of reproducer. If it's for real, we have a much
> > more serious problem; if it's a false positive, the right place to deal
> > with it would be elsewhere (perhaps on return from link_path_walk() with
> > a slashes-only pathname), but in any case it should only be done after we
> > manage to understand what's going on.
> Reproducer is pretty simple:
> Now if that is actually valid or not, I don't know...

Lovely... That would get an empty path and non-directory for a starting
point, but it should end up with LAST_ROOT in nd->last_type. Which should
not be able to reach the readers of those fields... Which kernel had
that been on?