Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Thu, 17 Sep 2020 01:22:38 +0100
From:   Al Viro <viro@zeniv.linux.org.uk>
To:     Eric Biggers <ebiggers@kernel.org>
Cc:     Anant Thazhemadam <anant.thazhemadam@gmail.com>,
        linux-kernel-mentees@lists.linuxfoundation.org,
        syzbot+4191a44ad556eacc1a7a@syzkaller.appspotmail.com,
        linux-fsdevel@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH] fs: fix KMSAN uninit-value bug by initializing nd in
 do_file_open_root
Message-ID: <20200917002238.GO3421308@ZenIV.linux.org.uk>
References: <20200916052657.18683-1-anant.thazhemadam@gmail.com>
 <20200916054157.GC825@sol.localdomain>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20200916054157.GC825@sol.localdomain>
Precedence: bulk

On Tue, Sep 15, 2020 at 10:41:57PM -0700, Eric Biggers wrote:

> Looking at the actual KMSAN report, it looks like it's nameidata::dir_mode or
> nameidata::dir_uid that is uninitialized.  You need to figure out the correct
> solution, not just blindly initialize with zeroes -- that could hide a bug.
> Is there a bug that is preventing these fields from being initialized to the
> correct values, are these fields being used when they shouldn't be, etc...

False positive, and this is the wrong place to shut it up.

->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.