Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Sun, 9 Oct 2022 10:42:59 +0200
From:   Christian Brauner <brauner@kernel.org>
To:     "Christian A. Ehrhardt" <lk@c--e.de>
Cc:     Tejun Heo <tj@kernel.org>,
        syzbot <syzbot+534ee3d24c37c411f37f@syzkaller.appspotmail.com>,
        gregkh@linuxfoundation.org, linux-kernel@vger.kernel.org,
        syzkaller-bugs@googlegroups.com,
        Yosry Ahmed <yosryahmed@google.com>
Subject: Re: [syzbot] general protection fault in kernfs_get_inode
Message-ID: <20221009084039.cw6meqbvy4362lsa@wittgenstein>
References: <000000000000385cbf05ea3f1862@google.com>
 <00000000000028a44005ea40352b@google.com>
 <Y0CbtVwW6+QIYJdQ@slm.duckdns.org>
 <Y0HBlJ4AoZba93Uf@cae.in-ulm.de>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
In-Reply-To: <Y0HBlJ4AoZba93Uf@cae.in-ulm.de>
Precedence: bulk

On Sat, Oct 08, 2022 at 08:29:40PM +0200, Christian A. Ehrhardt wrote:
> 
> Hi (from another Christian),
> 
> On Fri, Oct 07, 2022 at 11:35:49AM -1000, Tejun Heo wrote:
> > (cc'ing Christian and quoting whole body)
> > 
> > Christan, I can't repro it here but think what we need is sth like the
> > following. The problem is that cgroup_is_dead() check in the fork path isn't
> > enough as the perm check depends on cgrp->procs_file being available but
> > that is cleared while DYING before DEAD. So, depending on the timing, we can
> > end up trying to deref NULL pointer in may_write.
> > 
> > diff --git a/kernel/cgroup/cgroup.c b/kernel/cgroup/cgroup.c
> > index 8ad2c267ff471..603b7167450a1 100644
> > --- a/kernel/cgroup/cgroup.c
> > +++ b/kernel/cgroup/cgroup.c
> > @@ -4934,6 +4934,10 @@ static int cgroup_may_write(const struct cgroup *cgrp, struct super_block *sb)
> >  
> >  	lockdep_assert_held(&cgroup_mutex);
> >  
> > +	/*if @cgrp is being removed, procs_file may already be gone */
> > +	if (!cgrp->procs_file.kn)
> > +		return -ENODEV;
> > +
> >  	inode = kernfs_get_inode(sb, cgrp->procs_file.kn);
> >  	if (!inode)
> >  		return -ENOMEM;
> 
> I had syzbot hit the same crash here and as can be seen from the
> reproducer the root cause is that the target cgroup (specified
> via CLONE_INTO_CGROUP) is a version 1 cgroup. These cgroups just
> don't initialize ->procs_file.kn.
> 
> This is a regression from v6.0 caused by 
> 
>    f3a2aebdd6 ("cgroup: enable cgroup_get_from_file() on cgroup1")   

Yeah, this patch is wrong in its simple form and definitely breaks CLONE_INTO_CGROUP.
CLONE_INTO_CGROUP can only work with cgroup2 fds. It absolutely cannot
work with cgroup1 fds. The semantics would be terrible as controllers
can be mounted into separate hierarchies.

> 
> Unless we want to revert this patch the correct fix might be
> something like this:
> 
> diff --git a/kernel/cgroup/cgroup.c b/kernel/cgroup/cgroup.c
> index f487c54a0087..67474b1ae087 100644
> --- a/kernel/cgroup/cgroup.c
> +++ b/kernel/cgroup/cgroup.c
> @@ -6249,6 +6249,11 @@ static int cgroup_css_set_fork(struct kernel_clone_args *kargs)
>  		goto err;
>  	}
>  
> +	if (!cgroup_on_dfl(dst_cgrp)) {
> +		ret = -EBADF;
> +		goto err;
> +	}

That seems like a good enough patch to me.