Re: [Patch bpf-next] sock_map: fix a potential use-after-free in sock_map_close()

From: Cong Wang <xiyou.wangcong@gmail.com>
To: John Fastabend <john.fastabend@gmail.com>
Cc: Linux Kernel Network Developers <netdev@vger.kernel.org>,
	bpf <bpf@vger.kernel.org>, Cong Wang <cong.wang@bytedance.com>,
	syzbot <syzbot+7b6548ae483d6f4c64ae@syzkaller.appspotmail.com>,
	Daniel Borkmann <daniel@iogearbox.net>,
	Jakub Sitnicki <jakub@cloudflare.com>,
	Lorenz Bauer <lmb@cloudflare.com>
Subject: Re: [Patch bpf-next] sock_map: fix a potential use-after-free in sock_map_close()
Date: Thu, 8 Apr 2021 21:08:05 -0700	[thread overview]
Message-ID: <CAM_iQpXqfAXPoesdXskH7BaE206zc5QDEsfjFnfSRck2FH+wLg@mail.gmail.com> (raw)
In-Reply-To: <606f9f2b26b1a_c8b9208a4@john-XPS-13-9370.notmuch>

On Thu, Apr 8, 2021 at 5:26 PM John Fastabend <john.fastabend@gmail.com> wrote:
>
> Cong Wang wrote:
> > From: Cong Wang <cong.wang@bytedance.com>
> >
> > The last refcnt of the psock can be gone right after
> > sock_map_remove_links(), so sk_psock_stop() could trigger a UAF.
> > The reason why I placed sk_psock_stop() there is to avoid RCU read
> > critical section, and more importantly, some callee of
> > sock_map_remove_links() is supposed to be called with RCU read lock,
> > we can not simply get rid of RCU read lock here. Therefore, the only
> > choice we have is to grab an additional refcnt with sk_psock_get()
> > and put it back after sk_psock_stop().
> >
> > Reported-by: syzbot+7b6548ae483d6f4c64ae@syzkaller.appspotmail.com
> > Fixes: 799aa7f98d53 ("skmsg: Avoid lock_sock() in sk_psock_backlog()")
> > Cc: John Fastabend <john.fastabend@gmail.com>
> > Cc: Daniel Borkmann <daniel@iogearbox.net>
> > Cc: Jakub Sitnicki <jakub@cloudflare.com>
> > Cc: Lorenz Bauer <lmb@cloudflare.com>
> > Signed-off-by: Cong Wang <cong.wang@bytedance.com>
> > ---
> >  net/core/sock_map.c | 3 ++-
> >  1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/net/core/sock_map.c b/net/core/sock_map.c
> > index f473c51cbc4b..6f1b82b8ad49 100644
> > --- a/net/core/sock_map.c
> > +++ b/net/core/sock_map.c
> > @@ -1521,7 +1521,7 @@ void sock_map_close(struct sock *sk, long timeout)
> >
> >       lock_sock(sk);
> >       rcu_read_lock();
>
> It looks like we can drop the rcu_read_lock()/unlock() section then if we
> take a reference on the psock? Before it was there to ensure we didn't
> lose the psock from some other context, but with a reference held this
> can not happen.

Some callees under sock_map_remove_links() still assert RCU read
lock, so we can not simply drop the RCU read lock here. Some
additional efforts are needed to take care of those assertions, which
can be a separate patch.

Thanks.