Re: general protection fault in tls_write_space

[John Fastabend]

Jakub Kicinski wrote:
> On Tue, 13 Aug 2019 10:17:06 -0700, John Fastabend wrote:
> > > Followup of commit 95fa145479fb
> > > ("bpf: sockmap/tls, close can race with map free")
> > > 
> > > --- a/net/tls/tls_main.c
> > > +++ b/net/tls/tls_main.c
> > > @@ -308,6 +308,9 @@ static void tls_sk_proto_close(struct so
> > >  	if (free_ctx)
> > >  		icsk->icsk_ulp_data = NULL;
> > >  	sk->sk_prot = ctx->sk_proto;
> > > +	/* tls will go; restore sock callback before enabling bh */
> > > +	if (sk->sk_write_space == tls_write_space)
> > > +		sk->sk_write_space = ctx->sk_write_space;
> > >  	write_unlock_bh(&sk->sk_callback_lock);
> > >  	release_sock(sk);
> > >  	if (ctx->tx_conf == TLS_SW)  
> > 
> > Hi Hillf,
> > 
> > We need this patch (although slightly updated for bpf tree) do
> > you want to send it? Otherwise I can. We should only set this if
> > TX path was enabled otherwise we null it. Checking against
> > tls_write_space seems best to me as well.
> > 
> > Against bpf this patch should fix it.
> > 
> > diff --git a/net/tls/tls_main.c b/net/tls/tls_main.c
> > index ce6ef56a65ef..43252a801c3f 100644
> > --- a/net/tls/tls_main.c
> > +++ b/net/tls/tls_main.c
> > @@ -308,7 +308,8 @@ static void tls_sk_proto_close(struct sock *sk, long timeout)
> >         if (free_ctx)
> >                 icsk->icsk_ulp_data = NULL;
> >         sk->sk_prot = ctx->sk_proto;
> > -       sk->sk_write_space = ctx->sk_write_space;
> > +       if (sk->sk_write_space == tls_write_space)
> > +               sk->sk_write_space = ctx->sk_write_space;
> >         write_unlock_bh(&sk->sk_callback_lock);
> >         release_sock(sk);
> >         if (ctx->tx_conf == TLS_SW)
> 
> This is already in net since Friday:

Don't we need to guard that with an

  if (sk->sk_write_space == tls_write_space)

or something similar? Where is ctx->sk_write_space set in the rx only
case? In do_tls_setsockop_conf() we have this block

	if (tx) {
		ctx->sk_write_space = sk->sk_write_space;
		sk->sk_write_space = tls_write_space;
	} else {
		sk->sk_socket->ops = &tls_sw_proto_ops;
	}

which makes me think ctx->sk_write_space may not be set correctly in
all cases.

Thanks.

> 
> commit 57c722e932cfb82e9820bbaae1b1f7222ea97b52
> Author: Jakub Kicinski <jakub.kicinski@xxxxxxxxxxxxx>
> Date:   Fri Aug 9 18:36:23 2019 -0700
> 
>     net/tls: swap sk_write_space on close
>     
>     Now that we swap the original proto and clear the ULP pointer
>     on close we have to make sure no callback will try to access
>     the freed state. sk_write_space is not part of sk_prot, remember
>     to swap it.
>     
>     Reported-by: syzbot+dcdc9deefaec44785f32@xxxxxxxxxxxxxxxxxxxxxxxxx
>     Fixes: 95fa145479fb ("bpf: sockmap/tls, close can race with map free")
>     Signed-off-by: Jakub Kicinski <jakub.kicinski@xxxxxxxxxxxxx>
>     Signed-off-by: David S. Miller <davem@xxxxxxxxxxxxx>
> 
> diff --git a/net/tls/tls_main.c b/net/tls/tls_main.c
> index 9cbbae606ced..ce6ef56a65ef 100644
> --- a/net/tls/tls_main.c
> +++ b/net/tls/tls_main.c
> @@ -308,6 +308,7 @@ static void tls_sk_proto_close(struct sock *sk, long timeout)
>         if (free_ctx)
>                 icsk->icsk_ulp_data = NULL;
>         sk->sk_prot = ctx->sk_proto;
> +       sk->sk_write_space = ctx->sk_write_space;
>         write_unlock_bh(&sk->sk_callback_lock);
>         release_sock(sk);
>         if (ctx->tx_conf == TLS_SW)