Re: [bpf PATCH v2 6/6] bpf: sockmap/tls, close can race with map free

[John Fastabend]

Jakub Kicinski wrote:
> On Tue, 09 Jul 2019 20:33:58 -0700, John Fastabend wrote:
> > Jakub Kicinski wrote:
> > > On Mon, 08 Jul 2019 19:15:18 +0000, John Fastabend wrote:  
> > > > @@ -352,15 +354,18 @@ static void tls_sk_proto_close(struct sock *sk, 
> > > > long timeout)
> > > >         if (ctx->tx_conf == TLS_BASE && ctx->rx_conf == TLS_BASE)
> > > >                 goto skip_tx_cleanup;
> > > >  
> > > > -       sk->sk_prot = ctx->sk_proto;
> > > >         tls_sk_proto_cleanup(sk, ctx, timeo);
> > > >  
> > > >  skip_tx_cleanup:
> > > > +       write_lock_bh(&sk->sk_callback_lock);
> > > > +       icsk->icsk_ulp_data = NULL;  
> > > 
> > > Is ulp_data pointer now supposed to be updated under the
> > > sk_callback_lock?  
> > 
> > Yes otherwise it can race with tls_update(). I didn't remove the
> > ulp pointer null set from tcp_ulp.c though. Could be done in this
> > patch or as a follow up.
> 
> Do we need to hold the lock in unhash, too, or is unhash called with
> sk_callback_lock held?
> 

We should hold the lock here. Also we should reset sk_prot similar to
other paths in case we get here without a close() call. syzbot hasn't
found that path yet but I'll add some tests for it.

        write_lock_bh(...)
        icsk_ulp_data = NULL
        sk->sk_prot = ctx->sk_proto;
        write_unlock_bh(...)

Thanks