Re: [PATCH] netfilter: check the length of the data before dereferencing it

[Pablo Neira Ayuso <pablo@netfilter.org>]

On Sun, Apr 01, 2012 at 06:55:22PM +0200, Eric Dumazet wrote:
> On Sun, 2012-04-01 at 18:43 +0200, Pablo Neira Ayuso wrote:
> > On Sun, Apr 01, 2012 at 10:22:50PM +0800, Changli Gao wrote:
> > > We should check the length of the data before dereferencing it when parsing
> > > the TCP options.
> > > 
> > > Signed-off-by: Changli Gao <xiaosuo@gmail.com>
> > > ---
> > >  net/netfilter/nf_conntrack_proto_tcp.c |    4 ++++
> > >  1 file changed, 4 insertions(+)
> > > diff --git a/net/netfilter/nf_conntrack_proto_tcp.c b/net/netfilter/nf_conntrack_proto_tcp.c
> > > index 361eade..9e446c5 100644
> > > --- a/net/netfilter/nf_conntrack_proto_tcp.c
> > > +++ b/net/netfilter/nf_conntrack_proto_tcp.c
> > > @@ -404,6 +404,8 @@ static void tcp_options(const struct sk_buff *skb,
> > >  			length--;
> > >  			continue;
> > >  		default:
> > > +			if (length < 2)
> > > +				return;
> > >  			opsize=*ptr++;
> > >  			if (opsize < 2) /* "silly options" */
> > >  				return;
> > 
> > length is always multiple of 4:
> > 
> > int length = (tcph->doff*4) - sizeof(struct tcphdr);
> > --
> 
> initial value yes, but it can change in the loop.

Indeed, then I think we need a similar patch for tcp_parse_options()
in net/ipv4/tcp_input.c