Re: [PATCH] net: check if protocol extracted by virtio_net_hdr_set_proto is correct

[Jason Wang]

On 2021/2/19 10:55 下午, Willem de Bruijn wrote:
> On Fri, Feb 19, 2021 at 3:53 AM Jason Wang <jasowang@xxxxxxxxxx> wrote:
> > On 2021/2/18 11:50 下午, Willem de Bruijn wrote:
> > > On Thu, Feb 18, 2021 at 10:01 AM Balazs Nemeth <bnemeth@xxxxxxxxxx> wrote:
> > > > For gso packets, virtio_net_hdr_set_proto sets the protocol (if it isn't
> > > > set) based on the type in the virtio net hdr, but the skb could contain
> > > > anything since it could come from packet_snd through a raw socket. If
> > > > there is a mismatch between what virtio_net_hdr_set_proto sets and
> > > > the actual protocol, then the skb could be handled incorrectly later
> > > > on by gso.
> > > >
> > > > The network header of gso packets starts at 14 bytes, but a specially
> > > > crafted packet could fool the call to skb_flow_dissect_flow_keys_basic
> > > > as the network header offset in the skb could be incorrect.
> > > > Consequently, EINVAL is not returned.
> > > >
> > > > There are even packets that can cause an infinite loop. For example, a
> > > > packet with ethernet type ETH_P_MPLS_UC (which is unnoticed by
> > > > virtio_net_hdr_to_skb) that is sent to a geneve interface will be
> > > > handled by geneve_build_skb. In turn, it calls
> > > > udp_tunnel_handle_offloads which then calls skb_reset_inner_headers.
> > > > After that, the packet gets passed to mpls_gso_segment. That function
> > > > calculates the mpls header length by taking the difference between
> > > > network_header and inner_network_header. Since the two are equal
> > > > (due to the earlier call to skb_reset_inner_headers), it will calculate
> > > > a header of length 0, and it will not pull any headers. Then, it will
> > > > call skb_mac_gso_segment which will again call mpls_gso_segment, etc...
> > > > This leads to the infinite loop.
> >
> > I remember kernel will validate dodgy gso packets in gso ops. I wonder
> > why not do the check there? The reason is that virtio/TUN is not the
> > only source for those packets.
> It is? All other GSO packets are generated by the stack itself, either
> locally or through GRO.


Something like what has been done in tcp_tso_segment()?

    if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
                /* Packet is from an untrusted source, reset gso_segs. */

        skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);

        segs = NULL;
                goto out;
        }

My understanding of the header check logic is that it tries to dealy the 
check as much as possible, so for device that has GRO_ROBUST, there's 
even no need to do that.


> But indeed some checks are better performed in the GSO layer. Such as
> likely the 0-byte mpls header length.
>
> If we cannot trust virtio_net_hdr.gso_type passed from userspace, then
> we can also not trust the eth.h_proto coming from the same source.


I agree.


> But
> it makes sense to require them to be consistent. There is a
> dev_parse_header_protocol that may return the link layer type in a
> more generic fashion than casting to skb_eth_hdr.
>
> Question remains what to do for the link layer types that do not implement
> header_ops->parse_protocol, and so we cannot validate the packet's
> network protocol. Drop will cause false positives, accepts will leave a
> potential path, just closes it for Ethernet.
>
> This might call for multiple fixes, both on first ingest and inside the stack?


It's a balance between performance and security. Ideally, it looks to me 
the GSO codes should not assume the header of dodgy packet is correct 
which means it must validate them before using them. I'm not sure if it 
needs a lot of changes or not.

For security reason, it's better to do a strict check during first 
ingest. But it bascially suppress the meaning of NETIF_F_GSO_ROBUST 
somehow. And it needs some benchmark to see if it can cause obvious 
performance regression.

Thanks


>