On Thu, Apr 04, 2019 at 12:18:08AM +0300, Julian Anastasov wrote:
>
> Hello,
>
> On Wed, 3 Apr 2019, Simon Horman wrote:
>
> > On Sun, Mar 31, 2019 at 01:26:21PM +0300, Julian Anastasov wrote:
> > > Recognize UDP tunnels in received ICMP errors and
> > > properly strip the tunnel headers. GUE is what we
> > > have for now.
> > >
> > > Signed-off-by: Julian Anastasov <ja@xxxxxx>
> > > ---
> > > net/netfilter/ipvs/ip_vs_core.c | 58 +++++++++++++++++++++++++++++++++
> > > 1 file changed, 58 insertions(+)
> > >
> > > diff --git a/net/netfilter/ipvs/ip_vs_core.c b/net/netfilter/ipvs/ip_vs_core.c
> > > index 4447ee512b88..0b624e8e7982 100644
> > > --- a/net/netfilter/ipvs/ip_vs_core.c
> > > +++ b/net/netfilter/ipvs/ip_vs_core.c
> > > @@ -39,6 +39,7 @@
> > > #include <net/tcp.h>
> > > #include <net/udp.h>
> > > #include <net/icmp.h> /* for icmp_send */
> > > +#include <net/gue.h>
> > > #include <net/route.h>
> > > #include <net/ip6_checksum.h>
> > > #include <net/netns/generic.h> /* net_generic() */
> > > @@ -1579,6 +1580,34 @@ ip_vs_try_to_schedule(struct netns_ipvs *ipvs, int af, struct sk_buff *skb,
> > > return 1;
> > > }
> > >
> > > +/* Check the UDP tunnel and return its header length */
> > > +static int ipvs_udp_decap(struct netns_ipvs *ipvs, struct sk_buff *skb,
> > > + unsigned int offset, __u16 af, __be16 dport,
> > > + const union nf_inet_addr *daddr, __u8 *proto)
> > > +{
> > > + struct ip_vs_dest *dest = ip_vs_find_tunnel(ipvs, af, daddr, dport);
> > > +
> > > + if (!dest)
> > > + goto unk;
> > > + if (dest->tun_type == IP_VS_CONN_F_TUNNEL_TYPE_GUE) {
> > > + struct guehdr _gueh, *gueh;
> > > +
> > > + gueh = skb_header_pointer(skb, offset, sizeof(_gueh), &_gueh);
> > > + if (!gueh)
> > > + goto unk;
> > > + if (gueh->control != 0 || gueh->version != 0)
> > > + goto unk;
> > > + /* Later we can support also IPPROTO_IPV6 */
> > > + if (gueh->proto_ctype != IPPROTO_IPIP)
> > > + goto unk;
> > > + *proto = gueh->proto_ctype;
> > > + return sizeof(struct guehdr) + (gueh->hlen << 2);
> >
> > I think that the gue-specific portions of the above, which is most of
> > ipvs_udp_decap() should be a separate helper which is part of net/gue.h or
> > net/ipv4/fou.c - this seems to be a subset of the logic in gue_udp_recv().
>
> Yes but fou.c strips the headers in all cases while
> in IPVS we should do it only when connection is found because
> we do not want to mess with non-IPVS traffic.
>
> >
> > > + }
> > > +
> > > +unk:
> > > + return -1;
> > > +}
> > > +
> > > /*
> > > * Handle ICMP messages in the outside-to-inside direction (incoming).
> > > * Find any that might be relevant, check against existing connections,
> > > @@ -1658,8 +1687,37 @@ ip_vs_in_icmp(struct netns_ipvs *ipvs, struct sk_buff *skb, int *related,
> > > if (cih == NULL)
> > > return NF_ACCEPT; /* The packet looks wrong, ignore */
> > > ipip = true;
> > > + } else if (cih->protocol == IPPROTO_UDP) { /* Can be UDP encap */
> >
> > Can we consider factoring out the logic in this else-if clause
> > out into a function to avoid the use of goto?
>
> We can even safely return NF_ACCEPT instead of the goto
>
> >
> > > + struct udphdr _udph, *udph;
> > > + __u8 iproto;
> > > + int ulen;
> > > +
> > > + if (unlikely(cih->frag_off & htons(IP_OFFSET)))
> > > + return NF_ACCEPT;
> >
> > I think a comment is warranted regarding the behaviour implemented
> > for fragments.
>
> I'll add in v2.
>
> >
> > > + /* Error for our tunnel must arrive at LOCAL_IN */
> > > + if (!(skb_rtable(skb)->rt_flags & RTCF_LOCAL))
> > > + return NF_ACCEPT;
> > > + offset2 = offset + cih->ihl * 4;
> > > + udph = skb_header_pointer(skb, offset2, sizeof(_udph), &_udph);
> > > + if (!udph)
> > > + goto check;
> >
> > > + offset2 += sizeof(struct udphdr);
> > > + ulen = ipvs_udp_decap(ipvs, skb, offset2, AF_INET, udph->dest,
> > > + raddr, &iproto);
> > > + if (ulen < 0)
> > > + goto check;
> > > + /* Skip IP + UDP + ulen */
> > > + offset = offset2 + ulen;
> > > + /* Now we should be at the original IP header */
> > > + cih = skb_header_pointer(skb, offset, sizeof(_ciph), &_ciph);
> > > + if (cih && cih->version == 4 && cih->ihl >= 5 &&
> > > + iproto == IPPROTO_IPIP)
> > > + ipip = true;
> > > + else
> > > + return NF_ACCEPT;
> >
> > Skipping over tunnel headers and determining the inner IP protocol really
> > feels like something that should be handled code common to other UDP
> > encapsulation implementations.
>
> We can easily add simple FOU in ipvs_udp_decap() by
> returning 0 and correct *proto. Or you prefer to use common
> code from other files to parse the headers? Currently, there
> is no any GUE func that can be used for this.
My feeling is that using common code, even new common code, would
be better.
