Re: IPv4 multicast and mac-vlans acting weird on 3.0.4+

[Eric Dumazet <eric.dumazet@gmail.com>]

Le mercredi 05 octobre 2011 à 13:19 -0700, Ben Greear a écrit :
> On 10/05/2011 01:17 PM, Eric Dumazet wrote:
> > Le mercredi 05 octobre 2011 à 13:09 -0700, Ben Greear a écrit :
> >> On 10/05/2011 12:54 PM, Eric Dumazet wrote:
> >>> Le mercredi 05 octobre 2011 à 09:46 -0700, Ben Greear a écrit :
> >>>> This is on a hacked 3.0.4 kernel...
> >>>>
> >>>> I am seeing an issue where an IPv4 mcast receiver will not receive
> >>>> a 1473 or larger byte mcast message, but will receive a 1472.  The difference
> >>>> being that 1473 ends up being two packets on the wire.  It works on
> >>>> 802.1Q VLANs, VETH interfaces and real Ethernet.  It does not work
> >>>> on a mac-vlan hanging off the VETH.
> >>>>
> >>>> I see packets received on the macvlan in tshark, and they appear correct.  No
> >>>> obvious errors in the macvlan port stats or netstat -s,
> >>>> and the 'ss' tool doesn't appear to support UDP sockets at all.
> >>>>
> >>>> So, I'm about to go digging into the code, but if anyone has any
> >>>> suggestions for places to look, please let me know!
> >>>>
> >>>
> >>> Well, problem is defragmentation and macvlan cooperation.
> >>>
> >>> Multicast messages are broadcasted on all macvlan ports.
> >>>
> >>> But IP defrag will probably deliver a single final frame.
> >>>
> >>> We probably need to handle defrag in macvlan before broadcasting to all
> >>> ports.
> >>
> >> I see packets get to this code in ip_input.c (line 467 or so),
> >> and that printk is mine of course.
> >>
> >> 	if ((dev&&  strcmp(dev->name, "rddVR10#0") == 0) ||
> >> 	    (dev&&  strcmp(dev->name, "rddVR10") == 0)) {
> >> 		printk("calling ip_rcv_finish through NF_HOOK, dev: %s, len: %i\n",
> >> 		       dev->name, skb->len);
> >> 	}
> >>
> >> 	return NF_HOOK(NFPROTO_IPV4, NF_INET_PRE_ROUTING, skb, dev, NULL,
> >> 		       ip_rcv_finish);
> >>
> >> But, the macvlan packets never make it to the ip_rcv_finish method.
> >>
> >> I do see a big and a little packet entering this code.
> >>
> >> I have no firewall rules that I'm aware of, though there
> >> is some conn-track logic (though not associated with the
> >> mac-vlan interface):
> >
> > Say you have 10 vlans on your eth0, how many times do you want one
> > incoming multicast frame being delivered to your application listening
> > on 0.0.0.0:port ?
> 
> How would it work for two Ethernet devices on the same LAN?  I'd
> say that mac-vlans should mimic that case.
> 
> And in my case, I'm binding hard to a device & IP address,
> so my app should get it once regardless.
> 

OK, but before frame being delivered to your app, it must be
re-assembled by net/ipv4/inet_fragment.c & net/ipv4/ip_fragment.c
machinery.

This machinery uses :

static int ip4_frag_match(struct inet_frag_queue *q, void *a)
{
        struct ipq *qp;
        struct ip4_create_arg *arg = a;

        qp = container_of(q, struct ipq, q);
        return  qp->id == arg->iph->id &&
                        qp->saddr == arg->iph->saddr &&
                        qp->daddr == arg->iph->daddr &&
                        qp->protocol == arg->iph->protocol &&
                        qp->user == arg->user;
}

All frames broadcasted (because of multicast code in macvlan) on vlans
have same saddr/daddr/protocol (and user).

So kernel will discard all redundant copies of frames and deliver one
copy only to upper stack.

Check commit 7736d33f4262d437c5 (packet: Add pre-defragmentation support
for ipv4 fanouts) for a possible hint :

We could perform the re-assembly in macvlan code, before doing the
"broadcast the frame on all ports" part.