[ovs-dev] [PATCH RFC 0/3] - send garp on localnet

[Ramu Ramamurthy]

>
> On a separate but related topic, northd programs localnet to not respond to
> regular
> arp requests coming from the network via a localnet port.
>
> How can IP traffic be initiated from a VM outside OVN control or from the
> physical network to a VM inside OVN controlled HV, connected via a localnet
> port ?
>

Traffic initiated from VM1 outside OVN to VM2 inside OVN would flow as follows:

VM1 would first arp for VM2's IP. This arp is a broadcast, and would
enter br-int via the
localnet patch on ALL hypervisors connected to that localnet. br-int
"floods" that arp
to all locally attached VIFs on that localnet datapath. VM2 happens to
be one of those VIFs.
VM2 receives that arp, and replies back.

As a result of the above interaction, external switches would learn
(port->source mac) bindings
and the external router would learn the arp mappings (IP->mac)

The problem is that port->source mac binding is cached for 5 minutes
at switches, and the arp
binding is cached for 4 hours by some routers !!

During that period, consider the following 2 scenarios:

1) VM2 moved between hypervisors due to migration and stayed silent.
VM2 cannot be
accessed externally due to stale port->mac mappings for several
minutes. Eventually,
the mac-learning entry ages out and the switch treats VM2's mac as unknown and
floods it as unknown unicast.

OR

2) VM2 got deleted, and VM3 is booted within OVN, and VM3 has the same IP as VM2
BUT a different mac, AND VM3 stayed silent.
VM3 cannot be accessed externally due to stale arp mapping at a router
for several hours !.
Eventually, the router would ageout its arp entry, and then, the
router would ARP for VM3's IP,
and traffic to VM3 is successfully established.

So, the problem addressed by this code change ONLY handles the corner
cases of stale mac-learning
entries at switches or stale arp entries at routers.


> Are you depending on flooding the network with gARPs periodically (to handle
> new external hosts) for all VMs on all HVs inside OVN, attached to localnet
> ?

No, the gARP is sent only a few times when a logical port on localnet comes up
to handle the special corner case (of stale mac-learning or arp
caches) described above.