On Thu, Aug 10, 2017 at 02:41:30PM +0000, Mark Michelson wrote:
> I'm curious about the current behavior when ovn-controller originates a
> packet and transmits it over a link-local IPv6 multicast address. Will the
> packet be delivered only to nodes on the same chassis, or can the packet
> also reach on-link nodes on other chassis?

When an OVN logical switch receives an Ethernet broadcast or multicast
packet from a given port, it forwards it to all of its other logical
switch ports, regardless of chassis.  There are special cases for ARP
and IPv6 ND, to avoid forwarding them across chassis for known IP
addresses.  I don't think that we have any special cases for IPv6
link-local multicast.  Limiting these within a chassis would probably
require new mechanisms, because logical flows, by design, do not respect
physical boundaries.

> The reason I ask is that I'm working towards expanding OVN's IPv6 neighbor
> discovery capabilities, and the first step on that path is being able to
> send periodic router advertisements (RAs). In a traditional physical
> network, each interface on a router may send periodic RAs to its link-local
> all-nodes address (ff02::1). This way all on-link nodes receive the RA, and
> the world is a happy place.
> OVN is not a traditional physical router, though :)
> With OVN, each hypervisor has an ovn-controller instance running, blurring
> the lines about what exactly constitutes a "router interface". From an
> ovn-nb point of view, a logical router port constitutes a router interface.
> However, in practice, I'm not sure if this holds true once you get down to
> the ovn-controller level. Let's say that a single link (via a logical
> switch) is spread across three hypervisors. If the ovn-controller on each
> of those hypervisors sends an RA to the link-local all-nodes address, would
> that result in nodes on each hypervisor receiving three RAs? Or would each
> node only receive one RA since it only receives the RA from the local
> ovn-controller?

I believe that this would result in 3x advertisements.

> I suspect the answer to this all comes down to how openflow flows get
> installed on each OVS instance. What I'm curious about is the current
> behavior of OVN in this regard.

I guess that it would be more efficient for each chassis to individually
generate these router advertisements and ensure that they do not
propagate beyond the physical chassis boundary?  I can think of ways to
do that, especially if ovn-controller were generating them.  This is
because when ovn-controller generates packets, it can tag them with
additional metadata that indicate the purpose.  For example, to allow it
to send a chassis-limited packet, we could implement a new metadata bit
that suppresses output through tunnels, and ovn-controller could set
that bit.

> P.S. Someone please tell me if ovs-discuss would have been a more suitable
> list for this question. I figured since it was related to a feature I was
> developing that this would be the proper list, though.

Seems like a legit dev discussion to me.
dev mailing list

Reply via email to