[ovs-dev] [PATCH net 0/2] vxlan: Set a large MTU on ovs-created vxlan devices

[Jesse Gross]

On Wed, Jan 6, 2016 at 4:14 PM, Hannes Frederic Sowa
<hannes at stressinduktion.org> wrote:
> Hi,
>
>
> On 07.01.2016 00:57, Jesse Gross wrote:
>>
>> On Wed, Jan 6, 2016 at 3:25 PM, David Wragg <david at weave.works> wrote:
>>>
>>> David Miller <davem at davemloft.net> writes:
>>>>>
>>>>> Prior to 4.3, openvswitch vxlan vports could transmit vxlan packets of
>>>>> any size, constrained only by the ability to transmit the resulting
>>>>> UDP packets.  4.3 introduced vxlan netdevs corresponding to vxlan
>>>>> vports.  These netdevs have an MTU, which limits the size of a packet
>>>>> that can be successfully vxlan-encapsulated.  The default value for
>>>>> this MTU is 1500, which is awkwardly small, and leads to a conspicuous
>>>>> change in behaviour for userspace.
>>>>>
>>>>> These two patches set the MTU on openvswitch-crated vxlan devices to
>>>>> be 65465 (the maximum IP packet size minus the vxlan-on-IPv6
>>>>> overhead), effectively restoring the behaviour prior to 4.3.  In order
>>>>> to accomplish this, the first patch removes the MTU constraint of 1500
>>>>> for vxlan netdevs without an underlying device.
>>>>
>>>>
>>>> Is this really the right thing to do?
>>>
>>>
>>> I'm certainly open to suggestions of better ways to solve the problem.
>>
>>
>> One option is to simply set the MTU on the device from userspace.
>>
>> The reality is that the code you're modifying is compatibility code.
>> Maybe we should make this change to preserve the old behavior for old
>> callers (although, again, it should not be just for VXLAN). But no new
>> features or tunnel types will be supported in this manner.
>>
>> New or updated userspace programs should work by simply creating and
>> adding tunnel devices to OVS. That won't go through this path at all
>> so you're going to need to find another approach in the near future in
>> any case.
>
>
> I don't see any other way as to make MTUs part of the flow if we want to
> have correct ip_local_error notifications. And those must also work across
> VMs, so openvswitch in quasi brouting mode would need to emit ICMP PtBs
> (hopefully with a correct source address, otherwise uRPF kills them before
> reaching the applications) or do error signaling via virtio_net.

I actually implemented this a long time ago and then there was some
additional discussion on this about a year ago. I agree it's the right
solution overall but it's not entirely clearly to me how to get the
details correct.

> Either the openvswitch user space can feed those information to the datapath
> or the ovs dataplane can do a lookup on the outer ip address while filling
> out the metadata_dst and caching it in the flow or we just keep the dst in
> the flow anyway. So a net_device used by ovs has no real mtu anymore.

I agree that the concept of MTU is much more complicated than a single
number on a device, we just have to find the right way to model it.