[rhos-list] Quantum: Packet Fragmentation Issue?

[Thomas Graf]

On 06/25/2013 09:07 PM, Gary Kotton wrote:
> On 06/25/2013 08:54 PM, Johnray Fuller wrote:
>> Hello,
>>
>>
>> I appear to have an issue with packet fragmentation
>>
>> When we try to ssh from one VM to another where the VMs run on
>> different hosts on the source host the physical link (eth4) shows:
>>
>> 13:24:19.520812 fa:16:3e:dd:3c:4e (oui Unknown) > fa:16:3e:6c:eb:80
>> (oui Unknown), ethertype 802.1Q (0x8100), length 78: vlan 2, p 0,
>> ethertype IPv4, (tos 0x0, ttl 64, id 31721, offset 0, flags [DF],
>> proto TCP (6), length 60)
>> 10.0.0.11.60025 > 10.0.0.12.ssh: Flags [S], cksum 0xdb91 (correct),
>> seq 3087184310, win 14600, options [mss 1460,sackOK,TS val 11061855
>> ecr 0,nop,wscale 6], length 0
>>
>> While on the receiving end we see:
>>
>> 13:24:20.555105 3e:6c:eb:80:fa:16 (oui Unknown) > 3c:4e:fa:16:fa:16
>> (oui Unknown), ethertype Unknown (0x3edd), length 78:
>> 0x0000:  3c4e 0800 4500 003c 7be9 4000 4006 aabc <N..E..<{. at .@...
>> 0x0010:  0a00 000b 0a00 000c ea79 0016 b802 b1b6  .........y......
>> 0x0020:  0000 0000 a002 3908 db91 0000 0204 05b4  ......9.........
>> 0x0030:  0402 080a 00a8 ca5f 0000 0000 0103 0306  ......._........
>>
>
> This looks like it could be related to VLAN splinters when using
> openvswitch. Are you using Openvswitch? Maybe
> http://git.openvswitch.org/cgi-bin/gitweb.cgi?p=openvswitch;a=blob_plain;f=FAQ;hb=HEADcan
> help.

This is not related to VLAN splinters but VLANs usage in general.

We have seen this before and it usually is caused by an intermediate
device on the host having the same MTU as the interface inside the VM.
Typically both have 1500. The VM outputs 1500 sized frames, OVS adds
a VLAN header and that exceeds the MTU of any device on the host.

Fix is to either

a) decrease the MTU inside the VM by 4 if VLAN tagged packets are to
    go out to a physical ethernet

b) increase MTU of all intermediate interfaces on the host by at least
    4 to avoid fragmentation.

c) increase MTU of all soft devices on the host and enable jumbo frames
    on the physical ethernet device.

I would choose b) for Neutron if tunneling is being used. If external
VLANs are in play option c) is nice with a fallback to frags if jumbo
frames are unsupported.



>
>> It seems that encapsulation causes the packet to break. Does anyone
>> have any ideas on how to troubleshoot this?
>>
>> These VMs are on different hosts.
>>
>> We tried increasing the mtu on both hosts' eth4, but still no joy.
>>
>> We found the following, https://review.openstack.org/#/c/31518/ ,
>> which might be related, but this patch was abandoned.
>>
>> Any assistance would be  greatly appreciated.
>>
>> J
>>
>