Re: [Xen-users] domU ignores MTU

[Bartosz Lis <bartoszl@xxxxxxxxxxxxx>]

Hi Cristoph,

I had the same problem with some combinations of:
hypervisor: xen 3.1.3 or 3.0.4, 
Dom0: linux 2.6.22 patched with xen 3.1
DomU: linux 2.6.19 patched with xen 3.0.4 or linux 2.6.22 patched with xen 3.1

The problem does not occur with: hypervisor: xen 3.0.4, Dom0&DomU: linux 
2.6.19 patched with xen 3.0.4.

I think the problem is caused by offloading TCP stream for segmentation from 
TCP/IP stack to hardware controller (TCP Segmenttion Offload - TSO). 
Generally this should enhance performance, but when there are no hardware 
controller (in domU) and instead we have a virtualization software and a 
virtual ethernet bridge (in dom0) things get complex.

DomU virtual NIC driver reports to DomU TCP/IP stack that it is TSO capable. 
Being so informed, TCP/IP stack does not bother very much in segmenting TCP 
stream, ignores MTU and passes as large as possible datagrams to NIC driver 
for proper segmentation. Developers, please correct me if I'm wrong: Xen NIC 
drivers don't do segmentation, they pass large datagrams as they are to Dom0 
bridge. If the datagram is addressed do dom0, that's OK. But when the 
datagram should go outside it cannot fit in MTU of Dom0 NIC, so it is dropped 
and an ICMP packet is sent to DomU asking to use smaller packets. This causes 
retransmissions, delays and a very slow transfer.

The problem could be solved by entering in domU:

/sbin/ethtool -K eth0 tso off

This makes DomU TCP/IP stack to segment packets according to MTU and should 
not prevent from sending jumbo packets if one sets a "jumbo" MTU. Data 
transfer rate between Dom0 and DomU probably gets slightly degraded.

I think, the problem should be addressed somewhere in Dom0 bridge or Xen 
virtual NIC code/configuration. If someone knows how to do that or better 
understans the problem, please let me know.

Kind regards,

-- 
Bartosz Lis @ Institute of Comp. Science, Technical University of Lodz, Poland
   bartoszl @ ics.p.lodz.pl

_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-users