Description of problem: Wondering why Xen's Virtual Interfaces virbr0 vifX.0 and on the Guest OS the vif which eth0 etc map to have tx-checksumming turned on? given that these are not actual physical NIC Cards which would/could offload the checksumming from the CPU?? The tx-checksumming was giving us big problems with SSH and FTP to our Xen Guest and we ultimate had to add an ethtools command to our rc.local to turn it off. Steps to Reproduce: 1. ethtool -k eth0 see that tx-checksumming is on you can see that is is on a virtual interface. [root@dropbox ~]# cat /etc/modprobe.conf alias eth0 xennet alias scsi_hostadapter xenblk Expected results: Since Xen bridge, virtual interfaces, and on guestOS virtual interfaces are not physical NIC Cards which could legitimately offload the checksumming from the CPU I would have expected that they would have tx-checksumming off by default.
The Xen VIF kernel driver *does* support tx-checksumming, in so much as it will pass packets straight through to the host NIC for checksumming & fragmentation. For guest<->guest traffic this allows checksumming/fragmentation overhead to be avoided entirely. This results in a throughput improvements over > x100. If you have specific scenarios in which tx checksumming causes problems please enter a bug providing details of exactly what is broken & the steps to reproduce the problem.