Mike McGrath wrote:
We've got issues. Many of our xen guests are experiencing tcp checksum
corruption. Research on the net suggests ethtool -K eth0 tx off will
fix the issue but it hasn't in our case. I've tried the koji guest on
many physical hosts, same issue. It's been on two FC6 dom0 hosts and a
RHEL5 dom0 host.
For now I'm going to stick koji on app1 as a physical host. Just
because that box has gotten to the point where it becomes unusable and
we cannot have that happening right now.
I've contacted some internal xen developers, some guys in ##xen on
freenode and the Fedora-xen list. Anyone who is able I'd encourage you
to take a look as well. Simple test:
"tcpdump -vvv | grep incomplete"
You'll see:
06:34:20.135530 IP (tos 0x0, ttl 64, id 39370, offset 0, flags [DF],
proto: TCP (6), length: 40) nat-pool.fedora.phx.redhat.com.49457 >
publictest8.fedora.phx.redhat.com.http: P, cksum 0x76af (incorrect (->
0x6100), 601:601(0) ack 6801 win 597
If worse comes to worse I know a workaround for this that eliminates the
problem, but also adds a bit more inconvenience. It involves creating a
private NAT network on each Xen host, with the extra real IP on the
hosts real ethernet device, and NAT.
Warren
