On Thu, Jul 06, 2006 at 04:14:31PM +0200, Florian Heigl wrote:
> Hi Ari,
> 
> On Thu, Jul 06, 2006 at 03:58:22PM +0300, Ari Suutari wrote:
> > 
> > 	Thank you ! Network works perfectly now.
> > 	I should have read the FAQ first instead of
> > 	assuming that this is netbsd -related.
> 
> Noone can expect this error, it took me two weeks to find out the issue,
> not by debugging, but by at last giving up and searching for "xen bridge
> packet loss network issue" (and a few swear words to boil down the results)

Hrm, I caught it after a tcpdump showed me the bad checksums
and googled for that.

The lesson learned "Even if your network doesn't physically exist, tcpdump
can be useful."
> 
> > 	All debian stuff was/is installed via apt-get, and
> > 	on netbsd dom0 xen is from pkgsrc (I assume it is unmodifed)
> 
> hmmm, then there is still some space to look at in relation to what
> our bridge interface does when a linux kernel turns on tcp offloading.

Linux does a few things on packet optimization that I would describe as
pathological.  DaveM would probably say it's efficient.

Since folks will get burned with a netbsd domU on a linux dom0.
You still need to disable tx checksum offloading on dummynets.
If you attempt this prior to xen starting up the bridge, it will
fail to create the bridge (And also keep you from removing the bridge).
I currently have a kludge to do the setting after bringing up the dummynet 
bridge (dual bridge configuration), but it seems to have a race condition,
and doesn't work after every reboot.

For the bridge that goes to a real ethernet adaptor, I leave the tcp
transmit checksumming off, but remind myself that I won't be able to
effectively reach the dom0.
Someone that has poked more than me might be able to tell me if I can
force tcp checksuming on the specific veths and vifs so that the most
efficient combination is chosen.
> 
> I have no idea of the real internals of xen :)

Or when it's an OS internal that is doing things to make life
miserable for doms under Xen.
> 
> flo
> 
> -- 
> 
> florian heigl			 http://deranfangvomen.de/
> 
> Grid Computing erfreut sich gerade bei Windows-Nutzern sehr regem Zuspruch, auch
> wenn die Rechnerbesitzer meist nichts von ihrem Glück wissen. --f. weimer

-- 
Chris Dukes
"The key to effective management is properly timed hovering."