[ovs-dev] [PATCH] xenserver: make ovs-xenserverd startup configurable and disable it for 5.6.0 onwards.

Ben Pfaff blp at nicira.com
Thu Jun 3 20:31:10 UTC 2010


On Thu, Jun 03, 2010 at 08:40:46PM +0100, Ian Campbell wrote:
> On Thu, 2010-06-03 at 20:17 +0100, Ben Pfaff wrote: 
> > I've unfortunately had a report here, which I've confirmed, that network
> > UUIDs are not being updated on 5.6.810-31078p.  Is this build new
> > enough that you would expect ovs-xenserverd to be unnecessary, or
> > should we expect to update to a newer build first?
> 
> I would have expected this to work on any build since I added the call
> to set the network uuids for internal networks explicitly, which was
> ages ago (before the switch to 5.6.810 numbering I think). I thought
> xapi had always replugged all PIFs at start of day (Rob, is my
> understanding here correct?).
> 
> Do you know if it is pool join or eject (or both) which has the issue?
> is it all network types or just External/Internal/CHIN? Do the logs show
> interface-reconfigure or ovs-vsctl being called when xapi restarts after
> whichever event it was?

The issue that was reported to me was on pool join.  The server in
question does not have any internal or CHIN networks.  I doubt that
there would be a problem with pool eject, since reboot should fix this
kind of problem.

There's a big cluster of log messages in /var/log/xensource.log.1 with
Async.pool.join in them dated 20100602T01:14:00, and none elsewhere, so
I assume that the pool join happened then.  There is a corresponding
"restarting xapi in different operating mode" in /var/log/messages.1
dated Jun  1 18:14:13, which is the same time allowing for our GMT
offset.  I don't see any calls to interface-reconfigure or ovs-vsctl at
that time.  There are some calls to interface-reconfigure several hours
later to bring up the members of a bond, taking the bond itself down as
a side effect, and that interface-reconfigure call fixed the network
UUIDs for the bond members.

There were some networking problems on this box, as you can see from the
messages about connections to the master failing.  But they were not
serious enough to prevent XAPI on the slave from talking to the master
to bring down the bond, as mentioned above, so it's not obvious to me
why they would be serious enough to prevent interface-reconfigure being
called when XAPI restarted.

I've made those two logs available at:
        http://benpfaff.org/~blp/join-logs.tar.gz
I could always run a full "server status report", of course, if it would
be more helpful.

> Hopefully we can find a way to fix this on the xapi side but the
> fallback would likely be to add an network-uuid update to the
> openvswitch-cfg-update plugin. In any case I don't want a loop
> polling /etc/xensource-inventory in the version integrated into
> XenServer.

I don't want that either.

Thanks!

Ben




More information about the dev mailing list