[ovs-dev] [BUG] SLB bonding & bond-rebalance-interval not working as expected

Zoltan Kiss zoltan.kiss at citrix.com
Fri Feb 8 20:16:23 UTC 2013


Hi,

On 30/01/13 14:44, Markus Schuster wrote:
> Hi,
>
> looks like I'm currently in SLB bonding bug hunting mood :) - I think I found
> an additional bug / strange behaviour with balance-slb bonding: Two VMs
> suffered from short term (a  few minutes) connectivity issues every now and
> then, so I started digging further into the issue. First thing I noticed is
> the MAC of the VM jumping between ports on the uplink switches like mad - I
> had to think of my other bug report [1] - but this time I saw no broad- or
> multicast frames but normal unicast frames sent out on both member ports of
> the bond for a few minutes every now and then.
>
> Long story short: XCP 1.6 configures SLB bonds to rebalance their traffic
> every 30 minutes. And it looks like Open vSwitch sometimes fails in migrating
> certain flows from one interface to the other. That causes some traffic to be
> sent via the "old" interface and some via the "new" interface.

This is a known problem in OVS 1.4.2 (the version released with 
XenServer 6.1): the flows are not properly tagged during rebalancing in 
userspace, and therefore the change doesn't propagate to the kernel flow 
tables. This was fixed by this backported patch in 1.4.3:

http://openvswitch.org/cgi-bin/gitweb.cgi?p=openvswitch;a=commitdiff;h=865f22b3b3cb953c48ed30dd21f16ea3dd53f04c

And the algorithm was also quite ineffective, it was fixed in 1.4.4:

http://openvswitch.org/cgi-bin/gitweb.cgi?p=openvswitch;a=commitdiff;h=ba60b421cc159b7e53f7e9a890aa130a35801083

There isn't any XS 6.1 public hotfix released for this.

Regards,

Zoltan Kiss



More information about the dev mailing list