On 2/3/21 12:06 PM, Numan Siddique wrote:
On Tue, Feb 2, 2021 at 3:15 PM Dumitru Ceara <dce...@redhat.com> wrote:The supported upgrade procedure is to always upgrade ovn-controllers first and OVN DBs and ovn-northd later. This leads however to the situation when ovn-controller might try to set the Port_Binding.up field while the Southbound DB isn't yet aware of this field which would trigger transaction failures and control plane/data plane outages. To avoid such situations ovn-controller only sets the Port_Binding.up field if it was explicitly set to 'false'. This ensures that the SB database was already upgraded. On the ovn-northd side, as soon as ovn-northd is upgraded it will update all existing Port_Bindings and explicitly set 'Port_Binding.up' to false, implicitly notifying ovn-controller that it is safe to write to the field. Reported-by: Numan Siddique <num...@ovn.org> Fixes: 4d3cb42b076b ("binding: Set Logical_Switch_Port.up when all OVS flows are installed.") Signed-off-by: Dumitru Ceara <dce...@redhat.com>Thanks Dumitru for addressing this issue.
Thanks for the review!
I see a problem with the approach taken here when the central nodes are upgraded first. I do understand that OVN recommends updating/upgrading ovn-controllers first. But from what I have seen, Openstack tripleo (and possibly Openshift too) update the central nodes first. Suppose if ovn-northd and DBs are upgraded first, then after this patch, ovn-northd sets the logical_switch_port.up to 'down' for all the logical ports until the ovn-controllers are upgraded. This could cause some control plane issues. Like Openstack neutron may notify Nova service of vif-unplugged events. We could see this issue even if ovn-controllers are configured with - "ovn-match-northd-version=true". Another approach to solve this problem would be - ovn-controller will not set the port_binding.up to true if the internal version in SB DB is lesser than the version in which this new column - port_binding.up was added. What do you think ?
An alternative is to add a "features" column to SB.Chassis. Newer ovn-controllers could register there that they support a specific feature, e.g., "port-up-notification".
In northd we could check the feature support for chassis that claimed a port binding and based on that take into account Port_Binding.up when setting LSP.up.
Does this sound ok? Thanks, Dumitru _______________________________________________ dev mailing list d...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-dev
