On Fri, 15 Apr 2016 18:30:59 -0300, Thadeu Lima de Souza Cascardo wrote: > Jiri has suggested that we require users to create the interfaces themselves, > by > using whatever method their OS has, and add them as netdev devices. That would > still require fetching some of the configuration from the device in order to > make it properly work with flow-based tunnels. In fact, if we set the remote > or > local IP addresses on those devices, this would require multiple interfaces > instead of only one just to be able to specify the same level of configuration > as ovsdb allows us to.
No need for multiple interfaces. We have 'external' parameter for 'ip link add' that will create the interface in metada mode. The reason I'm suggesting this is it solves the problem with cleanup in unexpected conditions (What happens with the interface automatically created by ovs when ovs crashes? What happens if user creates an interface named vxlan_sys_*? Relaying on a particular interface name doesn't look like a good design to me.) and it allows easy support of future vxlan features without modifying ovs. This would require querying the kernel whether the interface is in metadata mode on vport addition, sure. But that's very easy to do and I think it's the only configuration parameter that we need to know in ovs. The existing code that allows adding the ports through the compat genetlink interface would stay, of course, for backwards compatibility. To use new vxlan feature, the new way of setup would be needed. Jiri _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
