On Tue, Jan 20, 2015 at 12:18:59AM +0100, Thomas Graf wrote: > On 01/19/15 at 09:19am, Ben Pfaff wrote: > > I really prefer feature negotiation in any case, because it means that > > OVN can work with OVS that has custom version numbers or backports of > > individual features. > > You are absolutely right. Thinking about this more, I guess we don't > even have a choice at all. Regardless of the OVS user space version, the > datapath could support any sub or superset of that.
Userspace tries pretty hard to work around datapath limitations, though, so I think the actual restrictions are pretty minimal. > > The one feature that I'm most interested in using from OVN, though, is > > one that will only be in OVS 2.4: conjunctive match. If you look at the > > matching language that I want to implement in OVN, it can be naturally > > expressed with conjunctive match and it's much more difficult to > > implement efficiently and in full generality without it. That probably > > means that for older versions we'd have to compute full cross-products, > > which is OK for simpler kinds of conjunctive match. I think that means > > that, if OVN is used with OVS<2.4, then OVN users would have to restrict > > the ACLs they add to ones that can be efficiently implemented with > > cross-products. Maybe that's OK; I haven't thought it through fully. > > I realize this should probably be avoided at all cost but at some > point the only answer out for certain requests might be "operation not > supported" or "flow too complex" errors reported back. The main use that I expect for conjunctive flows that would produce large crossproducts is for end user specified ACLs, which wouldn't be hard to restrict. It would also be straightforward to explain the restriction to end users. _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
