On Wed, Oct 26, 2011 at 08:17:21PM -0700, Jesse Gross wrote: > Ben, what do you think of this approach of zeroing out the priority > (or any other comments)?
I'm a bit nervous about it. Really I was expecting this commit to reduce the number of special cases, but it adds a new suite of special cases to the classifier and ofproto-dpif.c. Some of the special cases would be less odd if they were pushed down further, e.g. cls_rule_init() wouldn't need to explicitly zero priority if flow_zero_wildcards() zeroed priority (which is reasonable if we consider priority a field that is always wildcarded). cls_rule_init_exact() would still need to zero it, but that ill-conceived function needs to die anyway (I have a branch that kills it off). I'm not sure why priority has to be 0 in classifier_lookup(), since by design classifier_lookup() ignores fields that are wildcarded and we can just consider priority to be always wildcarded. So, in the classifier, it seems to me that it would be more natural to either fully implement matches against priority, just for sake of generality and avoiding special cases, even though it increases the cost of lookup. Anyway, I've got a branch in the works that completely regularizes flow wildcard matching (e.g. every single bit in the flow structure becomes independently maskable) while reducing the cost; I don't however know when that branch is going to land. _______________________________________________ dev mailing list [email protected] http://openvswitch.org/mailman/listinfo/dev
