I really don't think this is important enough to retroactively apply to the 1.1 branch. We would want to apply it to master instead, it would get picked up in the next release. Please rewrite it against master and resend the patch.
rule_create() needs to initialize rule->resubmit to true to handle the possibility of rules that have not had their action's xlated at the time of flow_push_stats(). We should probably do the same for modify_flow(). Your patch should set the resubmit field of the intermediate rules in the resubmit tree. Currently if rule A resubmits into rule B, and B has no facets, this patch won't set the resubmit flag of B even if it does in fact have resubmit actions. I liked Ben's idea of having validate_actions() return a flag indicating resubmits. I think it would be a bit cleaner then forcing xlate_actions() to populate the rule->resubmit and dealing with the possibility that xlate_actions() hasn't been called on a given rule at a given time. I think the patch would be a lot simpler and easier to think about this way. This approach is probably fine though. > I've just checked that the patch works, I'll evaluate the > performance later on. One disapointement is that, while it did cut the > number of calls to xlate_actions(), there remains still a large number > of them. This isn't surprising to me as flow_push_stats() is only called once per second on a given flow, and only on flows which have handled traffic. Have you seen an actual performance improvement? This feels like premature optimization to me. Ethan On Tue, Jun 7, 2011 at 09:25, Ben Pfaff <b...@nicira.com> wrote: > On Mon, Jun 06, 2011 at 06:20:02PM -0700, Jean Tourrilhes wrote: >> Just for curiosity, I've implemented the patch along the line >> of Ben suggestions, see below. I'm wondering if the compiler is smart >> enough to pack bools, if not, I may use bitfields... > > Adding a "bool" there should not increase the size of the structure > since there is already some padding. > > Bit-fields usually make code slower since the compiler isn't very good > at optimizing access to them. > > Your patch looks good to me. Ethan, would you mind taking a look too, > since you've been thinking about this? If it looks good to you then > I'll commit it to master. > _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
