Thanks a lot Ben for your review! On Fri, May 19, 2017 at 12:16 AM, Ben Pfaff <[email protected]> wrote:
> On Wed, May 10, 2017 at 08:35:45AM +0000, Daniel Alvarez wrote: > > This patch introduces a new type of OVN ports called "localport". > > These ports will be present in every hypervisor and may have the > > same IP/MAC addresses. They are not bound to any chassis and traffic > > to these ports will never go through a tunnel. > > > > Its main use case is the OpenStack metadata API support which relies > > on a local agent running on every hypervisor and serving metadata to > > VM's locally. This service is described in detail at [0]. > > > > An example to illustrate the purpose of this patch: > > > > - One logical switch sw0 with 2 ports (p1, p2) and 1 localport (lp) > > - Two hypervisors: HV1 and HV2 > > - p1 in HV1 (OVS port with external-id:iface-id="p1") > > - p2 in HV2 (OVS port with external-id:iface-id="p2") > > - lp in both hypevisors (OVS port with external-id:iface-id="lp") > > - p1 should be able to reach p2 and viceversa > > - lp on HV1 should be able to reach p1 but not p2 > > - lp on HV2 should be able to reach p2 but not p1 > > > > Explicit drop rules are inserted in table 32 with priority 150 > > in order to prevent traffic originated at a localport to go over > > a tunnel. > > > > [0] https://review.openstack.org/#/c/452811/ > > > > Signed-off-by: Daniel Alvarez <[email protected]> > > Thanks for working on this! > > This seems reasonable, but I'm torn about one aspect of it. I'm not > sure whether my concern is a kind of premature optimization or not, so > let me just describe it and we can discuss. > > This adds code that iterates over every local lport (suppose there are N > of them), which is nested inside a function that is executed for every > port relevant to the local hypervisor (suppose there are M of them). > That's O(N*M) time. But the inner loop is only doing something useful > for localport logical ports, and normally there would only be 1 or so of > those; at any rate a constant number. So in theory this could run in > O(M) time. > > I see at least two ways to fix the problem, if it's a problem, but I > don't know whether it's worth fixing. Daniel? Russell? (Others?) > > Yes, I also thought about this but I don't know either if it's a problem or not. If we want to impose at most one logical localport per datapath then we would have O(M) time (mimic localnet ports) but that's something I didn't want to do unless you think it makes more sense. > Thanks, > > Ben. > _______________________________________________ dev mailing list [email protected] https://mail.openvswitch.org/mailman/listinfo/ovs-dev
