> At present we have two IP interfaces of note: the loopback interface > and logical interfaces (bge0:1, hme0:3, etc.)
I'm not sure how to parse this. There are some IP interfaces that are IFF_VIRTUAL. Loopback is one; vni is another. All IP interfaces can have more than one address hosted on them, represented as logical interfaces. > The presence of these creates a collection of interaction issues - we > can't snoop on them, can't use ipfilter on them, etc. To address these > issues we need to add more special case code into IP. Starting with Clearview, you will be able to snoop on any IP interface. That will include all of the addresses that are hosted on it. > If we're going to go forward pushing the idea that IP interfaces > are independant of network interface drivers, should we have > an architecture that allows them to be developed/used? I'm not sure what this means. Since Solaris 2.0, IP interfaces have been independent of the underlying device driver. However, IFF_VIRTUAL IP interfaces need to have those semantics implemented inside IP. We could have some sort of callout API from IP, but I'm not sure what problem that would actually be solving. Are there third parties who want to implement new flavors of IFF_VIRTUAL IP interfaces? > Is there merit in building an interface that not only 3rd parties > can use to develop IP interfaces but also to better integrate > these with networking tools (such as snoop) and the bigger > networking picture? We've already spent a fair amount of time working through the intended semantics for IP interfaces with Clearview, and it's the essence of what the project is about. Reworking the logical interface abstraction (which is not part of Clearview) to be less confusing has been discussed on and off and there is general consensus that work is needed here, but that it's unclear what we can do in a backwards-compatible manner. It needs more investigation. -- meem _______________________________________________ networking-discuss mailing list [email protected]
