On Wed, May 30, 2007 at 12:52:15PM -0400, James Carlson wrote: > Nicolas Williams writes: > > The RFC explains why it's bad to manually configure LLAs. And though it > > doesn't say MUST NOT, it does say SHOULD NOT, and that provides enough > > cover. And while this SHOULD NOT does allow us to allow manual > > configuration of LLAs, we ought to have a good reason to ignore that > > SHOULD NOT, and any reason rejected by the IETF is not a very good > > reason for us to use for justifying an opposite decision. > > I think that misses both the intent and the language of section 1.6. > > The intent of that section is to prohibit *administrators* from using > the LLA subnet for their own allocation purposes. They shouldn't try > to treat it as roughly equivalent to RFC 1918, which many will likely > be tempted to do. The implementations might not have the right > duplicate detection support to make such usage safe (fortunately, not > an issue with Solaris), and there are special problems that can occur > if the addresses are (mis)configured into a DHCP server.
Except that normative RFC2119 language is supposed to be aimed at implementors -- administrators don't read RFCs, or shouldn't have to anyways. But yes, the language is not clearly aimed at implementors. One might argue that it tells implementors to tell users not to do this or that. But I think the authors probably meant that the implementors shouldn't allow this or that. We could ask them, or read the mailing list archives. But I think we can avoid that. > The one point of contention I see here is whether LLA itself is > defined by the use of the subnet (as Erik has been arguing) or if it's > defined by the subnet plus the random-probe-and-assign mechanism (as > Kacheong is saying). The latter definition is tighter and allows for > the (mis)use of the LLA subnet as a compatible evolution of the > system, but I can see how others would assume that the subnet itself > is architecturally defined (by the IETF) to be off limits. I'm in that last camp. I don't mind letting users manually configure LLAs, but I think that mixing manual LLA assignments with automatic LLA assignment should be declared to have undefined results (the ifconfig lla may fail, the manual LLA interfaces may be ifconfig'ed inet 0.0.0.0, etc...) -- if you want manual LLA don't be surprised that automatic LLA doesn't mix with it. > In any event, I don't think this is all that serious an issue, as any > of the alternatives in this area that've been proposed would work > fine. Agreed, it's not. I see you've derailed the case, but I think it's close to agreement anyways. Nico --
