Question on address configuration

Robert Elz kre@munnari.OZ.AU
Mon, 14 Jan 2002 21:22:30 +0700


    Date:        Thu, 10 Jan 2002 09:48:08 -0500
    From:        Edward Lewis <lewis@tislabs.com>
    Message-ID:  <v03130301b86356715630@[199.171.39.21]>

  | Your points up to this are accurate.  But I wouldn't agree that there is no
  | relationship to finding DNS servers.

You're arguing yet another point, unrelated to either of the other ones.
Or perhaps slightly related to the original, but not at all to the
other one.

  | Reading the replies and thinking more about this situation (and why two of
  | us IPv6 newcomers found this), we realized that this problem is rather
  | DNS-centric.

No, that's just one use.

  | Any other server out there can use autoconfig with the MAC
  | address and rely on DNS for name to number mapping.

Sure, they can.   But it isn't an issue of who can do what, but who
must do what.   I can use MAC address based tokens for my systems, and
usually, that's what I'll do.  But I have a choice, I can use anything I
like, any numbering scheme I prefer, and if I choose to number my systems
1 2 3 ... in the expectation that I can keep doing that for eternity,
using a new unique number for for each new system (or even redefined system)
without ever running out of the 64 bit space, then I should be permitted
to do just that.   I might need to use a DHCP server to make my admin
load manageable of course, or I might just configure the number in the
system if I have a centralised systems receiving dept, where everything
is delivered and locally configured before being sent to the end user.

If I'm using a DHCP server though, at the very least I'm going to have to
configure its address, not because it needs to be stable, but because it
has no DHCP server to request an address from.   And it can't simply use
its MAC address, as someday that might clash with my 1 2 3 ... (after I
eventually have to give up on keeping the meaning of the "global" bit...)

In any case, this is an issue of choice, in IPv6 we have 3 methods of
address generation - autoconfigured from MAC address, assigned by a
DHCP server, and manually configured.   We shouldn't be causing the
3rd one to be harder than it should otherwise be, nor should we be
causing other problems (like making it much harder to renumber a network)
just because people would like to manually configure the part of their
addresses that they get to assign (in both v4 and v6 you're stuck with
the prefix offered to you if you want to interoperate, but in both you
can be as imaginative as you like with the remaining bits).

Certainly DNS servers, DHCP servers, routers, and even SMTP servers (thanks
all the same, but I don't want a 2 day mail dead spot while my old DNS
records TTL times out...) are the most likely candidates for configured
addresses, but any host can have it.   Implementations that I'm aware of
all allow this already - all they don't do is the compromise where the
prefix comes from RA messages (just like fully autoconfigured addresses),
and the token comes from configuration.   This really should be added.

kre