Erik Nordmark wrote:
> Makes no sense.
> The behavior of 169.254/16 is standardized in RFC 3927.
The behavior of IPv4 LLA for one interface is standardized.
But the behavior of using IPv4 LLA for more than one interface
is not defined. While we don't know how to make it work
automatically, it does not mean that a careful manual set up
cannot be done.
> Given that Solaris hasn't yet implemented that RFC, whatever behavior
> somebody sees when configuring the system with such an address is
> undefined.
> As part of implementing that RFC we should make the behavior of such
> addresses conform with the RFC. It doesn't make sense to allow the user
> to get the old (undefined) behavior as an option - nobody would want that.
Are you saying that IPv4 LLA with more than one interface
is well defined? One may argue that it may not work in
every situations, but it does not mean that it should be
prohibited.
> My parenthetical remark about Bonjour was not the concern. The text
> before the parenthesis is my concern.
> The architectural invariant I'm concerned about is the relationship
> between what gethostbyname/getaddrinfo returns for "foo" on a remote
> system, and on system "foo" what an application will see with
> SIOCGLIFCONF. You are proposing to change that invariant with this case.
Note that the proposal will still show the LLA interface
if a flag is set. But the question is whether most apps
actually use SIOCGLIFCONF. It seems that we are looking
at a small set of apps. And for this set of apps, since
they are already using SIOCGLIFCONF, I guess they can be
changed to better cope with IPv4 LLA.
And is it true that the above is really an invariant?
Suppose an app A in a remote host using gethostbyname() to
lookup a name X via DNS and A gets a list. Can we really
say that the machine named X have that list of addresses up
all the time when A does a gethostbyname()?
> You are proposing to change an architectural invariant for IP addresses
> and names.
There is a fixed mapping between IP address and name. But I
don't think it is an invariant that there must be an interface
using a particular IP address all the time.
> Changing the default behavior of SIOCGLIFCONF to return LLAs, as you've
> correctly pointed out, causes a different set of problems.
>
> So you are stuck between the proverbial rock and a hard place.
What I've been trying to do is to minimize the impact. We
know that there is an impact. So what we need to do is to
make it safe. I think the current proposal meets this criterion.
--
K. Poon.
kacheong.poon at sun.com
