a possible solution to this problem that does not involve changing the
protocol would be using SRV dns records.
---------------
how it would work
=================
whenever an ssh connection is created, instead of looking up the A/AAAA
record of the hostname, _ssh._tcp.${hostname} is queried for SRV
records. if one exists, the request is resolved further by the usual
rules of SRV records; otherwise, an A/AAAA record is used.
a SRV record allows load balancing (when equal priorities are used),
fallback (with different priorities) and selecting the port. the latter
feature is what helps here: the server would have different daemons
running on different ports, and does name based dispatch in dns instead
of ssh.
how it (doesn't) affect security
================================
ssh is used to relying on insecure dns before creating a secured
connection. if an SRV record is used, the original hostname (without the
_ssh._tcp part, ie. what the user requested) should be used as the
default HostKeyAlias. this allows migration to and from plain dns hosts
without disturbing existing known_hosts files.
basically, treat the dns resolver like a rogue dns server. if it
redirects us to a place where the server can present the correct
fingerprint, we're fine, if not, we detect it.
of course, if more than one SRV record exists, the same host key has to
be present on multiple servers. this is the same situation as with
migrating servers where admins copy the old ssh key to the new machine;
whether this is acceptable or not is up to the administrator's
discretion, and most likely depends on how he distributes his
known_hosts data.
i don't know how this would interact with ssh keys verified via dns-sec,
if that is a thing.
existing implementations
========================
so far, i've only seen _ssh._tcp records in SRV record examples, for
discovering local services (when it's not ssh itself that uses it,
programs like nautilus that display connectable machines nearby), and an
ssh wrapper at [1].
both the original problem and this possible solution are things that
would have to be taken upstream.
as a side effect, this would add SRV support to all other protocols that
rely on ssh (eg rsync and git). they might want to override the prefix
(maybe _ssh+git._tcp), but then again, that prefix would have to be
stored in known_hosts.
-----------------
andras, would this solve your problem?
best regards
chrysn
[1] https://gist.github.com/taylor/1372925
Attachment:
signature.asc
Description: Digital signature