[Dovecot] Enhanced Kerberos support

[Greg Troxel]

Timo Sirainen <tss at iki.fi> writes:

>> SSH recently added this enhancement to address this common need:
>> 
>>       GSSAPIStrictAcceptorCheck
>>               Determines whether to be strict about the identity of the GSSAPI acceptor a client authenticates
>>               against. If “yes” then the client must authenticate against the host service on the current hostname.
>>               If “no” then the client may authenticate against any service key stored in the machine’s default
>>               store. This facility is provided to assist with operation on multi homed machines.  The default is
>>               “yes”.  Note that this option applies only to protocol version 2 GSSAPI connections, and setting it
>>               to “no” may only work with recent Kerberos GSSAPI libraries.
>
> Somehow this doesn't sound a very good idea.

This says "the host service on the current hostname", and I interpret
this as the principal "host/$hostname at MY.REALM", where $hostname is the
value returned by gethostname(3)/hostname(1).  There is no DNS involved
in this at all.

The alternative is to accept authentication to any principal either of
the form "host/$HOST@$REALM", as long as that key is stored in the
machine's keytab.

None of this involves DNS lookups.

>> I've heard that other daemons support multi-names by instead of using gethostname(), obtain the hostname of the
>> interface that the request came in on.
>
> I guess this would mean a PTR DNS lookup for the local IP? I've wanted
> to avoid DNS lookups in Dovecot so far, but proxying would also want to
> use them..

Yes, you could do this, allowing authentication to various names,
depending on the interface.  But I would think it's better to have an
option to either a) just allow the name that's configured as hostname,
or b) allow any host/ key that's in the keytab.

I don't see that it's useful from a security viewpoint to refuse
authentication that's done to host/foo when the request is received on
an interface that has an IP address that doesn't map to foo.  Actually,
I'd say that it isn't meaningful, for TCP at least, to talk about the
interface on which a request was received, and even for UDP packets can
arrive on arbitrary interfaces due to routing changes, and generally
these have no security consequences.

> I guess blocking DNS lookups for local IPs should be pretty safe and
> fast.

Why?  If the local DNS responder is hosed, it will be messy.  But this
is much less scary than lookups on random addresses.


What problem are we trying to solve?  The problem I can see is that if a
server is known by two names, clients may attempt to authenticate to
both of those names, and that should work (assuming both names have
service keys present in the keytab).

Are people trying to run some inside/outside split mailserver that's
both inside and outside a firewall?