directory-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Enrique Rodriguez" <enriqu...@gmail.com>
Subject Re: [SASL] SASL questions
Date Sun, 04 Mar 2007 02:13:39 GMT
On 3/3/07, Alex Karasulu <akarasulu@apache.org> wrote:
> On 3/1/07, Enrique Rodriguez <enriquer9@gmail.com> wrote:
> > ...
> > sasl-regexp
> >           uid=(.*),cn=example.com,cn=gssapi,cn=auth
> >           uid=$1,ou=users,dc=example,dc=com
>
> I'm having a bit of a problem understanding this.  It seems you're matching
> for some UID value in one DN (the first) and using that to map to another DN
> (2nd one in your example).
>
> I'm probably being a bit near sighted here and missing something.  Did we
> not want to use the regex to map a krb5 principal name to an LDAP DN?
>
> BTW is this example an OpenLDAP thing or a convention used to organize
> krb principal entries?

This is OpenLDAP config.  While we do, ultimately, want to map a
Kerberos principal to a DN, Kerberos (GSSAPI) is only one
authentication mechanism.  For example, users may also use DIGEST-MD5.
 The first DN is sort of a "virtual DN" where the root
"cn=gssapi,cn=auth" says the mechanism is GSSAPI.  It could also be
"cn=digest-md5,cn=auth", giving you a layer of indirection.  This
"virtual DN" format is:

uid=<username>,cn=<realm>,cn=<mech>,cn=auth

This allows you to have different mappings based on mechanism type.
For example, perhaps machine accounts must use GSSAPI while user
accounts can use GSSAPI or DIGEST-MD5 and machines and users are in
separate OU's?

> > The alternative would be to specify a baseDN, like we do for other
> > lookups.  We then search for the principal name and use the found DN.
> > Our configuration could be:
> >
> > gssapiBaseDn = ou=users,dc=example,dc=com
>
> Do you think not using a regex for KRB5 principal to LDAP entry DN mappings
> will limit the flexibility of the server?  Might also be nice to support
> both mechanisms
> but that might be a little too much to grok for just getting something out
> the door.
>
> Sounds to me like the use of the base dn configuration parameter is much
> easier to implement, understand and convey to our users.

I do think the regex is important for flexibility.  1..n "base DN"s
would be easier, but we certainly can do regex.  Also, if
per-mechanism mapping is desirable then we need to figure out how to
allow that.

One thing I think we should be adamant about from the beginning with
this LDAP+Kerberos integration is that we must have multiple realm
support.  You must be able to have, for example, users under
ou=users,dc=apache,dc=org AND ou=users,dc=example,dc=com where the
realms are, respectively, APACHE.ORG and EXAMPLE.COM.  Plus, as Quanah
pointed out, you'll have user, machine, and service principals in
separate subtrees, as well.  I found it interesting that your
ApacheCon tutorial talks about how the core Partition Nexus performs
"call routing based on namespace."  Multi-realm is a no-brainer to
implement; in fact, the old OSGi build allowed this.  The limitation
today is only that there's no config support for it; we allow one base
DN.

I don't want to complicate the 1st round of this, but ... this
multi-base lookup applies to multiple zones for DNS and multiple
subnets for DHCP.  The DNS RFCs call this a "catalog."  Again, maybe a
core service (or two) will work for pre-processing everything for
everybody.

As you noted, it's best to get something out the door, but I wanted to
get some awareness going.

> > 2)  If the best way from #1 is a lookup on Kerberos principal name at
> > a baseDN, where is the best place to invoke that lookup?  I think I
> > have to do the lookup in the LDAP protocol provider, since the
> > Authenticator.authenticate () method requires an LdapDN.
>
> I guess you're going to have to add the work flow for managing the SASL bind
> within the PP.  The standard authentication pathway which uses these
> Authenticators
> may be useless to you.

Yeah, the PP will handle workflow, but it will also coordinate with a
MINA filter that handles wrapping and unwrapping data, once the Bind
has completed.

Inbound -> SaslFilter (off) -> decoder -> BindHandler <-> backend 1, 2
Outbound <- SaslFilter (off) <- encoder <--------'

All subsequent LDAP operations:
Inbound -> SaslFilter (on) -> decoder -> SearchHandler <-> backend
Outbound <- SaslFilter (on) <- encoder <-------'

#1 is a user lookup to authenticate.  Should be an Authenticator, once
we figure out how to derive a DN for the user based on principal w/
realm.
#2 is an LdapContext establishment for that user and stored in the session.

> It's been a while since I've been in this code so bear with me.  I guess I
> can give
> a better answer if I started fishing around in the code to get reacclimated.
>  If you
> need me to do that just let me know and I can make some time.

Cool, thanks.  I'll get some code committed and we can start to review
the more hairy integration issues.  BTW, where should I start putting
work?  I'm thinking the server-sasl module but I think longer term it
will just roll into the PP.  From the server-sasl or protocol-ldap
modules we can review the touchpoints with the core.  server-sasl
seemed to me to be an effort to segregate SASL since it was a JDK 1.5
thing.  Also, FYI, I don't think I'll be able to make this a drop-in
MINA subproject, like SSL, since there is a handler that must talk to
a user store.  Anyway, we can review it later.

> > If I set env
> > property Context.SECURITY_PRINCIPAL to the Kerberos principal name
> > from GSSAPI, is there any way to do a lookup and convert that prior to
> > calling Authenticator.authenticate(), but in the backend, not the
> > protocol provider?
>
> We could setup code withing the authentication service (interceptor) to
> detect
> sasl binds and automatically map the Krb5 principal name to an LDAP name.

Cool, makes sense.  As I mentioned above, a common way to map
zones/realms/subnets and handled in a common place would be very
powerful.

> > Would we gain anything?  FWIW, the Kerberos
> > provider currently does lookups in the protocol provider.
>
> You could do that or alter the LDAP infrastructure between the pp and the
> partition (the interceptors) to handle sasl binds.
>
> > BTW, the regex is sounding easier and probably more performant.
>
> Oh? Why do you think?

Hmm, a misstatement on my part.  I was thinking the regex would be
more performant as you would build a DN and look directly for it,
while a lookup at a baseDN would mean a broader search.  And I have
lookup code from Kerberos but nothing for regex's so regex support
would be a new effort, and thus "harder."  Mostly my own lack of
experience with regex's.

> > 2)  Any opinion on the 'authenticatorType' to use?  Doco seems to
> > indicate that the choices are "none," "simple," and "strong."
> > However, it might be better (ie more modular) to have an authenticator
> > for each SASL type, eg "sasl-gssapi" and "sasl-digest-md5."  Even with
> > 2 SASL mechanisms supported we could be looking at one large
> > Authenticator.  Would that be a pain for embedders, in which case we
> > could use "strong" and have a separate env property if we decide to
> > have multiple authenticators?
>
> Hmmm I think adding strong type specifiers for specific SASL mechanisms
> might require a little refactoring in how Authenticators work.  I just need
> to take a better look at the code again.

OK, I'll wait.

> > 3)  I'm planning on adding GSSAPI.  What other SASL types are actually
> used?
>
> I think that's the most useful and most expected of mechanisms.  However md5
> is also popular.
>
> <dreaming>
> It would be cool would be to at some point implement the SASL mechanism for
> utilizing OTPs.
> </dreaming>

I looked at that while reviewing SASL doco.  IIUC, it isn't a generic
mechanism for any type of OTP, but rather a specific OTP algorithm.
We couldn't reuse it for HOTP, if that's what you're thinking.  HOTP
or HOTP+PIN would be great to add with Binds, though.

> I think I need to dive back into this code to be more useful to you.
> Furthermore it's not
> in the best shape it could be in.

It's been a great exercise for getting familiar with the core.

Enrique

Mime
View raw message