[Fedora-directory-devel] Re: Please Review: Add LDAPI (LDAP over unix domain sockets)

[Howard Chu]

>  Date: Wed, 21 Feb 2007 12:16:04 -0800
>  From: Pete Rowley <prowley(a)redhat.com&gt; 
Sorry, I don't mean to beat too much on a dead horse, but ...

>  Howard Chu wrote:
> > > Yes. But again, the Heimdal KDC does an explicit SASL/EXTERNAL Bind to 
> > > request this privilege. There is no assumption of automagic 
> > > authorization. 
>  I guess we can add that. Rich and I have already talked about that as
> a TBD. 
>  While observing RFC4513 is a good thing, and this implementation does
> so 
>  when auto-bind is switched off, I believe these kinds of decisions are 
>  the domain of site administrative policy and not of standards documents. 
>  Further, a client in the anonymous bind state has no practical knowledge 
>  of the effects of that state on server responses in any case, nor can it 
>  be sure that binding as a non-anonymous user has any effect on those 
>  responses, nor indeed does auto-bind necessarily remove or add any 
>  privilege for the client - that is all administrative policy and 
>  undefined by any RFC. This is just one more administrative policy option. 
Certainly it's true that LDAP has nothing specific to say about 
authorization. But don't confuse authorization with authentication; they're 
two separate things. A client can easily determine that it's in an anonymous 
state, using the WhoAmI extended op.

If "ldapwhoami -x -H ldapi:///" doesn't return a zero-length ID, then your 
server is broken.

>  In addition, LDAP is defined as it is in no small part to the
> underlying 
>  assumption of TCP and designed around the practical methods of 
>  authentication given that assumption, strictly speaking LDAPI isn't LDAP 
>  (it's not even platform agnostic), and LDAPI has other methods at its 
>  disposal. 
And again, don't confuse the protocol with the transport. Fuzzy thinking like 
that is one of the reasons we have LDAP today, instead of just DAP over TCP. 
Sure the RFCs only define LDAP in terms of TCP, but obviously any reliable, 
ordered, stream transport will work just as well. And in a proper library 
implementation, such a transport can be substituted between any client and 
server without needing to alter any other code in the client or server.

>  While I understand your concern, the feature is an option, not a 
>  requirement. 
Yes, but it shouldn't be an option at all. You provide the option in the 
expectation that someone will use it. The presence of such an option 
therefore encourages client writers to depend on non-standard behaviors. You 
might say "this is a site-local policy decision" but it doesn't just affect
a 
single site. You implicitly create lock-in with features like this, and you 
have no idea what other servers such clients will eventually be talking to.
I would expect thinking like this from Microsoft or Sun, but not from an open 
source developer. And no matter how much we may think our respective server 
is the best in the world, and that no one would ever have any reason to talk 
to any other server, one need only look at SunOne to see that nothing lasts 
forever. Providing options that break standards like this *hurts* your users, 
it doesn't help them.

The standard mechanism for using an LDAP session with previously established 
credentials is to use a SASL/EXTERNAL Bind. Encouraging any other behavior is 
flat out wrong.

Yes, it's OK to provide options that completely change the behavior of the 
server. But it's only OK to do that *when the client explicitly asks for it*. 
I could define a control that causes all strings to be returned in Morse 
code, and that would be perfectly fine, because a client has to use the 
control before it gets such an outlandish behavior.

Some folks may be wondering why I'm spending so much time on this point, 
since it's your server and I'm not contributing to its code. Just remember 
that network protocols are about interoperability; nothing you do exists in a 
vacuum. Every mistake anyone makes affects everyone - just look at the 
dynamic group mess if you need an example...
-- 
   -- Howard Chu
   Chief Architect, Symas Corp.  http://www.symas.com
   Director, Highland Sun        http://highlandsun.com/hyc
   Chief Architect, OpenLDAP     http://www.openldap.org/project/

[Andrew Bartlett]

On Thu, 2007-02-22 at 13:55 -0800, Howard Chu wrote:
>  > Date: Wed, 21 Feb 2007 12:16:04 -0800
>   > From: Pete Rowley <prowley(a)redhat.com&gt;
>  
>  Sorry, I don't mean to beat too much on a dead horse, but ...
>  
>  > Howard Chu wrote:
>  >> > Yes. But again, the Heimdal KDC does an explicit SASL/EXTERNAL Bind to 
>  >> > request this privilege. There is no assumption of automagic 
>  >> > authorization.
>  
>  > I guess we can add that. Rich and I have already talked about that as a TBD.
>  
>  > While observing RFC4513 is a good thing, and this implementation does so 
>  > when auto-bind is switched off, I believe these kinds of decisions are 
>  > the domain of site administrative policy and not of standards documents. 
>  > Further, a client in the anonymous bind state has no practical knowledge 
>  > of the effects of that state on server responses in any case, nor can it 
>  > be sure that binding as a non-anonymous user has any effect on those 
>  > responses, nor indeed does auto-bind necessarily remove or add any 
>  > privilege for the client - that is all administrative policy and 
>  > undefined by any RFC. This is just one more administrative policy option.
>  
>  Certainly it's true that LDAP has nothing specific to say about 
>  authorization. But don't confuse authorization with authentication; they're 
>  two separate things. A client can easily determine that it's in an anonymous 
>  state, using the WhoAmI extended op.
>  
>  If "ldapwhoami -x -H ldapi:///" doesn't return a zero-length ID, then your
>
>  server is broken. 
Agreed.  Now I should implement that extended operation in our server
some day...

>  Yes, but it shouldn't be an option at all. You provide the option
> in the 
>  expectation that someone will use it. The presence of such an option 
>  therefore encourages client writers to depend on non-standard behaviors. You 
>  might say "this is a site-local policy decision" but it doesn't just affect
> a 
>  single site. You implicitly create lock-in with features like this, and you 
>  have no idea what other servers such clients will eventually be talking to.
>  I would expect thinking like this from Microsoft or Sun, but not from an open 
>  source developer. And no matter how much we may think our respective server 
>  is the best in the world, and that no one would ever have any reason to talk 
>  to any other server, one need only look at SunOne to see that nothing lasts 
>  forever. Providing options that break standards like this *hurts* your users, 
>  it doesn't help them.
>  
>  The standard mechanism for using an LDAP session with previously established 
>  credentials is to use a SASL/EXTERNAL Bind. Encouraging any other behavior is 
>  flat out wrong. 
>  Some folks may be wondering why I'm spending so much time on this
> point, 
>  since it's your server and I'm not contributing to its code. Just remember 
>  that network protocols are about interoperability; nothing you do exists in a 
>  vacuum. Every mistake anyone makes affects everyone - just look at the 
>  dynamic group mess if you need an example... 
I strongly agree with Howard on this one.  I want Samba to be able to
bind to either OpenLDAP or Fedora DS and have as few differences as
possible.  

And where OpenLDAP has done something first, or it's way of doing things
is more sane, I ask that Fedora DS follow that lead.  I need less, not
more 'if <vendor>' code...

(In this vane, I so wish nsUniqueID was in standard GUID format, rather
than %08x-%08x-%08x-%08x...)

If this needs to be made easier for client apps, then work on making
setting up ldapi:// connections with an EXTERNAL bind trivial to do from
the Perl Net::LDAP module and python.  That will help users, no matter
the LDAP server they choose to deploy. 

Andrew Bartlett

-- 
Andrew Bartlett                                http://samba.org/~abartlet/
Authentication Developer, Samba Team           http://samba.org
Samba Developer, Red Hat Inc.                  http://redhat.com

[Pete Rowley]

Andrew Bartlett wrote:
>  And where OpenLDAP has done something first, or it's way of doing
> things
>  is more sane, I ask that Fedora DS follow that lead.  I need less, not
>  more 'if <vendor>' code...
>
>     
Your if vendor code would be zero. Presumably Samba would be enabled 
with access in line with its operational requirements. Bearing in mind 
that Samba runs as root, it is likely to find that any machine it is 
installed on has anonymous access for root, just like it is allowed to 
actually run as root.
>  (In this vane, I so wish nsUniqueID was in standard GUID format,
> rather
>  than %08x-%08x-%08x-%08x...)
>     
I believe we have already discussed this and that I have already agreed 
that we would look at it.

-- 
Pete

[Andrew Bartlett]

On Thu, 2007-02-22 at 18:18 -0800, Pete Rowley wrote:
>  Andrew Bartlett wrote:
>  > And where OpenLDAP has done something first, or it's way of doing things
>  > is more sane, I ask that Fedora DS follow that lead.  I need less, not
>  > more 'if <vendor>' code...
>  >
>  >   
>  Your if vendor code would be zero. Presumably Samba would be enabled 
>  with access in line with its operational requirements. Bearing in mind 
>  that Samba runs as root, it is likely to find that any machine it is 
>  installed on has anonymous access for root, just like it is allowed to 
>  actually run as root. 
I'm not quite sure what you mean here, but what I don't want is a
situation where the admin runs Samba4 against a Fedora DS instance, and
forgets to explicitly set 'nsslapd-ldapiautobind: off'.  Samba would end
up proxying anonymous access as root!

It certainly seems an odd default.

Or worse still, there be a disagreement between applications as to if
this is a setting they want, or a setting they don't want.  Instead,
have applications that want EXTERNAL auth ask for it, just as they have
to ask for it for OpenLDAP.

Andrew Bartlett

-- 
Andrew Bartlett                                http://samba.org/~abartlet/
Authentication Developer, Samba Team           http://samba.org
Samba Developer, Red Hat Inc.                  http://redhat.com

[Pete Rowley]

Andrew Bartlett wrote:
>  On Thu, 2007-02-22 at 18:18 -0800, Pete Rowley wrote:
>    
> > Andrew Bartlett wrote:
> >     
> >> And where OpenLDAP has done something first, or it's way of doing things
> >> is more sane, I ask that Fedora DS follow that lead.  I need less, not
> >> more 'if <vendor>' code...
> >>
> >>   
> >>       
> > Your if vendor code would be zero. Presumably Samba would be enabled 
> > with access in line with its operational requirements. Bearing in mind 
> > that Samba runs as root, it is likely to find that any machine it is 
> > installed on has anonymous access for root, just like it is allowed to 
> > actually run as root.
> >     
>
>  I'm not quite sure what you mean here, 
I mean typically services are not
allowed to run as root, but apparently 
Samba must so Samba is configured to do so if the site needs Samba. In 
exactly the same way, as an example only, auto bind for root might be 
often mapped to some administrative user in the directory, but clearly 
that would not be desirable if one wanted Samba to run on the machine. 
Options would then be: don't configure root as anything other than 
anonymous, or, if that was not acceptable, configure samba to use LDAP, 
not LDAPI, or configure samba to have root OS privilege, but make use of 
the autobind feature that allows to more finely distinguish between OS 
users with the same uid and have Samba identified by its own unique 
entry with its own unique security context. None of those options 
involve an #ifdef vendor or even the slightest whiff of a branch in your 
code.

>  It certainly seems an odd default.
>
>     
Agreed, but that is moot at this point.

-- 
Pete

[Pete Rowley]

Howard Chu wrote:
> > Date: Wed, 21 Feb 2007 12:16:04 -0800
>  > From: Pete Rowley <prowley(a)redhat.com&gt;
>
>  Sorry, I don't mean to beat too much on a dead horse, but ...
>  
;)

OK. That is a pretty powerful argument, and my rereading of the RFCs 
reveals that there is about zero wiggle room on this precise point. 
However, for us this is an experimental feature on top of an 
experimental feature (ldapi itself), and in that light I propose to do 
the following:

1. #ifdef WITH_AUTOBIND the auto bind code
2. not #define WITH_AUTOBIND
3. offer no configure mechanism for turning on autobind

This allows experimenters to do their experimenting but requires an 
unusual step (in an autotools world) on the part of builders to add the 
feature, and it will not appear in the binary builds of FDS offered for 
download.

Regards

-- 
Pete