logo_kerberos.gif

Difference between revisions of "Samba4 Port: iptables Remapping"

From K5Wiki
Jump to: navigation, search
Line 1: Line 1:
 
<pre>
 
<pre>
 
Q: Background problem to solve: to port Samba4 (an OSS replacement
 
Q: Background problem to solve: to port Samba4 (an OSS replacement
for active directory) from Heimdal-krb to MIT-krb.
+
for Active Directory) from Heimdal-krb to MIT-krb.
   
The Samba4 people have built their AD-surrogate server
 
  +
Foreground problem:
(called Samba), configured to use the Heimdal KDC via a library
+
* Samba, the AD-surrogate server, uses the Heimdal KDC via a
interface, so that the KDC runs in the same process as all of
+
library interface, so that the KDC runs in the same process
the other AD services (SMB, netlogon, cifs, samr, etc). The
+
as all of the other AD services (SMB, netlogon, cifs, samr,
Samba4 people repeatedly insist that they won't accept an MIT
+
etc).
integration that doesn't work in exactly the same way.
+
* The Samba4 people repeatedly insist that they won't accept
The MIT-krb people respond that spec'ing such a libkdc interface,
+
an MIT integration that doesn't work in exactly the same way.
getting approval from the krbdev community, and building the
+
* The MIT-krb people respond that spec'ing such a libkdc
libkdc interface will take "a long time".
+
interface, getting approval from the krbdev community, and
  +
building the libkdc interface will take "a long time".
  +
 
Part of the technical disagreement is two conflicting realities:
 
Part of the technical disagreement is two conflicting realities:
* Unix deployments of MIT-krb take it for granted that you
+
* Unix deployments of MIT-krb take it for granted that you don't
don't run non-kdc services on the same box that hosts a kdc;
+
run non-kdc services on the same box that hosts a kdc; while
while
 
 
* Samba4 people know that Windows clients inflexibly expect that
 
* Samba4 people know that Windows clients inflexibly expect that
 
kdc, cifs, netlogon, ntlm, samr, etc all have to be served
 
kdc, cifs, netlogon, ntlm, samr, etc all have to be served
Line 21: Line 21:
 
So, I'm looking for a better solution.
 
So, I'm looking for a better solution.
 
Specifically, I believe the Samba4 people are more devoted to
 
Specifically, I believe the Samba4 people are more devoted to
uniformity than to their libkdc approach; they might accept an
+
uniformity than to their libkdc approach; they might accept
alternate solution, as long as both MIT-krb & Heimdal kdc will
+
an alternate solution, as long as both MIT-krb & Heimdal kdc
work identically with that one solution.
+
will work identically with that one solution. So, here's what
So, here's what I'm hoping can be made to work:
+
I'm hoping can be made to work:
 
* I want to keep the Samba4 kdc (Heimdal or MIT's) in a
 
* I want to keep the Samba4 kdc (Heimdal or MIT's) in a
 
separate process on the Samba-server host;
 
separate process on the Samba-server host;
Line 30: Line 30:
 
krb-service, I want the Samba server to pretend that it is
 
krb-service, I want the Samba server to pretend that it is
 
tunnelling the client request to the kdc, via an IPC connection.
 
tunnelling the client request to the kdc, via an IPC connection.
* A curlicue is that in unix krb deployments, the kdc ignores
 
  +
* The kdc needs to receive the krb-protocol packets with the
the client's ip-address for authentication purposes, but for
+
client's source-address intact. in unix krb deployments,
AD, the kdc usually enforces access-control on the client's
+
the kdc ignores the client's IP-address for authentication
IP-address. this means that the tunnel needs to present the
+
purposes, but for AD, the kdc usually enforces access-control
krb-protocol packets to the kdc, with the client's source-
+
on the client's IP-address.
address intact.
 
 
Is there some way to make this pseudo-tunnelled approach work,
 
Is there some way to make this pseudo-tunnelled approach work,
 
without reimplementing the TCP/IP stack?
 
without reimplementing the TCP/IP stack?
Line 67: Line 67:
 
Q: Doesn't matter; the krb protocol is GRAS (generally recognized
 
Q: Doesn't matter; the krb protocol is GRAS (generally recognized
 
as safe).
 
as safe).
A: xinetd wouldl add some overhead.
+
A: xinetd would add some overhead.
 
Q: I believe kdc performance isn't a critical issue in Samba
 
Q: I believe kdc performance isn't a critical issue in Samba
 
deployments; AD users typically deply 1 AD-server per subnet
 
deployments; AD users typically deply 1 AD-server per subnet
Line 76: Line 76:
 
of user-accounts). So in unix deployments, performance-hits
 
of user-accounts). So in unix deployments, performance-hits
 
would be bad.
 
would be bad.
A: Sounds like kernel-level redirection might be best
+
A: Sounds like kernel-level redirection (iptable-remapping)
Q: kernel-level redirect is the iptable-remapping option?
+
might be best. you need to use the routing capabilities of
A: Yes; you would want to use the routing capabilities of the
+
the linux kernel.
linux kernel.
 
 
A: So the samba-server-host would be a router for the kdc port.
 
A: So the samba-server-host would be a router for the kdc port.
 
Q: Is this feature universal across linux distros?
 
Q: Is this feature universal across linux distros?
A: Yes; linux makes for a very good router.
+
A: Yes; linux makes for a very good router; Cisco uses it for
A: Cisco uses it for low-end models
+
low-end models
 
Q: Do I have hope of seeing this feature on Solaris or on *bsd?
 
Q: Do I have hope of seeing this feature on Solaris or on *bsd?
A: I think most Unix boxes can act as routers
+
A: I think most Unix boxes can act as routers.
 
A: What you are wanting to do is to route just one port and
 
A: What you are wanting to do is to route just one port and
 
leave others alone.
 
leave others alone.

Revision as of 14:40, 31 August 2009

Q: Background problem to solve: to port Samba4 (an OSS replacement
   for Active Directory) from Heimdal-krb to MIT-krb. 

   Foreground problem:  
   * Samba, the AD-surrogate server, uses the Heimdal KDC via a 
     library interface, so that the KDC runs in the same process 
     as all of the other AD services (SMB, netlogon, cifs, samr, 
     etc).  
   * The Samba4 people repeatedly insist that they won't accept 
     an MIT integration that doesn't work in exactly the same way.
   * The MIT-krb people respond that spec'ing such a libkdc 
     interface, getting approval from the krbdev community, and 
     building the libkdc interface will take "a long time".
   
   Part of the technical disagreement is two conflicting realities:  
   * Unix deployments of MIT-krb take it for granted that you don't 
     run non-kdc services on the same box that hosts a kdc; while
   * Samba4 people know that Windows clients inflexibly expect that
     kdc, cifs, netlogon, ntlm, samr, etc all have to be served 
     from the same ip-address.

   So, I'm looking for a better solution.
   Specifically, I believe the Samba4 people are more devoted to 
   uniformity than to their libkdc approach;  they might accept 
   an alternate solution, as long as both MIT-krb & Heimdal kdc 
   will work identically with that one solution.  So, here's what 
   I'm hoping can be made to work:  
   * I want to keep the Samba4 kdc (Heimdal or MIT's) in a 
     separate process on the Samba-server host;
   * When the Samba server catches a client request for AD-style 
     krb-service, I want the Samba server to pretend that it is 
     tunnelling the client request to the kdc, via an IPC connection.
   * The kdc needs to receive the krb-protocol packets with the 
     client's source-address intact.  in unix krb deployments, 
     the kdc ignores the client's IP-address for authentication 
     purposes, but for AD, the kdc usually enforces access-control 
     on the client's IP-address. 
   Is there some way to make this pseudo-tunnelled approach work,
   without reimplementing the TCP/IP stack?
   I'm told that tun/tap kernel modules might help, but I dunno.
   One asset in my favor:  the kdc knows how to deliver krb service
   via UDP; though AD clients use only TCP-mediated krb-service. 
 
A: How about:
   * iptables redirection:  iptables can take a port and redirect
     it.  You can also tunnel between machines with ipsec.
   * xinetd can also do redirection and tunneling
   * sshd or stunnel can do it (stunnel might be better than sshd).
Q: So you're saying, the samba4 server catches (readrecv()) the
   client's kdc-request, and the samba4 server can then tunnel
   the client's request to the kdc, inside some other IPC-
   connection protocol, so that the client & the kdc don't know
   about the tunnel?  is this easy to set up?
A: I was thinking iptables could probably do the proxy work and
   send the packet to another machine.
   if that doesn't work, then xinetd might be possible
Q: The Samba4 people really-really want everything on one host;
   ease-of-administration is one of their selling-points.
A: The only issue with doing it in iptables is that you have to
   make sure not to delete tha NAT rules when locking down the
   system
Q: But having a host-to-host tunnel solution available for
   krb-admins who want it, would be good from MIT's POV, and
   might be acceptable to the samba4 people, as non-default
   behavior.
A: Would the traffic be something you wouldn't want 3rd parties
   to see?  in other words, clear text or encrypted?
Q: Doesn't matter;  the krb protocol is GRAS (generally recognized
   as safe).
A: xinetd would add some overhead.
Q: I believe kdc performance isn't a critical issue in Samba
   deployments;  AD users typically deply 1 AD-server per subnet
   or per building, so it's rare for an AD server to see heavy
   kdc-traffic.
Q: In unix-based krb deployments, you usually see 1 master kdc
   & a couple slaves handling a large organization (thousands
   of user-accounts).  So in unix deployments, performance-hits 
   would be bad.
A: Sounds like kernel-level redirection (iptable-remapping) 
   might be best.  you need to use the routing capabilities of 
   the linux kernel.
A: So the samba-server-host would be a router for the kdc port.
Q: Is this feature universal across linux distros?
A: Yes;  linux makes for a very good router;  Cisco uses it for 
   low-end models
Q: Do I have hope of seeing this feature on Solaris or on *bsd?
A: I think most Unix boxes can act as routers.
A: What you are wanting to do is to route just one port and 
   leave others alone.
Q: I think so;  I believe the Samba service catches different
   services on different ports, as usual.
A: If that's the case, it should be easy
Q: So you're suggesting that the Samba server can route kdc
   traffic to the kdc, on the same host or on a different host,
   in such a way that the client doesn't have to know the kdc's
   real IP-address / port?
A: Yes
Q: And in such a way that the kdc does see the client's real
   IP-address, just as if the kdc readrecv()'ed the client's
   original kdc-request-packet?
A: http://www.netfilter.org/documentation/HOWTO/netfilter-extensions-HOWTO-4.html#ss4.5
A: ^^^ might be useful
A: Now the question is did it make it into iptables mainline by now