samba - Jan 2004 - A possible error in the way nmbd fails to respond to certain request packets

I have just speant several hours trying to nut out an error with a new 
installation of SAMBA 3, and the results may indicate something has changed 
with the way the Name Server (nmbd) responds to certain request. It could also 
be network stack related, but I am about at my limit of skills to diagnose it 
any further.

I have just installed Fedora Core 1 to a test server in order to play around 
with SAMBA 3. The exact version is SAMBA 3.0.0-15.

My network consists of three machines. All the machines are in a workgroup 
called DRWHO.
192.168.1.50  : TESTSERVER - The new Fedora machine
192.168.1.100 : K9 - A Windows 98 client
192.168.1.1   : TARDIS - An older SAMBA Server (Linux Kernel 2.4.9 / SAMBA 
2.0.10)

I used the SWAT wizard to configure the new server and set up a single share 
point called "pub", which just allowed guest access to /tmp.

I then thought I would check out the new "nbmlookup -S <name>"
tool to see what
it would say about the different machines on the network. The results were as 
expected, so I won't repeat them here. I then went to the Win98 machine to 
compare these results with a "nbtstat -a <name>" command, and
this is where the
fun started.

Basically the Win98 machine kept responding saying that TESTSERVER was an 
unknown host. It correctly reported the results for itself and the older SAMBA 
server, TARDIS.

My first thought was that Fedora had set up some firewalling rules, even though 
I had turned it off during the configuration. But "iptables -L" showed
that
everything was set to ACCEPT.

So I fired up Ethereal on the new Fedora box, and watched what went across the 
network when I queried TESTSERVER from both its local console and from the 
Win98 box. From the local console the results were:-
Src=192.168.1.50 Dst=192.168.1.255 (broadcast) Name Query NB
TESTSERVER<00>
followed by a response, then
Src=192.168.1.50 Dst=192.168.1.50 Name Query NBSTAT TESTSERVER<00>
followed by a response.

Everything as I expected to see it. However, from the Win98 box I just got 
three request packets, spaced about a second apart, but no response.

A quick check from the console of TESTSERVER showed that nmbd had bound itself 
to the network correctly:-
[root@TestServer root]# netstat -na
udp        0      0 192.168.1.50:137        0.0.0.0:*
udp        0      0 0.0.0.0:137             0.0.0.0:*
udp        0      0 192.168.1.50:138        0.0.0.0:*
udp        0      0 0.0.0.0:138             0.0.0.0:*

So, if Ethereal was showing the query packets turning up on the network 
interface, and nmbd was listening on that interface, why wasn't it
responding?

So after much more fruitless time, running the nmbd daemon in interactive mode, 
with various levels of debug, searching google, and other online resources, I 
eventually (finally!) noticed in the Ethereal dumps that the exact request line 
from the Win98 box was:-
Src=192.168.1.100 Dst=192.255.255.255 (broadcast) Name Query NB
TESTSERVER<00>

It seems that there was (and always has been) an error in my DHCP configuration 
file that meant that the Win98 machine was getting a 255.0.0.0 netmask, even 
though the rest of the network was working off a 255.255.255.0 netmask.

While this was an error, I can't see why it shouldn't have worked. The 
destination address (IP=192.255.255.255 / Ethernet=ff:ff:ff:ff:ff:ff) should 
still be acceptable to a service listening on 0.0.0.0, and there was certainly 
enough routing information to return a response back to the Win98 box on 
192.168.1.100.

Also, the Win98 box has been successfully working with the older SAMBA server 
for several years, so it seems that *something* has changed. It could be that 
the new Fedora IP stack was dropping the request before it got to SAMBA, or it 
might be that SAMBA was ignoring the request - I couldn't tell which. Also, 
there might be a very good reason why such changes were made, in which case 
this message can just server as a warning to others.

Thanks very much for your time, and thanks for the great software.
Aaron Power.