Problem with port 0

[Bruce M. Simpson]

Ricardo Nabinger Sanchez wrote:
> On Fri, 12 Jan 2007 12:03:17 -0500
> "Hug Me" <hugme at hugme.org> wrote:
>
>   
>> We believe FreeBSD is not allowing a UDP source port of 0 and the kernel is
>> dropping the packet before it ever reaches the tftp server but are unable to
>> verify this hypothesis. I was hoping someone here could help shed some light
>> on the problem.
>>     
>
> But port 0 has special meaning to the kernel (ie, "give me some random
> port").  Also, it is a reserved one.  Please check IANA:
>
> 	http://www.iana.org/assignments/port-numbers
>
> I'm afraid you'll have to select another port number.
>   
Nope. A source port of 0 is perfectly legal for UDP.

I did an experiment with rpcbind whereby I performed a UDP based rpcinfo 
query from one FreeBSD machine to another, captured the traffic, and 
used tcpreplay to inject it from source port 0.

At first I thought the INPCB hash lookup was doing the wrong thing, then 
I ktrace'd rpcbind and it was apparent that it was in fact being 
delivered to rpcbind from udp_input().

rpcbind tries to reply to destination port 0. This was verified with 
kdump and rpcbind -d. This quite rightly fails, and, indeed, we reject 
this from the socket code.

So, FreeBSD appears to handle a UDP source port of 0 ok based on these 
tests.

The most likely explanation for the failure in this case, without 
looking further, is that inetd or the tftpd implementations can't handle 
source port 0.

BMS