[nsd-users] Wrong source IP for reply if 'ip-address' is not specified

[Anand Buddhdev]

On 13/12/2017 22:26, Paul Wouters wrote:

Hi Paul,

> We have a different opinion on what is "normal behaviour". I believe the
> normal behaviour is to reply using the IP address you received the
> packet from, eg using:
> 
> err = setsockopt(s, SOL_IP, IP_PKTINFO, &opt, sizeof(opt));
> 
> or
> 
> err = setsockopt(s, IPPROTO_IP, IP_RECVDSTADDR, &opt, sizeof(opt));

I don't know if these options are available in non-Linux socket
implementations, and is probably the reason that NSD doesn't use them.
But I'm sure Wouter can comment more definitively.

I know the questions will come, so let me try to anticipate them and
answer them. Someone might ask why this isn't necessary with BIND. This
is because BIND attempts to detect the capability of the OS it's running
on, and compensate for the cases where these advanced options are not
present. This may make it easier for an operator, but at the expense of
more code complexity. I really do prefer NSD's simpler approach.

Note also that in nsd.conf, this cause is very clearly noted. From nsd.conf:

ip-address: <ip4 or ip6>[@port]
    NSD will bind to the listed ip-address. Can be give multiple times
    to bind multiple ip-addresses. Optionally, a port number can be
    given.   If  none  are  given  NSD listens to the wildcard
    interface. Same as commandline option -a.  For servers with multiple
    IP addresses that can be used to send traffic to the internet, list
    them one by one, or the source address of replies could be wrong.

    This is because if the udp socket associates a source address of
    0.0.0.0 then the kernel picks an ip-address with which to send to
    the internet, and it picks the wrong one.  Typically needed for
    anycast  instances.   Use  ip-transparent  to  be  able  to  list
    addresses that turn on later (typical for certain load-balancing).

Regards,
Anand