[Nut-upsdev] [Nut-upsuser] Question on simultaneous IPv4 and IPv6 "any address" listening

[Greg Troxel]

Jim Klimov <jimklimov+nut at gmail.com> writes:

> Thanks to everyone for a fruitful discussion, links and ideas.
>
> The result is nearing a merge at
> https://github.com/networkupstools/nut/pull/2013 and seems to not upset CI
> on any platform, including Windows (which behaves funny WRT binding to the
> same host:port as many times as you ask).
>
> Ultimately the chosen logic is that if there was a `LISTEN * <port>` in
> `upsd.conf`, the depending on CLI settings (-4/-6) or lack thereof we try
> PIv4, IPv6 or both, with the bigger logic for "both" being:
> * try to get IPv4 "ANY", to know we can do so
> * release IPv4, try to get IPv6 and then IPv4 again

As I commented on github, I think this is unnecessarily complicated and
thus not what we should do.

> If in the end neither socket works, declare a fatal error (could not
> fulfill the config requirement).
> If we could get IPv4 initially, and could not after getting IPv6, do not
> bother - assume a dual-stack system (and log it so).

You mean "a system that defaults to mapped addresses" :-)

> Also as part of this change, NUT would ask (although not insist) for the
> IPV6_V6ONLY socket option when preparing IPv6 connections, except when
> handling `LISTEN *`.

I don't know what you mean by insist.  To me the right logic is, after
opening a socket to be used for v6, very pseudocodish:

#ifdef V6ONLY_define_for_setsockopt
  call setsockopt to set v6only=1
  if that fails log an error
  /*
   * deal with bug reports later; it really seems buggy/odd to have the
   * setsockopt and then error on setting it to 1.
   */
  /*
   * It would be fine to ifdef out the v6only setting on openbsd if
   * that turns out ot be necessary.  But it wouldn't surprise me if on
   * openbsd setting it to 1 succeeds and setting it to 0 is an error,
   * in which case there is no need to special case it.
   */
#else
  /*
   * The operating system does not implement the setsockopt to control
   * whether mapped addresses are used.  Assume that it does not support
   * mapped addresses at all.
   */
#endif

and then for listen * to

  listen on v6, log erorrs
  listen on v4, log errors

and that's it.

I'm not sure what will then error; when we figure out what does we can
see what to do about it.

Overall I lean to doing things the simplest and most obviously correct
way possible and then to accomodate as necessary.

> Finally, I noticed that if some configuration hostname resolves to more
> than one address, the first one bound wins and others are ignored. This
> behavior was here before, the PR change just logs that this happens.

Not sure what you mean first bound wins/etc.  If you mean the hostname
is looked up and then the first address is picked arbirtarily and that
address is listened too, then ack that this is a pre-existing likely
bug, agreed that noting it is good, and agreed that there is no reason
to have to address that in this PR.