Nut upsuser - Apr 2024 - RFE to extend "LISTEN" directive to support host-colon-port (as single token)

Jim Klimov via Nut-upsdev <nut-upsdev at alioth-lists.debian.net> writes:
>   A recent discussion in the issue tracker brought up the idea to allow the
> `LISTEN` keyword to also accept a single "host:port" token (e.g.
if there
> is only one argument, with at least one colon, and the last colon is
> followed only by numbers, split it into host and port) :
> https://github.com/networkupstools/nut/issues/2424
Is the point that people want to use different ports, and the current
situation lets you choose an address but not a port?

Assuming so, why would there be a restriction to a single host if there
is port, while one could have multiple listen addresses if not?  I would
think the : scheme should apply to each argument, with lack of : being
an implict :[normalport].

For me, the reason to use explicit listen is because you don't like *,
and if you are using IP addresses you might well want to listen to v4
and v6.

This raises the issue of whether "host" expands to all IP addresses
associated with a domain name.
>   There are also certain cons, primarily about parsing such stuff reliably
> and consistently in different code bases (now also with augeas and nutconf
> to worry about). The actual "production" parsing in NUT data
server code
> should be trivial.
I find this whole "our config needs to be generally machine readable but
we aren't just changing it to a machine-readable format" to be odd.  
If
we want to play in some world where that happens we should just flip to
yaml or something :-)
>   On a somewhat related note, should the port part be constrained to
> numbers, or should it also pass through the naming database (resolve via
> typically /etc/services on POSIX systems) if it is a non-numeric string,
> similar to how we resolve host names into IP address numbers?
sure, non-numeric port could go through getservbyname(3) and then fail
if not the expected protocol.  Don't talk about /etc/services but the
posix interface, except it's from 4.2BSD and I'm not sure it's been
specified by POSIX :-)

On the other hand, if you are using an alternate port, it's because you
are not doing the normal thing, and the idea that you specify 'ntp'
because you want nut on 123 doesn't really make sense.   And if you know
some name it's not hard to look it up by name.   So all in all, I vote
for no, it's a number.
>   What would the community say, is any of this worth spending time on?
> Would anyone volunteer to roll up the sleeves for that? :)
No and no, IMHO.   What is the problem being solved?  Are there actual
people who want to use a different port?  What are their reasons?
I can believe there might be a scenario, but it seems speculative.

Thanks for sharing your take on this. (Sorry about likely mixing historic
standards, was not in position to cross-check while posting)

Just to clarify: using a different port *is* possible since forever, with
`LISTEN host port` (as two arguments to the directive); the question was if
having a way to spell it as one argument as `LISTEN host:port` would solve
some shortcomings/ease adoption more than introduce some new problems :)

With recent releases addressing this area, a host name resolving to several
IP addresses should be recognized, but at the moment this would only emit a
warning about the situation and the first seen IP address number would get
attempted for bind(). If this proves to be a problem, should not be too
hard to address (need to inject entries into an internal list tracking the
sockets which is originally sized by amount of LISTEN lines; there is
already precedent for injection of IPv4 and IPv6 where a single `LISTEN *`
directive and avoided dual-stack mode are in place).

As for practical use of non-default ports, NIT (NUT Integration Testing)
scripts come to mind and do that extensively (especially where the same CI
host/agent can be running different scenarios in parallel, so any one
hard-coded port is prone to conflict).

Other practical reasons might include security by obscurity (like running
web consoles or ssh on strange ports), running different NUT data servers
(e.g. real drivers on one, and "dummy-ups" or "clone*"
relays on the
other), or attempts to avoid conflicts with uncooperative software. Can't
think of much more quickly :)

Hope this helps,
Jim Klimov


On Mon, Apr 29, 2024 at 2:31?PM Greg Troxel via Nut-upsuser <
nut-upsuser at alioth-lists.debian.net> wrote:
> Jim Klimov via Nut-upsdev <nut-upsdev at alioth-lists.debian.net>
writes:
>
> >   A recent discussion in the issue tracker brought up the idea to
allow
> the
> > `LISTEN` keyword to also accept a single "host:port" token
(e.g. if there
> > is only one argument, with at least one colon, and the last colon is
> > followed only by numbers, split it into host and port) :
> > https://github.com/networkupstools/nut/issues/2424
>
> Is the point that people want to use different ports, and the current
> situation lets you choose an address but not a port?
>
> Assuming so, why would there be a restriction to a single host if there
> is port, while one could have multiple listen addresses if not?  I would
> think the : scheme should apply to each argument, with lack of : being
> an implict :[normalport].
>
> For me, the reason to use explicit listen is because you don't like *,
> and if you are using IP addresses you might well want to listen to v4
> and v6.
>
> This raises the issue of whether "host" expands to all IP
addresses
> associated with a domain name.
>
> >   There are also certain cons, primarily about parsing such stuff
> reliably
> > and consistently in different code bases (now also with augeas and
> nutconf
> > to worry about). The actual "production" parsing in NUT data
server code
> > should be trivial.
>
> I find this whole "our config needs to be generally machine readable
but
> we aren't just changing it to a machine-readable format" to be
odd.   If
> we want to play in some world where that happens we should just flip to
> yaml or something :-)
>
> >   On a somewhat related note, should the port part be constrained to
> > numbers, or should it also pass through the naming database (resolve
via
> > typically /etc/services on POSIX systems) if it is a non-numeric
string,
> > similar to how we resolve host names into IP address numbers?
>
> sure, non-numeric port could go through getservbyname(3) and then fail
> if not the expected protocol.  Don't talk about /etc/services but the
> posix interface, except it's from 4.2BSD and I'm not sure it's
been
> specified by POSIX :-)
>
> On the other hand, if you are using an alternate port, it's because you
> are not doing the normal thing, and the idea that you specify 'ntp'
> because you want nut on 123 doesn't really make sense.   And if you
know
> some name it's not hard to look it up by name.   So all in all, I vote
> for no, it's a number.
>
> >   What would the community say, is any of this worth spending time on?
> > Would anyone volunteer to roll up the sleeves for that? :)
>
> No and no, IMHO.   What is the problem being solved?  Are there actual
> people who want to use a different port?  What are their reasons?
> I can believe there might be a scenario, but it seems speculative.
>
> _______________________________________________
> Nut-upsuser mailing list
> Nut-upsuser at alioth-lists.debian.net
> https://alioth-lists.debian.net/cgi-bin/mailman/listinfo/nut-upsuser
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://alioth-lists.debian.net/pipermail/nut-upsdev/attachments/20240429/4b860c32/attachment.htm>