_socket.py: fix for systems where AI_NUMERICSERV is not defined

[barracuda156]

Rationale: on some platforms AI_NUMERICSERV may not be defined. In particular, old versions of macOS, possibly some other. While affected users are probably few, it is a trivial fix, which makes trio usable (for example, yewtube app works then).

Credits to @CoolCat467
Closes: #3133

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 99.62%. Comparing base (a32de81) to head (5260922).
Report is 1 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #3135   +/-   ##
=======================================
  Coverage   99.62%   99.62%           
=======================================
  Files         122      122           
  Lines       18357    18358    +1     
  Branches     1221     1221           
=======================================
+ Hits        18289    18290    +1     
  Misses         47       47           
  Partials       21       21           

Files with missing lines	Coverage Δ
src/trio/_socket.py	100.00% <100.00%> (ø)

[CoolCat467]

Looks good, but needs newsfragment (that is why continuous integration is failing). newsfragments folder's README file has more information.

[barracuda156]

@CoolCat467 That one CI which fails is unrelated, right?

[A5rocks]

@CoolCat467 That one CI which fails is unrelated, right?

Docs are failing because your newsfragment is using markdown not rST. Try doing socket.AI_NUMERICSERV instead of AI_NUMERICSERV (for the automatic linking) or doubling the backticks.

It's a bit confusing because the logs show two warnings, but the one you should care about is:

/home/docs/checkouts/readthedocs.org/user_builds/trio/checkouts/3135/docs/source/history.rst:42: WARNING: py:obj reference target not found: AI_NUMERICSERV [ref.obj]

I don't think the other one makes it fail, though I may be wrong.

[jakkdl]

perhaps good to add a comment explaining why this is needed / when the workaround can be removed

[CoolCat467]

From the original issue, what we have learned so far is that it doesn't exist in MacOS 10.6 on PowerPC. Not sure immediately what version that changed in.

[barracuda156]

@CoolCat467 It does not exist on any macOS on any arch before 10.6 release. In practice, this will mostly be 10.5 and 10.4, on ppc and i386. (I.e. the fix has nothing to do with the architecture, it is only dependent on SDK.)

[jakkdl]

I think you need to be a bit more specific here, otherwise future people that stumble on this comment will have to do a bunch of research to figure out why/when it might be missing. And as I said in the issue, perhaps even a note on why it's needed despite being deprecated since a decade+

[A5rocks]

Maybe just linking the issue will work? That has info.