gh-154271: Define _XOPEN_SOURCE=600 on Solaris to enable socket ancillary data and forkserver

[serhiy-storchaka]

On Solaris/illumos the X/Open socket ancillary-data API (CMSG_*, sendmsg(), recvmsg()) is declared only under _XOPEN_SOURCE>=600, but configure.ac left it undefined there (historically because it hides other features — gh-57208, bpo-1759169). So the socket module built without sendmsg/recvmsg, making multiprocessing.reduction.HAVE_SEND_HANDLE False and the forkserver start method unavailable.

Define _XOPEN_SOURCE=600 on Solaris; __EXTENSIONS__ (from AC_USE_SYSTEM_EXTENSIONS) re-enables the platform features it would otherwise hide — the same combination the OpenIndiana system Python builds with. Verified on OpenIndiana (illumos): sendmsg/recvmsg/CMSG_* are built, forkserver becomes available, full build clean, and test_socket / test_multiprocessing_forkserver / test_concurrent_futures pass.

Related: GH-154272 (test-robustness fix for platforms that genuinely lack forkserver).

🤖 Generated with Claude Code

• Issue: Solaris/illumos: _XOPEN_SOURCE left undefined disables socket.sendmsg and the forkserver start method #154271

[serhiy-storchaka]

@kulikjak, could you please look at this and verify if it is a correct change for real Solaris?

[kulikjak]

Thanks for the ping, Serhiy.

Interestingly, on Oracle Solaris, sendmsg and forkserver are available even without your change:

$> LD_LIBRARY_PATH=./build/sparcv9/ ./build/sparcv9/python
Python 3.16.0a0 (main, Jul 22 2026, 10:34:32) [GCC 15.2.0] on sunos5
Type "help", "copyright", "credits" or "license" for more information.
>>> import socket; hasattr(socket.socket, 'sendmsg')
True
>>> import multiprocessing; multiprocessing.get_all_start_methods()
['forkserver', 'fork', 'spawn']

The header files are likely slightly different now compared to Illumos. That said, I built and tested it with your change and everything seems to work as before.

One minor thing I noticed is that the configure now says:

checking whether it is safe to define __EXTENSIONS__... yes
checking whether _XOPEN_SOURCE should be defined... no

and then it's defined, which might be a little confusing.

[serhiy-storchaka]

Thanks for testing it!

Then the gap is illumos-specific -- I reworded the comment and the NEWS entry.

The confusing check is Autoconf's own, from AC_USE_SYSTEM_EXTENSIONS, and asks a different question -- whether to define _XOPEN_SOURCE to 500 for mbstate_t. It answers "no" on unpatched main too.