✅ Deploy Preview for python-cpython-preview canceled.

IMHO, we should keep all cached process-level resource availability/capability checks as static global variables. I'm specifically talking about the following group (and any others we might add in the future):

Making any of them per-interpreter is unnecessary since the value will always be uniform for the process. Moving them to _PyRuntimeState did not seem correct because the value is independent of Python's runtime and should not change across init/fini cycles.

tl;dr I'd rather we keep dup3_works as a static variable and not move it to the module state.

In the case of dup3_works the only concern I had was if there were any potential races, since the value is set lazily after runtime init. Looking at it again, there really isn't a race there. The only real race would be if two threads both see a -1 value (line 9463, if (dup3_works == -1)) and set the variable. However, that has no consequences if we assume word-size values like this are written atomically, which they should be. So let's not worry about it.

In the case of dup3_works the only concern I had was if there were any potential races, since the value is set lazily after runtime init. Looking at it again, there really isn't a race there

Okay, I understood :)

Thanks for working on this though!