posix: sysconf: match _SC* defines with newlib/picolib

[nslowell]

It's possible for newlib/picolib libc libraries to internally call sysconf() which would execute zephyr's implementation. However, if the _SC* defines do not have matching values, then the incorrect switch case executes.

This issue arises when using newlib/picolib libc that includes sysconf implementation for ARM. With current defaults, the zephyr sysconf() overrides the original libc sysconf() so we must ensure proper operation.

[keith-packard]

Is there some reason not to just use the picolibc/newlib values when those C libraries are in use?

[cfriedt]

@nslowell - can we switch to #defines instead of enum?

The major benefit of that is getting warnings / errors if the symbol is redefined.

But also, the enum was mainly there because originally there was no specific ordering required.

[keith-packard]

@nslowell - can we switch to #defines instead of enum?

And use the #defines that come from libc when available?

[cfriedt]

And use the #defines that come from libc when available?

That could be achieved with something like #include_next <sys/sysconf.h> assuming that file exists in newlib / picolibc.

[keith-packard]

That could be achieved with something like #include_next <sys/sysconf.h> assuming that file exists in newlib / picolibc.

Yup, although the standard file is <unistd.h>, so Zephyr should use that if possible. If that's not possible, then using <sys/unistd.h> could work.

[nslowell]

Appreciate the feedback @cfriedt and @keith-packard.

Switching from enum to #defines that match newlib/picolib seemed doable for POSIX_SYSCONF_IMPL_FULL. I did have to adjust how z_sysconf() is used to prevent it from breaking due to preprocessor expansion, like I mentioned in #87833 (comment)
Example:

#define z_sysconf(x) (long)CONCAT(__z_posix_sysconf_, x) /* <--- add trailing underscore here */

switch(x) {
	case _SC_ADVISORY_INFO:
		return z_sysconf(SC_ADVISORY_INFO); /* <--- drop the leading underscore here */

However, for POSIX_SYSCONF_IMPL_MACRO, there is no way around the sysconf breakage, so the #defines have to be wrapped within #ifdef POSIX_SYSCONF_IMPL_FULL.

Secondly, I really thought getting this code to reference newlib/picolib's unistd.h was going to be a bit of a headache, but I guess I got lucky. It's possible there's some scenario this breaks, but it seemed to work with all libc options (minimal, newlib, and picolib).

I'm hoping #88195 will merge soon, so this can run the full test suite and find any potential issues with this change.

Thanks again for the feedback.

[nslowell]

@keith-packard, @cfriedt
As I had feared, trying to #include_next <unistd.h> to pull in defines from newlib/picolib creates a mountain of declaration conflicts I simply was not able to climb over. This PR now has a few commits, the first few being an initial attempt to resolve issues but they just continued to pile up.

Secondly, there's no way to maintain the macro version of sysconf() and have successful tests without isolating it and keeping the original enum declaration for it. The new #defines that match newlib/picolib values will be only for the full implementation.

Hope this is sufficient. I've got all tests passing (except for timeouts which are unrelated and pass when run locally)

[nslowell]

@cfriedt @keith-packard I think this is good to go now.

[cfriedt]

LGTM - thanks for all of your help @nslowell 🙏

[cfriedt]

In case there is another revision, I would just bracket (...) this entire expression out of paranoia, but not blocking.