Various missing constants (and some functions) #1325

hi again, I have some more super fun things to report

as a short preamble, fixing the missing constants seems to be mostly just a continuation of the work in #193 also i apologize if this should really be multiple issues, I did it all in one bc I thought it would've felt needlessly spammy since the majority of the work here is more or less the same thing even if concerns different upstream files. also same as before I'm just reporting this here bc the most immediate issues is (aside from the generally missing RDS_* macros) for y'all. I also wrote some patches just to examine ig if it rly is just continuing the existing work and it unsurprisingly does seem to be, the same nuances apply for avoiding header clashes but aside from that seems to be all the same thing more or less.

The tldr is in terms of missing constants is as follows

• #ifdef const: exported under #ifdef NAME, so the build's libc headers must define the name.
• #ifdef group (X): exported without its own guard, inside a block guarded by X, so it is missing because X is.
• glibc include: the header CPython takes it from only exists in newer glibc, so the include itself is skipped.
• #ifdef const (argument): an argument that raises NotImplementedError unless its C macro is defined.
• configure link test: an argument that needs a glibc wrapper function, detected by linking at configure time.

Process and timers (os)

Vectored I/O flags (os.preadv / os.pwritev)

Later flags (#139090 RWF_DONTCACHE, #153856 RWF_NOSIGNAL) will follow the same pattern: posixmodule does not include <linux/fs.h>, so they only appear once glibc's headers have them.

Clocks, polling and memory (time, select, mmap)

File locks and descriptors (fcntl)

File sealing (fcntl)

Write-life hints (fcntl)

TCP options (socket)

All are guarded one by one in socketmodule.c:9155-9193 and come from glibc's <netinet/tcp.h>; <linux/tcp.h> cannot be included next to it.

Protocols, address families and socket levels (socket)

Missing functions

preadv()/pwritev() refuse any non-zero flags #ifndef HAVE_PREADV2 / HAVE_PWRITEV2 (posixmodule.c:12418-12423, 13124-13129) through argument_unavailable_error(). HAVE_PREADV2 is configure's link test for the glibc 2.26 wrapper, which a 2.19 sysroot fails. Here the existing syscall fallback strat should work. i.e. just

#if defined(HAVE_PREADV2) || (defined(__linux__) && defined(__NR_preadv2))
#  define _Py_HAVE_PREADV2
#endif
#if defined(HAVE_PWRITEV2) || (defined(__linux__) && defined(__NR_pwritev2))
#  define _Py_HAVE_PWRITEV2
#endif
#ifdef __linux__
// Mirroring `LO_HI_LONG(val)` from (sysdeps/unix/sysv/linux/sysdep.h:93-95)
#  define _Py_LO_HI_LONG(off) (long)(off), (long)((uint64_t)(off) >> 32)
#endif

and then at the HAVE_{PREADV2/PWRITEV2} sites the approaches reduces to just

#ifndef _Py_HAVE_*V2
   // arg unavail error for nonzero flags 
#else
#ifdef _Py_HAVE_*V2
    // syscall avail on non-zero flags, yippeee!
#endif

The usual point applies that these changes mean for older host libc's or kernels OSError's would be raised with these fixes applied.

Related upstream issues/notes

• test_pwritev_flags has never run: it is gated on hasattr(posix, 'os.RWF_SYNC'), always False since bpo-31368 (2018), bc it should be 'RWF_SYNC' and its expected data is wrong (b'xxtest1tt2' instead of b'xxtest1tt2t3' validated against just calling it directly).
• I think it would make sense to probably add tests to CPython to assert that these macros are actually present i.e. the fallbacks correctly trigger when building against lacking libc, unsure if there is any prior instance of this kind of testing but I assume it might be nice even if it is basically just testing like basic preprocessor semantics. One example that motivates it would be the addition of PIDFD_NONBLOCK, was added in 2022 but nothing CPython included actually defined it so it was always just gone untill 2 years later the actual header file was added to make it work, I would assume had this been given like a presence check it would've raised the error.
• I was curious if the PIDFD_NONBLOCK defect ,i.e. the missing header for macro defs, appeared anywhere else and I found that the full RDS_* group here which comes from rds.h exhibits the same issue i.e. its missing rds.h, there's one related open issue based on what I can tell virtually nobody uses these constants hence why this has gone essentially unreported since they were added in 2011; turns out CPython does a namespace pollution lint test this might complicate adding the header, (could use __has_include (C23, alrdy used in some places) maybe?), but again low prio. There is also MADV_SOFT_OFFLINE defined in <asm-generic/mman-common.h> at mmapmodule.c:2379-2381 but header is missing.