Horton

Re: Cleaning up historical portability baggage

Andres Freund <andres@anarazel.de>

View thread View on postgr.es
From: Andres Freund <andres@anarazel.de>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: Thomas Munro <thomas.munro@gmail.com>, Andrew Dunstan <andrew@dunslane.net>, Robert Haas <robertmhaas@gmail.com>, pgsql-hackers <pgsql-hackers@postgresql.org>
Date: 2022-08-11T17:14:12Z
Lists: pgsql-hackers 
Hi,

On 2022-08-11 10:52:51 -0400, Tom Lane wrote:
> Thomas Munro <thomas.munro@gmail.com> writes:
> > The most interesting things to say about these ones are:
> >  * The concept of a no-Unix-socket build is removed.  We should be
> > able to do that now, right?  Peter E seemed to say approximately that
> > in the commit message for 797129e5.  Or is there a thought that a new
> > operating system might show up that doesn't have 'em and we'd wish
> > we'd kept this stuff well marked out?
> 
> I'm kind of down on removing that.  I certainly think it's premature
> to do so today, when we haven't even yet shipped a release that
> assumes we can always define it on Windows

I think what might be good next step is to have tests default to using unix
sockets on windows, rather than requiring PG_TEST_USE_UNIX_SOCKETS. The
pg_regress.c and Utils.pm hunks disabling it by default on windows don't
really make sense anymore.

#if !defined(HAVE_UNIX_SOCKETS)
    use_unix_sockets = false;
#elif defined(WIN32)

    /*
     * We don't use Unix-domain sockets on Windows by default, even if the
     * build supports them.  (See comment at remove_temp() for a reason.)
     * Override at your own risk.
     */
    use_unix_sockets = getenv("PG_TEST_USE_UNIX_SOCKETS") ? true : false;
#else
    use_unix_sockets = true;
#endif

and

    # Specifies whether to use Unix sockets for test setups.  On
    # Windows we don't use them by default since it's not universally
    # supported, but it can be overridden if desired.
    $use_unix_sockets =
      (!$windows_os || defined $ENV{PG_TEST_USE_UNIX_SOCKETS});


Tests don't reliably run on windows without PG_TEST_USE_UNIX_SOCKETS, due to
the port conflict detection being incredibly racy. I see occasional failures
even without test concurrency, and with test concurrency it reliably fails.


I don't really know what to do about the warnings around remove_temp() and
trapsig(). I think we actually may be overreading the restrictions. To me the
documented restrictions read more like a high-level-ish explanation of what's
safe in a signal handler and what not. And it seems to not have caused a
problem on windows on several thousand CI cycles, including plenty failures.


Alternatively we could just default to putting the socketdir inside the data
directory on windows - I *think* windows doesn't have strict path length
limits for the socket location. If the socket dir isn't in some global temp
directory, we don't really need signal_remove_temp, given that we're ok with
leaving the much bigger data directory around.  The socket directory
determination doesn't really work on windows right now anyway, one manually
has to set a temp directory as TMPDIR isn't normally set on windows, and /tmp
doesn't exist.

    char       *template = psprintf("%s/pg_regress-XXXXXX",
                                    getenv("TMPDIR") ? getenv("TMPDIR") : "/tmp");

both TEMP and TMP would exist though...



> -- I won't be too surprised if we get pushback on that after 15.0 is out.

From what angle? I think our default behaviour doesn't change because
DEFAULT_PGSOCKET_DIR is "". And OS compatibility wise we apparently are good
as well?


> But in general, Unix sockets seem like kind of an obvious thing that might
> not be there on some future platform.

Maybe not with the precise API, but I can't imagine a new platform not having
something very similar. It serves a pretty obvious need, and I think the
security benefits have become more important over time...

Greetings,

Andres Freund

Commits

  1. Make sure IOV_MAX is defined.

  2. Doc: Abstract AF_UNIX sockets don't work on Windows.

  3. Clean up inconsistent use of fflush().

  4. Remove configure probe for sockaddr_in6 and require AF_INET6.

  5. Remove configure probes for sockaddr_storage members.

  6. Remove configure probe for netinet/tcp.h.

  7. Fix macro problem with gai_strerror on Windows.

  8. Remove configure probe for net/if.h.

  9. mstcpip.h is not missing on MinGW.

  10. Remove dead ifaddr.c fallback code.

  11. Remove configure probe for gethostbyname_r.

  12. Remove replacement code for getaddrinfo.

  13. Remove configure probe for struct sockaddr_storage.

  14. Remove HAVE_UNIX_SOCKETS.

  15. Remove configurability of PPC spinlock assembly code.

  16. Remove configure probe for shl_load library.

  17. Remove configure probe for sys/select.h.

  18. Remove configure probes for sys/un.h and struct sockaddr_un.

  19. Remove configure probes for sys/ipc.h, sys/sem.h, sys/shm.h.

  20. Remove configure probe for sys/resource.h and refactor.

  21. Remove configure probe for sys/uio.h.

  22. Fix obsolete comments in instr_time.h.

  23. aix: fix misreading of condition in 8f12a4e7add

  24. aix: Remove checks for very old OS versions

  25. windows: Remove HAVE_MINIDUMP_TYPE test

  26. solaris: Remove unnecessary gcc / gnu ld vs sun studio differences

  27. Trust a few system headers to stand on their own

  28. Rely on __func__ being supported

  29. Simplify replacement code for strtof.

  30. Simplify gettimeofday for Windows.

  31. Simplify gettimeofday() fallback logic.

  32. Remove fallbacks for strtoll, strtoull.

  33. Remove configure probe for fdatasync.

  34. Remove configure probe for clock_gettime.

  35. Simplify replacement code for preadv and pwritev.

  36. Remove dead pread and pwrite replacement code.

  37. Remove dead setenv, unsetenv replacement code.

  38. Remove configure probes for poll and poll.h.

  39. Remove configure probe for link.

  40. Remove configure probes for symlink/readlink, and dead code.

  41. Remove configure probe for setsid.

  42. Remove configure probe for shm_open.

  43. Remove configure probe and related tests for getrlimit.

  44. Remove configure probe for dlopen, and refactor.

  45. Remove dead getpwuid_r replacement code.

  46. Remove dead handling for pre-POSIX sigwait().

  47. Remove dead getrusage replacement code.

  48. Cygwin and Mingw floating-point fixes.

  49. Support platforms where strtoll/strtoull are spelled __strtoll/__strtoull.

  50. Arrange to supply declarations for strtoll/strtoull if needed.

  51. Lock down regression testing temporary clusters on Windows.

  52. Secure Unix-domain sockets of "make check" temporary clusters.

  53. Add configure checks for strtoll, strtoull (or strto[u]q). Disable

  54. Portability fix for old SunOS releases: fflush(NULL)