Re: "could not adopt C locale" failure at startup on Windows

On Wed, Jun 10, 2015 at 10:09:38AM -0400, Tom Lane wrote:
> Noah Misch <noah@leadboat.com> writes:
> > I can reproduce this with "initdb --locale=C",
> > postgresql-9.4.3-1-windows-binaries.zip (32-bit), Windows 7 x64, and the
> > Windows ANSI code page set to CP936.  (Choose "Chinese (Simplified, PRC)" in
> > Control Panel -> Region and Language -> Administrative -> Language for
> > non-Unicode programs.)  It is neither necessary nor sufficient to change
> > Control Panel -> Region and Language -> Formats -> Format.  Binaries from
> > postgresql-9.4.3-1-windows-x64-binaries.zip do not exhibit the problem.  Note
> > that CP936 is a PG_ENCODING_IS_CLIENT_ONLY() encoding.
>
> Hm.  I could understand getting encoding difficulties in that environment,
> but it's hard to see why they'd manifest like this.  Can you trace through
> pg_perm_setlocale and figure out why it's reporting failure?

A faster test is to set LC_CTYPE=C in the environment and run "postgres
--version".  The root cause is a bug my commit 5f538ad introduced at the start
of the 9.4 cycle.  pg_perm_setlocale() now calls pg_bind_textdomain_codeset(),
which calls setlocale(LC_CTYPE, NULL).  POSIX permits that to clobber all
previous setlocale() return values, which it did here[1].  The ensuing
putenv("LC_CTYPE=<garbage bytes>") at the end of pg_perm_setlocale() fails
under Windows ANSI code page 936, because the garbage bytes often aren't a
valid CP936 string.  I would expect the same symptom on other multibyte
Windows locales.

While Windows was the bellwether, harm potential is greater on non-Windows
systems.  pg_perm_setlocale() sets the LC_CTYPE environment variable to help
PL/Perl avoid clobbering the process locale; see plperl_init_interp()
comments.  However, that function has bespoke code for Windows, on which
setting the environment variable doesn't help.  I don't know which other
platforms invalidate previous setlocale() return values on setlocale(LC_CTYPE,
NULL).  Therefore, I propose committing the attached diagnostic patch and
reverting it after about one buildfarm cycle.  It will make affected
configurations fail hard, and then I'll have a notion about the prevalence of
damage to expect in the field.

The actual fix is trivial, attached second.  This is for back-patch to 9.4.


[1] It does so in 32-bit "release" (non-debug), NLS builds done under Visual
Studio 2012 and Visual Studio 2013.  The official binaries used VS2013.  The
symptoms are slightly different under VS2012.  I did not test earlier
versions.  Debug builds and 64-bit builds were unaffected.