Re: mod_perl2 and Perl 5.14 with uselargefiles on 32-bit architectures (was: Early core dump)

On Wed, Sep 28, 2011 at 02:22:49PM -0700, Marco Walther wrote:
> OK, I think I found one problem.  The following two defines don't
> make it from the Perl make to the CCFLAGS for the mod_perl:-(
> `-D_LARGEFILE_SOURCE -D_FILE_OFFSET_BITS=64' (They are automatically
> added by the Configure for perl and listed in the perl -V output
> below).
> 
> That causes the my_perl structure to be of different sizes/offsets
> between perl and mod_perl. That works by accident with Perl 5.10.1
> and finally breaks with 5.14.[12]

We're running into this on Debian 32-bit architectures too
(http://bugs.debian.org/636651 [cc'd]), and the issue is one of the
blockers for our transition to Perl 5.14.

> Unfortunately even trying to run
> /opt/kenai/bin/perl Makefile.PL DEFINE='-D_LARGEFILE_SOURCE
> -D_FILE_OFFSET_BITS=64'
> is not enough:-( The defines still do not make it to the
> src/modules/perl/Makefile:-( But after changing that Makefile by
> hand and rebuilding, things seem to be working fine.

These cpp flags are stripped by lib/Apache2/Build.pm, see
has_large_files_conflict() and strip_lfs().

The mod_perl2 2.0.5 test suite works for me with Perl 5.14 if I hardwire
has_large_files_conflict() to return 0 and apply r1125476 from 2.0.6-dev:
 http://svn.apache.org/viewvc/perl/modperl/trunk/src/modules/perl/modperl_svptr_table.c?r1=932879&r2=1125476

The elaborate comments about large file issues in lib/Apache2/Build.pm
around strip_lfs() seem to be partly outdated; selectively quoting:

# on Unix systems where by default off_t is a "long", a 32-bit integer,
# there are two different ways to get "large file" support, i.e. the
# ability to manipulate files bigger than 2Gb:
#
# 1) you compile using -D_LARGEFILE_SOURCE -D_FILE_OFFSET_BITS=64.
[...]
# 2) you compile using -D_LARGEFILE64_SOURCE
[...]
# The problem that mod_perl has to work around is when you take a
# package built with approach (1), i.e. Perl, and any package which was
# *not* built with (1), i.e. APR, and want to interface between
# them. [1]
[...]
# Perl built with -Duselargefiles uses approach (1).
#
# APR HEAD uses (2) by default.
[...]
# [1]: In some cases, it may be OK to interface between packages which
# use (1) and packages which use (2).  APR HEAD is currently not such a
# case, since the size of apr_ino_t is still changing when
# _FILE_OFFSET_BITS is defined.

The last paragraph dates back to 2004, and the apr changelogs read:

> Changes for APR 1.2.12
>   *) Define apr_ino_t in such a way that it doesn't change definition
>   based on the library consumer's -D'efines to the filesystem.
>   [Lucian Adrian Grijincu <lucian.grijincu gmail.com>]

> Changes for APR 1.4.3
>   *) configure: Make definition of apr_ino_t independent of
>      _FILE_OFFSET_BITS even on platforms where ino_t is 'unsigned int'.
>      [Stefan Fritsch]

To summarize, it looks like Apache2::Build::strip_lfs() breaks with Perl
5.14 with -Duselargefiles on 32-bit architectures, and is not necessary
since at least apr 1.4.3, possibly earlier.

I'd like input on whether we should expect further pitfalls if we
build mod_perl2 with -D_LARGEFILE_SOURCE -D_FILE_OFFSET_BITS=64 on
Debian, i.e. stop stripping those flags in Apache2::Build.

Obviously, a more portable solution is needed for mod_perl 2.0.6.
Perhaps an explicit probe for sizeof(apr_ino_t) with different
_FILE_OFFSET_BITS definitions?

Cheers,
-- 
Niko Tyni   [email protected]