[yocto] perl 5.22 and 32 bit targets

Wed Mar 23 10:25:15 PDT 2016

On 2016-03-23 13:17, Jens Rehsack wrote:
>
>> Am 23.03.2016 um 13:07 schrieb Gary Thomas <gary at mlbassoc.com>:
>>
>> On 2016-03-23 12:43, Jens Rehsack wrote:
>>>
>>>> Am 23.03.2016 um 12:05 schrieb Gary Thomas <gary at mlbassoc.com>:
>>>>
>>>> On 2016-03-23 10:48, Jens Rehsack wrote:
>>>>>
>>>>>> Am 23.03.2016 um 10:14 schrieb Jens Rehsack <rehsack at gmail.com>:
>>>>>>
>>>>>>>
>>>>>>> Am 23.03.2016 um 10:09 schrieb Gary Thomas <gary at mlbassoc.com>:
>>>>>>>
>>>>>>> On 2016-03-23 09:57, Jens Rehsack wrote:
>>>>>>>>
>>>>>>>>> Am 23.03.2016 um 09:40 schrieb Gary Thomas <gary at mlbassoc.com>:
>>>>>>>>>
>>>>>>>>> On 2016-03-23 09:09, Gary Thomas wrote:
>>>>>>>>>> On 2016-03-23 06:36, Khem Raj wrote:
>>>>>>>>>>> On Tue, Mar 22, 2016 at 9:53 PM, Gary Thomas <gary at mlbassoc.com> wrote:
>>>>>>>>>>>> I hope this is the correct place to discuss this problem.  It
>>>>>>>>>>>> is all about a difference in behavior between a program built
>>>>>>>>>>>> using bitbake/OE (only OE-core is needed) vs building the program
>>>>>>>>>>>> on the target hardware itself.
>>>>>>>>>>>>
>>>>>>>>>>>> I've been struggling with this problem since perl was upgraded
>>>>>>>>>>>> to version 5.22.  I'm working on Amanda (Advanced Maryland Archive
>>>>>>>>>>>> tool) which is written primarily in perl and uses swig interfaces
>>>>>>>>>>>> to access native C functions.  This code works great when using
>>>>>>>>>>>> the previous perl (5.20.x) but fails on all 32 bit targets with
>>>>>>>>>>>> perl 5.22
>>>>>>>>>>>>
>>>>>>>>>>>> The interesting thing is that if I build Amanda on my target
>>>>>>>>>>>> directly (using SDK tools), it works perfectly even with perl
>>>>>>>>>>>> 5.22, so it seems that there is some [subtle] difference between
>>>>>>>>>>>> building using bitbake/OE than when built on the self-hosted
>>>>>>>>>>>> target.  I've compared the builds and the only thing I could
>>>>>>>>>>>> find (from the output of configure) is a difference in sizeof(off_t)
>>>>>>>>>>>> Sadly, when I tried to adjust this in the OE build, it didn't
>>>>>>>>>>>> make any difference, but perhaps I didn't make this change
>>>>>>>>>>>> correctly or completely.
>>>>>>>>>>>
>>>>>>>>>>> do you have largefile support turned on ? if you do then it might
>>>>>>>>>>> be detecting it wrongly during configure since we cache it to a
>>>>>>>>>>> non-largefile case
>>>>>>>>>>>
>>>>>>>>>>> so try to add something like
>>>>>>>>>>>
>>>>>>>>>>> EXTRA_OECONF += "${@bb.utils.contains('DISTRO_FEATURES', 'largefile',
>>>>>>>>>>> 'ac_cv_sizeof_off_t=8', '', d)}"
>>>>>>>>>>>
>>>>>>>>>>> while building perl or the affected program and see if that helps
>>>>>>>>>>
>>>>>>>>>> Thanks for the idea, but that didn't help.  I also forced some CFLAGS
>>>>>>>>>> to match, in particular:
>>>>>>>>>>   -D_LARGEFILE_SOURCE -D_FILE_OFFSET_BITS=64
>>>>>>>>>> but this didn't make any difference either.
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>> On a whim I just tried a little experiment where I took the *.o files
>>>>>>>>> from the perl subdirectory (where all the swig shims live) from a working
>>>>>>>>> (self-hosted) build and moved them to my bitbake/OE build.  I then touched
>>>>>>>>> all the *.o and *.lo files in the perl tree to force a relink. I then ran
>>>>>>>>> % bitbake amanda -C compile && bitbake core-image-base
>>>>>>>>> to my surprise, amanda works!  So the culprit lies somewhere within the
>>>>>>>>> swig generated glue.  I've tried comparing these files before and I didn't
>>>>>>>>> find anything other than cosmetic differences (mostly comments about the
>>>>>>>>> name of the file processed, etc).  I've added this subtree to "results"
>>>>>>>>> in my github layer in case someone can see what might be relevant.
>>>>>>>>>
>>>>>>>>> Any ideas what might be different and make this swig generated glue fail?
>>>>>>>>> Note that the swig interface files are rebuilt as part of the build process
>>>>>>>>> and both bitbake/OE and self-hosted are using the same swig version.
>>>>>>>>
>>>>>>>> I digged a bit through your layer (while my up2date scanner over meta-cpan
>>>>>>>> blocks my build chain :P) and realized that you use perl-5.20.0 as it was
>>>>>>>> in poky. A "simple" downgrade would be more reasonable ... if reason applys
>>>>>>>> here in general :)
>>>>>>>
>>>>>>> In practice, I am doing that.  However, I want to understand why perl 5.22
>>>>>>> breaks things and get it fixed.
>>>>>>
>>>>>> I did a diff between your 5.20 and poky's 5.22 and realize some fixes applied
>>>>>> in 5.22 regarding library path's aren't applied in your copy. Maybe swig relies
>>>>>> on wrong library locations and when we know, we can fix.
>>>>>>
>>>>>> So it's maybe not a 5.20 vs. 5.22 problem, it's maybe a weird swig setup problem.
>>>>>>
>>>>>>>> When you fail on cross-build and succeed in target build, try to compare the
>>>>>>>> C files and includes (even swig libraries) used.
>>>>>>>>
>>>>>>>> It smells more like a "wrong source" than a "perl problem" (and even when
>>>>>>>> I never would read any python thread, the same problem would likely occur
>>>>>>>> there, too ^^).
>>>>>>>>
>>>>>>>> Which perl headers are used in your build? To dig down, more logs would
>>>>>>>> be reasonable ...
>>>>>>>
>>>>>>> Everything comes from the same sources, same revisions, etc, as I'm using
>>>>>>> either a bitbake/OE build or the embedded (self-hosted) version from the
>>>>>>> same build plus SDK tools.
>>>>>>
>>>>>> And your SDK does not include any host tools? Did you prove the intermediate
>>>>>> amanda build files (eg. generated by SWIG) for relicts from wrong source?
>>>>>> Did you check the logs which include directories had been used?
>>>>>
>>>>> I give it a quick shot and got:
>>>>>
>>>>> ../../arm-poky-linux-gnueabi-libtool  --tag=CC   --mode=compile arm-poky-linux-gnueabi-gcc  -march=armv7-a -marm  -mthumb-interwork -mfloat-abi=hard -mfpu=neon -mtune=cortex-a9 --sysroot=/homes/sno/fsl-release-bsp/ornithologen-kann-man-mit-voegeln-eine-freude-machen/tmp/sysroots/curie -DHAVE_CONFIG_H -I. -I../../config -I../../common-src -I../../common-src -I../../xfer-src -I../../gnulib -I../../ndmp-src -I/homes/sno/fsl-release-bsp/ornithologen-kann-man-mit-voegeln-eine-freude-machen/tmp/sysroots/x86_64-linux/usr/lib/perl-native/perl/5.22.1/CORE   -fno-strict-aliasing -D_GNU_SOURCE -pthread -I/homes/sno/fsl-release-bsp/ornithologen-kann-man-mit-voegeln-eine-freude-machen/tmp/sysroots/curie/usr/include/glib-2.0 -I/homes/sno/fsl-release-bsp/ornithologen-kann-man-mit-voegeln-eine-freude-machen/tmp/sysroots/curie/usr/lib/glib-2.0/include -DSWIG  -O2 -pipe -g -feliminate-unused-debug-types -fipa-pta -ftree-partial-pre -ftree-loop-distribution -ftree-loop-distribute-patterns -fgcse-
 a
f
>> t
>> e
>>>> r
>>>> -reload -fgcse-sm -fgcse-las  -fno-strict-aliasing  -c -o xferwrap.lo xferwrap.c
>>>>>
>>>>> Do you see the "-I/homes/sno/fsl-release-bsp/ornithologen-kann-man-mit-voegeln-eine-freude-machen/tmp/sysroots/x86_64-linux/usr/lib/perl-native/perl/5.22.1/CORE" ?
>>>>>
>>>>> Any further blames of perl-5.22? :)
>>>>
>>>> I'm not sure what you are getting at here.  Indeed it is supposed to be
>>>> using perl-5.22.1 for everything.  The version 5.22.0 in meta-amanda layer
>>>> is just for testing and should not be used at the moment.
>>>
>>> Please compare word wise ...
>>>
>>> -I/homes/sno/fsl-release-bsp/ornithologen-kann-man-mit-voegeln-eine-freude-machen/tmp/sysroots/x86_64-linux/usr/lib/perl-native/perl/5.22.1/CORE
>>> -I/homes/sno/fsl-release-bsp/ornithologen-kann-man-mit-voegeln-eine-freude-machen/tmp/sysroots/curie/usr/include/glib-2.0
>>>
>>> Probably this helps getting my point:
>>>
>>> $ ls -l tmp/work
>>> total 149
>>> drwxrwxr-x  54 sno sno  54 Mar 22 23:49 all-poky-linux
>>> drwxrwxr-x 269 sno sno 269 Mar 22 23:49 arm926ejse-poky-linux-gnueabi
>>> drwxrwxr-x   3 sno sno   3 Mar 22 23:48 bohr-nand-poky-linux-gnueabi
>>> drwxrwxr-x  19 sno sno  19 Mar 22 23:49 bohr-poky-linux-gnueabi
>>> drwxrwxr-x   9 sno sno   9 Mar 22 21:38 cortexa9hf-vfp-neon-mx6qdl-poky-linux-gnueabi
>>> drwxrwxr-x 326 sno sno 326 Mar 23 11:28 cortexa9hf-vfp-neon-poky-linux-gnueabi
>>> drwxrwxr-x   3 sno sno   3 Mar 22 21:37 curie-emmc-poky-linux-gnueabi
>>> drwxrwxr-x  29 sno sno  29 Mar 22 21:50 curie-poky-linux-gnueabi
>>> drwxrwxr-x 196 sno sno 196 Mar 23 11:28 x86_64-linux
>>>
>>>> All of my comparisons and testing (see instructions in the README) are
>>>> for perl-5.22.1 only and in fact I've only been using the bits from bitbake/OE
>>>> and not building that any more (I found that it wasn't necessary - one only
>>>> needs to build amanda on the target)
>>>>
>>>> Sorry for the confusion.
>>>
>>
>> Huzzah!  This seems to be on the right track.  I adjusted the PERL_INC
>> value in perl/Makefile (the *.h files used by the swig interfaces) to
>> point to the target ones and not the perl-native ones and now it works :-)
>>
>> I'm not sure what's different between the two sets of include files, and
>> more interestingly why it worked with previous versions of perl, but this
>> does seem to be the culprit.
>>
>> Now I just need to figure out how to convince the 'configure' process to
>> do it correctly.
>
> IIRC, rrdtool is a good example to do it right. collectd had the same problem
> as amanda had - wrong perl used ...
> Maybe vice versa - I'm getting old and forget to remember :D

That did it.  I just had to add this line to my recipe
   export PERL_INC = "-I${STAGING_LIBDIR}${PERL_OWN_DIR}/perl/${@get_perl_version(d)}/CORE"
and a small patch to make it honor it (instead of always building the wrong path).  Now
it builds and runs great on a 32 bit target.

-- 
------------------------------------------------------------
Gary Thomas                 |  Consulting for the
MLB Associates              |    Embedded world
------------------------------------------------------------