[arch-dev-public] i686 pkg built in chroot on x86_64 doesn't work

[Paul Mattal]

On 03/09/2010 12:19 AM, Allan McRae wrote:
> On 09/03/10 14:40, Daniel J Griffiths (Ghost1227) wrote:
>> On 03/08/10 at 11:32pm, Paul Mattal wrote:
>>> On 03/08/2010 07:16 PM, Allan McRae wrote:
>>>> On 09/03/10 08:38, Daniel J Griffiths (Ghost1227) wrote:
>>>>> On 03/08/10 at 05:20pm, Paul Mattal wrote:
>>>>>> On 03/07/2010 02:33 PM, Paul Mattal wrote:
>>>>>>> On 02/25/2010 11:49 AM, Aaron Griffin wrote:
>>>>>>>> On Tue, Feb 23, 2010 at 7:17 PM, Daniel J Griffiths (Ghost1227)
>>>>>>>> <ghost1227 at archlinux.us> wrote:
>>>>>>>>> I've always thought the method of modifying your local mirrorlist,
>>>>>>>>> running mkarchroot, then reverting the changes to be more tedious
>>>>>>>>> than
>>>>>>>>> necessary for creation of i686 chroots on x86_64. My recent
>>>>>>>>> work with
>>>>>>>>> setting up a dedicated build server gave me plenty of time and an
>>>>>>>>> excuse to actually do something about it. As such, I've put
>>>>>>>>> together a
>>>>>>>>> little patch that allows specification of creation of an i686
>>>>>>>>> chroot
>>>>>>>>> at runtime. When set, this flag will automatically modify your
>>>>>>>>> local
>>>>>>>>> mirrorlist, create the requested i686 chroot, then revert the
>>>>>>>>> changes
>>>>>>>>> to the mirrorlist file. I don't know if others would find this
>>>>>>>>> useful
>>>>>>>>> or not, but I figured it wouldn't hurt to post it.
>>>>>>>>
>>>>>>>> Actually, I *thought* I added a flag to point to an alternate
>>>>>>>> pacman
>>>>>>>> config when building the chroot, to simplify this. So it'd be as
>>>>>>>> simple as:
>>>>>>>>
>>>>>>>> sed s/x86_64/i686/< /etc/pacman.d/mirrorlist>
>>>>>>>> /etc/pacman.d/mirrorlist-i686
>>>>>>>> sed s/mirrorlist/mirrorlist-i686/< /etc/pacman.conf>
>>>>>>>> /etc/pacman-i686.conf
>>>>>>>>
>>>>>>>> mkarchroot -C /etc/pacman-i686.conf ...yada yada...
>>>>>>>
>>>>>>> Having some uniform turnkey script to build an i686 chroot on an
>>>>>>> x86_64
>>>>>>> box as part of devtools would be useful.
>>>>>>>
>>>>>>> I just noticed today that in setting up my chroots, I had
>>>>>>> replaced my
>>>>>>> x86_64 in makepkg.conf with i686, but not replaced the x86-64 which
>>>>>>> occur in CFLAGS and CXXFLAGS. If others have done similar things, it
>>>>>>> might result in buggy or suboptimal packages.
>>>>>>>
>>>>>>> It would at least be nice of someone who knows a lot about
>>>>>>> building in
>>>>>>> chroots describes in detail what must be done for i686 chroots in
>>>>>>> x86_64
>>>>>>> in the wiki page, just in case there are important details I or
>>>>>>> others
>>>>>>> have missed.
>>>>>>
>>>>>> So here's one for the chroot gurus.
>>>>>>
>>>>>> It appears that even with all my settings fixed, tomcat doesn't work
>>>>>> when built in my i686 chroot on x86_64:
>>>>>>
>>>>>> http://bugs.archlinux.org/task/18604
>>>>>>
>>>>>> Building on an actual i686 box, even in a chroot, works.
>>>>>>
>>>>>> Can anyone guess why this might be? Are there some guidelines
>>>>>> someone can give for evaluating whether or not it's safe to build a
>>>>>> package under an i686 chroot on an x86_64 box?
>>>>>>
>>>>>> It seems there's an increased risk we're putting out broken packages
>>>>>> when we build i686 packages in a chroot on an x86_64 box.
>>>>>>
>>>>>> - P
>>>>> I've seen a few (rare) cases where a package built for i686 on an
>>>>> x86_64
>>>>> machine _must_ be run with linux32, just using a 32bit chroot doesn't
>>>>> cut it. This could be one of those cases. On a side note, perhaps I
>>>>> should
>>>>> add a flag on pkgbuild.com to enable this just for those off-the-wall
>>>>> cases...
>>>>
>>>> I'd say to always use linux32... You can get some very strange
>>>> configure
>>>> errors without it. I have "makechrootpkg64" as an alias to use "linux64
>>>> makechrootpkg" on my system.
>>>
>>> I've added linux32 to my build scripts on my x86_64 box, since this
>>> seems like a good idea in general; is there any reason *not* to run
>>> makechrootpkg for an i686 chroot on an x86_64 box with linux32 all
>>> the time? I could try to tool up a patch to makechrootpkg to
>>> automatically detect/do that.
>>>
>>> However, adding it doesn't solve the tomcat building problem; same
>>> result. This must be some other issue.
>>>
>>> - P
>> OK... take two at an explanation... the go package won't build in an
>> i686 chroot because the
>> build scripts natively use uname to determine arch... possibly
>> something similar in tomcat?
>
> linux32 will change uname appropriately.
>
> allan at arch ~
>  > linux32 uname -m
> i686
>
> allan at arch ~
>  > linux64 uname -m
> x86_64

So this is a weird one, but I think I just made some headway.

Apparently, this seems to be related to the ACTUAL CPU of the box I 
build on. When building on an i686 Arch box running on an AMD 64 CPU, 
the build generates the broken binaries! When I build on an Intel CPU, 
it works fine.

So it would appear this is at least in part Tomcat's build process's 
fault for not building the right thing based on the uname -m but rather 
looking directly at the processor.

Yuck. Perhaps I will at least add a note to the PKGBUILD suggesting this 
issue, so that the next poor person who decides to update tomcat doesn't 
trip over this same thing.

Thanks, all, for working this through with me.

- P