[ovs-dev] [PATCH] m4: Fix 32bit build with 64bit version of libunwind-dev installed.

Ilya Maximets i.maximets at ovn.org
Wed Oct 2 15:46:36 UTC 2019


On 02.10.2019 17:46, Ben Pfaff wrote:
> On Tue, Oct 01, 2019 at 06:40:10PM -0700, William Tu wrote:
>> On Tue, Oct 1, 2019 at 4:19 PM William Tu <u9012063 at gmail.com> wrote:
>>>
>>> On Tue, Oct 01, 2019 at 08:29:41PM +0300, Ilya Maximets wrote:
>>>> On 01.10.2019 19:56, Ben Pfaff wrote:
>>>>> On Tue, Oct 01, 2019 at 07:46:10PM +0300, Ilya Maximets wrote:
>>>>>> On 01.10.2019 19:17, Ben Pfaff wrote:
>>>>>>> On Tue, Oct 01, 2019 at 02:22:20PM +0300, Ilya Maximets wrote:
>>>>>>>> 32 and 64 bit versions of libunwind-dev package can not be installed
>>>>>>>> at the same time at least in Debian and Ubuntu.  AC_CHECK_LIB macro
>>>>>>>> doesn't know if we have '-m32' in OVS_CFLAGS, so we need to
>>>>>>>> additionally check for actual compiling possibility before enabling
>>>>>>>> the feature.  Otherwise 'make' will fail due to inability to
>>>>>>>> build with 64bit libunwind headers:
>>>>>>>>
>>>>>>>>       https://travis-ci.org/openvswitch/ovs/jobs/590580374
>>>>>>>>
>>>>>>>>       /usr/include/libunwind.h:23:28: fatal error:
>>>>>>>>           libunwind-x86.h: No such file or directory
>>>>>>>>
>>>>>>>> CC: William Tu <u9012063 at gmail.com>
>>>>>>>> Fixes: e2ed6fbeb18c ("fatal-signal: Catch SIGSEGV and print backtrace.")
>>>>>>>> Signed-off-by: Ilya Maximets <i.maximets at ovn.org>
>>>>>>>
>>>>>>> I think there's some confusion here on the use of options.  CFLAGS is
>>>>>>> for the user, OVS_CFLAGS is for the OVS configure script to use itself.
>>>>>>> .cirrus.yml and .travis/linux-build.sh are setting OVS_CFLAGS, but they
>>>>>>> should not because they are the "user", not the configure script.  If
>>>>>>> they set CFLAGS, instead, I believe that there would be no problem here.
>>>>>>>
>>>>>>> See commit 95626395a3db ("build: Respect CFLAGS and LDFLAGS passed to
>>>>>>> make") and the documentation it references.
>>>>>>
>>>>>> That's an interesting point.
>>>>>> However, 'configure' script makes decision about building with libunwind
>>>>>> and it needs to know  if we're going to build 32 or 64 bit binary to
>>>>>> enable or disable the feature. Setting CFLAGS for configure script breaks
>>>>>> it in a following way:
>>>>>>
>>>>>> $ CFLAGS='-m32' ./configure
>>>>>> ...
>>>>>> checking for gcc... gcc
>>>>>> checking whether the C compiler works... no
>>>>>
>>>>> Is this because the 32-bit C library isn't installed?  It should be easy
>>>>> to fix with "sudo dnf install glibc-devel.i686" or the distro-specific
>>>>> equivalent.
>>>>
>>>> Yes, you're right. It works after installation of gcc-multilib.
>>>>
>>>> I'll think about how to rework CI scripts tomorrow.
>>>>
>>>>>
>>>>> I don't see how it's reasonable to run "configure" with a toolchain
>>>>> different from the one we're going to use for building.  Adding -m32
>>>>> changes the ABI and the API.
>>>>>
>>>>> My usual way to do this is actually to use CC="gcc -m32".  That way the
>>>>> -m32, which is important, can't get separated at all!
>>>
>>> Thanks Ben and Ilya,
>>>
>>> I don't have a good solution, on my Xenial system, if I do 32-bit build
>>>
>>>    apt-get install libunwind-dev:i386
>>>    CC="gcc -m32" ./configure
>>>
>>> then it correctly detects libunwind and pass, for x86_64, I have to
>>>
>>>    apt-get remove libunwind-dev:i386
>>>    apt-get install libunwind-dev
>>>    ./configure
>>>
>>> so like Ilya mentioned, the 32 and 64 bit libunwind-dev can not be installed
>>> at the same time. To pass travis, we need to check whether it has '-m32' or not,
>>> then install different libunwind-dev package.
>>>
>>> I'm also ok with disabling this feature on 32-bit system.
>>>
>>> Regards,
>>> William
>>>
>> Hi,
>>
>> How about this fix? So for 32-bit system, we have to install libunwind-dev:i386,
>> in the meantime the libunwind-dev for 64-bit will be removed automatically.
>> I tested it and passed.
>> https://travis-ci.org/williamtu/ovs-travis/builds/592323952
> 
> Acked-by: Ben Pfaff <blp at ovn.org>
> 
> It might be worth waiting for Ilya's feedback, though.

Thanks William and Ben,

I think that we need both solutions.  Conditional installing of
32bit library will allow us to test with it.  However, I think,
that we need to eliminate usage of OVS_CFLAGS in our CI scripts
too, because this gives a bad example for users.  I personally
always used OVS_CFLAGS for passing '-Wno-cast-align' while
building with clang and dpdk, probably, just because first time
I saw that in CI scripts. I'm working on the fix.

The change itself is fine. One small nit is that maybe it worth
to move this command to linux-prepare.sh script to keep all the
external software related stuff in the same place?

Best regards, Ilya Maximets.


More information about the dev mailing list