[Pkg-scicomp-devel] Bug#459025: Bug#459025: parmetis_3.1-8+b1(unstable/amd64/xenophanes): FTBFS: Doesn't search for mpi.h in right include path

[Ondrej Certik]

On Jan 4, 2008 10:15 AM, Marc 'HE' Brockschmidt
<marc at marcbrockschmidt.de> wrote:
> Package: parmetis
> Version: 3.1-8+b1
> Severity: serious
>
> Heya,
>
> I tried rebuilding parmetis against the new mpich packages, so that the
> whole heap of packages can migrate to testing. This failed:
>
> | Automatic build of parmetis_3.1-8+b1 on xenophanes by sbuild/amd64 98-farm
> | Build started at 20080104-0031
> | ******************************************************************************
>
> [...]
>
> | make[2]: Entering directory `/build/buildd/parmetis-3.1/METISLib'
> | gcc -I. -I.. -I/build/buildd/parmetis-3.1/ -O3 -c coarsen.c
> | In file included from ./metis.h:27,
> |                  from coarsen.c:13:
> | ../parmetis.h:17:17: error: mpi.h: No such file or directory
> | In file included from ./metis.h:27,
> |                  from coarsen.c:13:
> | ../parmetis.h:59: error: expected declaration specifiers or '...' before 'MPI_Comm'
> [...]
>
> A complete build log can be found at
> http://experimental.debian.net/build.php?arch=amd64&pkg=parmetis&ver=3.1-8+b1
>
> I guess you will need to update the build scripts to look into
> /usr/lib/openmpi/include.

Thanks a lot!

The problem is, that openmpi package moved the *.h files out of
/usr/include to /usr/lib and created
a link from /usr/include/openmpi. However, that link doesn't always
work. And if the link doesn't work,
it's a FHS violation -  *.h files should be in /usr/include/... not in /usr/lib.

I reported this against the openmpi package here:

http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=457088

But I was not accepted friendly. Unfortunately, I just checked on my
computer, and /usr/include/openmpi
doesn't work, so I will have to report it again. Part of the problem
is probably in the broken update-alternatives
package (that sometimes fails to update the link), but nevertheless
it's the openmpi package, that doens't work.

So I think I should report it against the openmpi package, right?

Ondrej