Comment (by Max Morris):

When building HDF5 via the ExternalThorn, only serial HDF5 is built, so the issue is sidestepped. See src/build.sh:77 in ExternalThorn-MPI, there is no --enable-parallel flag when calling ./configure and I don't see any way for that flag to arrive there. The only instance where parallel HDF5 is linked against MPI is during the final Cactus linkage (ExternalThorn-MPI's src/detect.sh:269 checks for parallel HDF5 and adds MPI libdirs to its own), by which point hwloc is present on the link line anyway as I mentioned previously. If ExternalThorn-HDF5 were to build with --enable-parallel then I suspect it would suffer from the same issue as PETSc.

I still think you are right that the fix ought to live in the MPI thorn. I am currently building Cactus with such a fix to make sure it works before submitting a new PR.

--
Ticket URL: https://bitbucket.org/einsteintoolkit/tickets/issues/2969/externallibraries-petsc-link-hwloc-into