Building OpenMPI 4.1.1 against PMIx 4.1.0 leads to compilation errors (PMIX_PACKAGE_RANK)

[Micket]

Thank you for taking the time to submit an issue!

Background information

What version of Open MPI are you using? (e.g., v3.0.5, v4.0.2, git branch name and hash, etc.)

OpenMPI v4.1.1
PMIx v4.1.0

Describe how Open MPI was installed (e.g., from a source/distribution tarball, from a git clone, from an operating system distribution package, etc.)

Everything is build from source for the EasyBuild project.

Please describe the system on which you are running

Error seems reproducible on many (all?) platforms. Cf. easybuilders/easybuild-easyconfigs#13668

---

Details of the problem

Building against the new PMIx 4.1.0 I see compilation errors almost immediately

make[2]: Leaving directory `/local/EB/build/OpenMPI/4.1.1/GCC-11.2.0/openmpi-4.1.1/opal'
Making all in mca/common/ofi
make[2]: Entering directory `/local/EB/build/OpenMPI/4.1.1/GCC-11.2.0/openmpi-4.1.1/opal/mca/common/ofi'
  CC       common_ofi.lo
  LN_S     libmca_common_ofi.la
In file included from ../../../../opal/mca/pmix/base/base.h:22,
                 from common_ofi.c:29:
common_ofi.c: In function get_package_rank:
common_ofi.c:321:40: error: PMIX_PACKAGE_RANK undeclared (first use in this function)
  321 |     OPAL_MODEX_RECV_VALUE_OPTIONAL(rc, PMIX_PACKAGE_RANK,
      |                                        ^~~~~~~~~~~~~~~~~
../../../../opal/mca/pmix/pmix.h:153:56: note: in definition of macro OPAL_MODEX_RECV_VALUE_OPTIONAL
  153 |         if (OPAL_SUCCESS == ((r) = opal_pmix.get((p), (s), &(_ilist), &(_kv)))) {        \
      |                                                        ^
common_ofi.c:321:40: note: each undeclared identifier is reported only once for each function it appears in
  321 |     OPAL_MODEX_RECV_VALUE_OPTIONAL(rc, PMIX_PACKAGE_RANK,
      |                                        ^~~~~~~~~~~~~~~~~
../../../../opal/mca/pmix/pmix.h:153:56: note: in definition of macro OPAL_MODEX_RECV_VALUE_OPTIONAL
  153 |         if (OPAL_SUCCESS == ((r) = opal_pmix.get((p), (s), &(_ilist), &(_kv)))) {        \
      |                                                        ^
make[2]: *** [common_ofi.lo] Error 1
make[2]: Leaving directory `/local/EB/build/OpenMPI/4.1.1/GCC-11.2.0/openmpi-4.1.1/opal/mca/common/ofi'
make[1]: *** [all-recursive] Error 1
make[1]: Leaving directory `/local/EB/build/OpenMPI/4.1.1/GCC-11.2.0/openmpi-4.1.1/opal'
make: *** [all-recursive] Error 1

I seems like it might be as simple as a missing include.
Perhaps trying to use PMIx 4 isn't a good idea here? (I'm not sure about how it would interact with pmix3 software, like the slurm builds likely would be using)

[rhc54]

Has to be a missing include as PMIX_PACKAGE_RANK is definitely defined in PMIx v4.1.0. There should be no problem interacting with an earlier version of PMIx if Slurm is using one.

[Micket]

I'm a bit surprised to be the first to see this. I don't believe we are doing anything that special, so I would expect this compilation error to be caught immediately. I'm open to suggestions where the proper place is to patch in an extra include.

[rhc54]

I see the problem. This was backported from OMPI master/v5 where we have gone to direct PMIx symbols. In OMPI v4, we are still abstracting those symbols - i.e., PMIX_FOO becomes OPAL_PMIX_FOO. However, the configure code just checks for the existence of PMIX_PACKAGE_RANK, and then the common/ofi code directly references that symbol.

Probably the simplest fix would be to add the OPAL abstracted form of that symbol and have the code reference it. I'll post a PR to that effect and let folks comment on it.

[kloczek]

Testet ad it works for me.

[jsquyres]

Confirmed: I tested both the v4.1.x branch HEAD and the latest nightly v4.1.x snapshot tarball; both seem to work for me.