Ceph » CephFS

At first, we thought that NFS Command doubling is due to the import of NFS manager module into the Rook manager module (`rook_cluster.py`) where it is used to cast `mgr` to NFS to write a rados object at:

NFSRados(mgr_module.rados, service_id).write_obj('', f'conf-nfs.{spec.service_id}')

(here mgr_module is `mgr` casted to NFS module)

but this can be done without casting too because the `rados` is not specific to NFS module as it comes from the `MgrModule` that `nfs.module` and the `mgr` (which is of type `RookOrchestrator`) also inherits.

Therefore, John proposed some ideas to tackle this:

  1 "quick fix" if we simply remove the `from nfs.module import Module` line and fix the lines that use(d) it.
  2 Longer term it may be worth considering an approach where manager modules talk to each other using the {mgr,mon}_command functions.
    Something calling a new command like `nfs cluster pool-init` or like.
  3 Implment some way of blocking imports for unsafe modules? Maybe expose a python function from the C++ side indicating what module it is? Something like:

       if mgr_module_realname() != "nfs":
            raise ImportError("incorrect mgr module")

I tried out 1 but unfortunately it didn't work, reason being import of `nfs.module` in `nfs.__init__` [0] i.e. anything that imports any part of NFS will trigger the CLICommands, thus there seems to be no simple fix to get rid of this unless we try 2 or 3 but it would take some efforts and possibly some foundational changes to the way we treat imports. Therefore it was decided to keep this aside for sometime until we find some strong reason to pick this up.

Special shout-out to John Mulligan for all the debugging and helping find the root cause, most of this discussion took place in the Orchestration team weekly meeting on 6th June '23.

[0] https://github.com/ceph/ceph/blob/fd6c853512109a1e83a419c3c55d4316d8f6d010/src/pybind/mgr/nfs/__init__.py#L7