Comment # 3 on bug 990356 from Neil Brown

> I agree with rpc.idmpad being mandatory, not sure about rpc.mountd however.

Having myself written a large part of the code with which the kernel
communicates with mountd, I can assure you without a shadow of doubt that
rpc.mountd is required for any NFS service.
The Redhat documentation you found is correct.  The "over-the-wire" reference
means that the NFS client doesn't communication over the network directly to
mountd, it only communicates with the kernel.  But the kernel definitely
communicates with mountd.

If the NFS server systemd unit isn't starting when rpcbind.socket is masked,
that suggests that nfsserver.service still 'Requires' rpcbind, rather than
'Wants' it.

This was fixed in late May 2016
What does:

 rpm -q --changelog nfs-kernel-server | head

report?  If it doesn't contain

* Tue May 24 2016 nfbrown@suse.com
- 0001-systemd-Decouple-the-starting-and-stopping-of-rpcbin.patch
  0002-systemd-unit-files-fix-up-dependencies-on-rpcbind.patch
  Fix systemd dependencies to ensure rpcbind is started when needed.
  (bsc#975265)

then it is not up-to-date.  If it does, then this should work and I'll need to
look deeper.