rbdplugin: provision seperately from node controller

[pohly]

I'd like to deploy the rbdplugin a bit differently: instead of running it on each node, I want to handle provisioning in one pod that has external-provisioner+external-attacher+rbdplugin. Then rbdplugin can get deployed as NodeServer only on those nodes which are supposed to use it, or not at all (for example, when only testing provisioning).

This kind of deployment is currently possible, but could be simpler: the fix for #86 introduced a call to nsenter.NewNsenter that is executed before starting the gRPC service. It fails unless the provisioner pod has the /rootfs mount, which isn't actually needed. It would be better if the nsenter was done only when needed, for example when a NodeServer.PublishVolume call is invoked. The result could be cached for reuse in future calls.

[ShyamsundarR]

@pohly The current deployment manifests do this exactly. The provisioners are deployed as a deployment with leader election and the nodeplugin is deployed on all nodes. (rbd manifests are provided as an example, cephfs manifests are similar).

I am closing this issue as the current state reflects required intent of this request.