On Thu, Apr 25, 2019 at 03:00:22PM +0000, Trond Myklebust wrote: > The assumption is that if you have enough privileges to mount a > filesystem using the NFS client, then you would also have enough > privileges to run a userspace client, so there is little point in > restricting the NFS client. > > So the guiding principle is that a NFS client mount that is started in > a container should behave as if it were started by a process in a "real > VM". That means that the root uid/gid in the container maps to a root > uid/gid on the wire. > Ditto, if there is a need to run the idmapper in the container, then > the expectation is that processes running as 'user' with uid 'u', will > see their creds mapped correctly by the idmapper. Again, that's what > you would see if the process were running in a VM instead of a > container. > > Does that all make sense? Yes, thanks! I thought there was a problem that the idmapper depended on keyring usermodehelper calls that it was hard to pass namespace information to. Did that get fixed and I missed it or or forgot? > > So this means that any orchestrator software which may be setting up > NFS mounts as part of setting up the container has 2 options: > > 1. Either perform the mount outside the container namespaces, in which > case the container process uids/gids are mapped from their > user_namespace into the user_namespace of the orchestrator, and the > uids/gids on the wire will reflect those mapped uids/gids (so uid 0 > in the container will be mapped to uid xxx where xxx is decided by > the orchestrator). > 2. Perform the mount inside the container namespaces, in which case the > container process uids/gids go on the wire as-is. OK, great, that sounds perfect. --b.
