On Mon, 2022-02-28 at 11:07 -0500, Benjamin Coddington wrote: > On 17 Feb 2022, at 7:48, Benjamin Coddington wrote: > > > Trond and Anna, this patch makes obvious one problem with the udev > > approach. > > We cannot depend fully on stable uniquifiers. The conversation on > > the > > userspace side has come full circle to asserting we use a mount > > option. > > > > Do you still want us to keep this approach, or will you accept a > > mount > > option as: > > - first mount in a namespace sets the identifier > > - subsequent mounts with conflicting identifiers return an error > > > > If we keep this udev approach, this patch isn't necessary but does > > greatly > > reduce the chances of having clients with unstable identifiers. > > > > No one can be blamed for ignoring this, but it would be a relief to > > get this > > problem solved so it can stop burning our time. > > > > Please advise, > > Ben > > Attempts to work toward solutions in this area seem to be ignored, > the > questions still stand. Until we can sort out and agree on a > direction, > self-NACK to this patch. A new mount option doesn't solve any problems that we can't solve with the existing framework. The problem we're trying to solve isn't "How do I set the stable uniquifier in a container?" You can already do that by manually setting the correct value in /sys/fs/nfs/net/nfs_client/identifier. There is no need for a new kernel mount option. The real problem that we are trying to solve is: "How do I ensure that a stable uniquifier is always set by the user?" While you could have "mount" do that, it is not going to be solved by adding a mandatory mount option. It is far better to have a separate helper that generates and persists that stable uniquifier in a well- known location. If you want to call that helper from 'mount' when it sees that the /sys/fs/... pseudo-file is empty, then by all means, however the point is that the solution needs to apply to existing setups without any further changes. I'm happy to set up a kernel patch that initialises the pseudo-file based uniquifier for init_net from the module parameter. That could help ensure that we won't override any existing setting for the host. -- Trond Myklebust Linux NFS client maintainer, Hammerspace trond.myklebust@xxxxxxxxxxxxxxx
