Re: [PATCH 0/3] Enable the setting of a kernel module parameter from nfs.conf

[Trond Myklebust <trondmy@hammerspace.com>]

On Sat, 2021-04-17 at 12:33 -0400, Steve Dickson wrote:
> Hey!
> 
> On 4/15/21 8:40 PM, Trond Myklebust wrote:
> > Here is a skeleton example:
> > 
> > [root@leira rules.d]# cat /etc/udev/rules.d/50-nfs4.rules 
> > ACTION=="add" KERNEL=="nfs_client" ATTR{identifier}=="(null)"
> > PROGRAM="/usr/sbin/nfs4_uuid" ATTR{identifier}="%c"
> > 
> > [root@leira rules.d]# cat /usr/sbin/nfs4_uuid 
> > #!/bin/bash
> > #
> > if [ ! -f /etc/nfs4_uuid ]
> > then
> >         uuid="$(uuidgen -r)"
> >         echo -n ${uuid} > /etc/nfs4_uuid
> > else
> >         uuid="$(cat /etc/nfs4_uuid)"
> > fi
> > echo ${uuid}
> > 
> > 
> > Obviously, the /usr/sbin/nfs4_uuid would need to be fleshed out a
> > little more to ensure that the file /etc/nfs4_uuid actually
> > contains a
> > uuid in the right format, but you get the gist...
> > 
> > With the above additions, I end up with a repeatable
> > 
> > [root@leira rules.d]# modprobe nfs4
> > [root@leira rules.d]# cat /sys/fs/nfs/net/nfs_client/identifier 
> > 7f9f211b-0253-4ef8-a970-b1b0f600af02
> > [root@leira rules.d]# cat /etc/nfs4_uuid 
> > 7f9f211b-0253-4ef8-a970-b1b0f600af02
> 
> I see that this example does populate nfs_client/identifier and
> I'm sure we could beef up the mechanism but the may question
> is this.... 
> 
> How does populating nfs_client/identifier via udev
> actually setting the nfs4_unique_id parameter which is used to set
> the unique id? I look and i've must have missed it...
> 
> If the answer is we need to change the client to look a
> the nfs_client/identifier... then we should get rid of the
> nfs4_unique_id param all together... 
> 

Commit 39d43d164127 ("NFSv4: Use the net namespace uniquifier if it is
set") should default to using the nfs_client identifier if it is set.
Otherwise it falls back to using the nfs4_unique_id. So kernels >= 5.10
are ready to use this udev-based mechanism.

The reason why I added udev support is, as I said, because we need
something that works correctly inside containers. Unfortunately, module
parameters are system-wide, so the older mechanism works just fine
right up to the moment where you fire up 'docker', 'kubernetes' or
'podman' (which is an increasingly important use case for NFS).

We do need something a little more sophisticated than the naive script
that I provided. As I said, that was intended as a skeleton example
just to demonstrate the basic mechanism and how to configure udev. It
would be very good to have something similar to what I showed there be
installed by default when you install nfs-utils.

-- 
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@hammerspace.com