Re: [PATCH 2/2] NFS: add a clientid mount option

[Trond Myklebust <trondmy@kernel.org>]

On Mon, 2025-07-14 at 10:47 -0400, Chuck Lever wrote:
> On 7/14/25 9:31 AM, Christoph Hellwig wrote:
> > On Mon, Jul 14, 2025 at 09:24:20AM -0400, Chuck Lever wrote:
> > > On 7/14/25 2:30 AM, Christoph Hellwig wrote:
> > > > Add a mount option to set a clientid, similarly to how it can
> > > > be
> > > > configured through the per-netfs sysfs file.  This allows for
> > > > easy
> > > > testing of behavior that relies on the client ID likes locks or
> > > > delegations with having to resort to separate VMs or
> > > > containers.
> > > 
> > > The problem with approaches like this is that it becomes
> > > difficult
> > > to manage multiple mounts of the same server. Each of those
> > > mounts
> > > really cannot have a different clientid.
> > 
> > Having different clientids for multiple mounts from the same server
> > is the purpose and only reason for this option.
> 
> It would be helpful to explain exactly what test you are trying to do
> or
> what bug you are trying to explore. I can't think of a way that the
> current client code base would ever need to behave this way. So I
> assume
> you are trying to test some kind of server behavior. If that's the
> case,
> why not craft one or more pynfs test cases? (Or, maybe pynfs already
> handles this case).
> 
> 
> > > For testing, why can't you use the per-container clientid
> > > setting?
> > 
> > Because having to create a container is a lot of effort when all
> > that is needed is just a mount with a different clientid.
> 
> Since this is for development testing (?) I am hesitant to endorse
> adding it as part of the everyday administrative interface.
> Especially
> since this will break things (on purpose, of course). I don't relish
> having to support administrators coming to us complaining that some
> unimagined future use case is not working with the clientid= mount
> option.
> 
> If clientid= does get merged, though, what is your plan for an nfs(5)
> update?
> 

There is a lot of potential for tripping over your own shoelaces with
this mount option.

I can't think of any circumstances where an ordinary user should need
to set a different client identifier depending on the server. I too am
therefore sceptical that anyone will need this functionality other than
for kernel development purposes. It requires very deep knowledge of the
NFSv4 protocol both to understand what it does, and to stay out of
trouble when using it.

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