Re: [PATCH] nfsd: Fix nfs4_disable_idmapping option

[Chuck Lever]

On Fri, Sep 13, 2024 at 09:03:00AM +1000, NeilBrown wrote:
> On Fri, 13 Sep 2024, Pali Rohár wrote:
> > On Friday 13 September 2024 08:26:02 NeilBrown wrote:
> > > On Fri, 13 Sep 2024, Pali Rohár wrote:
> > > > NFSv4 server option nfs4_disable_idmapping says that it turn off server's
> > > > NFSv4 idmapping when using 'sec=sys'. But it also turns idmapping off also
> > > > for 'sec=none'.
> > > > 
> > > > NFSv4 client option nfs4_disable_idmapping says same thing and really it
> > > > turns the NFSv4 idmapping only for 'sec=sys'.
> > > > 
> > > > Fix the NFSv4 server option nfs4_disable_idmapping to turn off idmapping
> > > > only for 'sec=sys'. This aligns the server nfs4_disable_idmapping option
> > > > with its description and also aligns behavior with the client option.
> > > 
> > > Why do you think this is the right approach?
> > 
> > I thought it because client has same configuration option, client is
> > already doing it, client documentation says it and also server
> > documentation says it. I just saw too many pieces which agreed on it and
> > just server implementation did not follow it.
> > 
> > And to make mapping usable, both sides (client and server) have to agree
> > on the configuration.
> > 
> > So instead of changing also client and client's documentation it is
> > easier to just fix the server.
> > 
> > > If the documentation says "turn off when sec=sys" and the implementation
> > > does "turn off when sec=sys or sec=none" then I agree that something
> > > needs to be fixed.  I would suggest that the documentation should be
> > > fixed.
> > > 
> > > From the perspective of id mapping, sec=none is similar to sec=sys.
> > 
> > It is similar, but quite different. With sec=sys client sends its uid
> > and list of gids in every (RPC) packet and server authenticate client
> > (and do mapping) based on it. With sec=none client does not send any uid
> > or gid. So mostly uid/gid form is tight to sec=sys.
> > 
> 
> With sec=none I don't think that any mapping makes sense except to map
> all uids and gids to "none" or similar.

I tend to agree that "sec=none" on the server should be akin to
squashing all RPC users to the local "nobody" identity. But I
haven't looked closely at NFSD's implementation of this security
flavor.


> The documented purpose of nfs4_disable_idmapping is to "ease migration
> from NFSv2/v3".  That suggests that where relevant it should behave
> mostly like v2/v3.
> 
> I don't feel strongly about this.  You appear to be actually using
> AUTH_NONE authentication.  What behaviour would work best for your
> use-case, and why?

Or to ask another way, what isn't working for you, exactly?

The problem statement above says "The server doesn't work like the
client" but does not explain why that's a problem.

-- 
Chuck Lever