Re: [pnfs] [RFC 0/4] nfs-utils: nfsd support for minor version

["J. Bruce Fields" <bfields@fieldses.org>]

On Fri, Apr 17, 2009 at 08:35:36AM -0400, Steve Dickson wrote:
> J. Bruce Fields wrote:
> > On Thu, Apr 16, 2009 at 03:01:47PM -0400, Steve Dickson wrote:
> >> I do get your point, but as we did with the initial v4 support, 
> >> having the support on by default and then having away to turn it
> >> off is the correct approach... IMHO... 
> > 
> > I'd prefer it be off by default, for the obvious safety reasons.  (It's
> > under rapid development and particularly likely to have bugs).  The only
> > reason we had it on by default before was that we didn't add the
> > switching mechanism early enough.  (Well, and because we could keep it
> > off in the config.  But I'd rather be able to ship users a kernel that
> > supports 4.1 and give them the option of turning it on at runtime, than
> > make them build a new kernel.)
> I agree with not making people recompile kernels, which is the whole
> purpose behind the Fedora repos, but do I think you might be a bit 
> too cautious with exposing the technology. 
> 
> One, I've been running the kernels with everything enabled for a while
> now with no problems whatsoever... A few scary looking warnings now and
> then but nothing major. I also spent the majority of my time at Connectathon
> this years testing with kernel that were fully enabled. Not one problem
> WRT regression testing. Plus there is no better way to expose regression 
> problems (early on) than to enable the code.. IMHO... 
> 
> Second, its my understanding that clients have to explicitly  ask
> for 4.1 support. Are there any client out there that default to 
> 4.1 support? I would think not since there is only one client out there
> that defaults to V4.0. If there is a client that defaults to 4.1, then we 
> will a knob to turn that support off.

We can't really control what all (especially, all future) clients will
do.  Only the server can know how mature its own code is, so it's the
server's default that matters here.

There's also a greater risk of security problems with younger code, and
a vulnerable 4.1 server can be exploited without the need for a 4.1
client.

If rpc.nfsd itself is going to default to turning this on, then at the
very least I would strongly advise distributions for now to ship an
/etc/syconfig/ or /etc/defaults/ file that defaults to turning it off.

> Finally, with rpc.nfsd the precedence has already been set that we disable
> versions, not enabled them. For us to start enabling versions with rpc.nfsd
> we would have to come up with another command line flags (similar to Benny's
> patch). This means we would have one flag that enables versions and 
> another that would disable them... talk about confusion... and that's just
> not right... IMHO.... 

I agree that it's annoying to need another flag, but if that's the only
way to keep 4.1 defaulted off then on balance it's the right thing to
do.  I think it'd be OK to change the default later and remove the need
for the flag.  So only early adopters will need to know about it.

--b.

> So I say, lets stay with the precedence that was already set, enable
> the support by default but have a way to disable it...