Re: should we make "-o iversion" the default on ext4 ?

[Jeff Layton <jlayton@kernel.org>]

On Wed, 2022-07-20 at 12:29 -0400, Benjamin Coddington wrote:
> On 20 Jul 2022, at 12:15, Jeff Layton wrote:
> 
> > On Wed, 2022-07-20 at 11:56 -0400, Benjamin Coddington wrote:
> > > On 20 Jul 2022, at 10:38, Jeff Layton wrote:
> > > > On Wed, 2022-07-20 at 16:15 +0200, Lukas Czerner wrote:
> > > > > 
> > > > > Is there a different way I am not seeing?
> > > > > 
> > > > 
> > > > Right, implementing this is the difficult bit actually since this uses a
> > > > MS_* flag. If we do make this the default, we'd definitely want to
> > > > continue allowing "-o noiversion" to disable it.
> > > > 
> > > > Could we just reverse the default in libmount? It might cause this to
> > > > suddenly be enabled in some deployments, but in most cases, people
> > > > wouldn't even notice and they could still specify -o noiversion to turn
> > > > it off.
> > > > 
> > > > Another idea would be to introduce new mount options for this, but
> > > > that's kind of nasty from a UI standpoint.
> > > 
> > > Is it safe to set SB_I_VERSION at export time?  If so, export_operations
> > > could grow an ->enable_iversion().
> > > 
> > 
> > That sounds like it might be problematic.
> > 
> > Consider the case where a NFSv4 client has cached file data and the
> > change attribute for the file. Server then reboots, but before the
> > export happens a local user makes a change to the file and it doesn't
> > update the i_version.
> 
> Nfsd currently uses both ctime and i_version if its available, I'd expect
> that eliminates this case.
> 

Good point, that probably would. Still, I'd rather we just enable this
wholesale if we can get away with it. There's still some interest in
exposing i_version to userland via statx or the like, so I'd rather not
assume that only nfsd will care about it.
-- 
Jeff Layton <jlayton@kernel.org>