Re: [PATCH] xfs: reinstate the old i_version counter as STATX_CHANGE_COOKIE

[Jeff Layton <jlayton@kernel.org>]

On Wed, 2023-10-11 at 08:49 -0700, Darrick J. Wong wrote:
> On Wed, Oct 11, 2023 at 09:09:38AM -0400, Jeff Layton wrote:
> > On Fri, 2023-09-29 at 14:43 -0400, Jeff Layton wrote:
> > > The handling of STATX_CHANGE_COOKIE was moved into generic_fillattr in
> > > commit 0d72b92883c6 (fs: pass the request_mask to generic_fillattr), but
> > > we didn't account for the fact that xfs doesn't call generic_fillattr at
> > > all.
> > > 
> > > Make XFS report its i_version as the STATX_CHANGE_COOKIE.
> > > 
> > > Fixes: 0d72b92883c6 (fs: pass the request_mask to generic_fillattr)
> > > Signed-off-by: Jeff Layton <jlayton@kernel.org>
> > > ---
> > > I had hoped to fix this in a better way with the multigrain patches, but
> > > it's taking longer than expected (if it even pans out at this point).
> > > 
> > > Until then, make sure we use XFS's i_version as the STATX_CHANGE_COOKIE,
> > > even if it's bumped due to atime updates. Too many invalidations is
> > > preferable to not enough.
> > > ---
> > >  fs/xfs/xfs_iops.c | 5 +++++
> > >  1 file changed, 5 insertions(+)
> > > 
> > > diff --git a/fs/xfs/xfs_iops.c b/fs/xfs/xfs_iops.c
> > > index 1c1e6171209d..2b3b05c28e9e 100644
> > > --- a/fs/xfs/xfs_iops.c
> > > +++ b/fs/xfs/xfs_iops.c
> > > @@ -584,6 +584,11 @@ xfs_vn_getattr(
> > >  		}
> > >  	}
> > >  
> > > +	if ((request_mask & STATX_CHANGE_COOKIE) && IS_I_VERSION(inode)) {
> > > +		stat->change_cookie = inode_query_iversion(inode);
> > > +		stat->result_mask |= STATX_CHANGE_COOKIE;
> > > +	}
> > > +
> > >  	/*
> > >  	 * Note: If you add another clause to set an attribute flag, please
> > >  	 * update attributes_mask below.
> > > 
> > > ---
> > > base-commit: df964ce9ef9fea10cf131bf6bad8658fde7956f6
> > > change-id: 20230929-xfs-iversion-819fa2c18591
> > > 
> > > Best regards,
> > 
> > Ping?
> > 
> > This patch is needed in v6.6 to prevent a regression when serving XFS
> > via NFSD. I'd prefer this go in via the xfs tree, but let me know if
> > you need me to get this merged this via a different one.
> 
> Oh!   Right, this is needed because the "hide a state in the high bit of
> tv_nsec" stuff got reverted in -rc3, correct?  So now nfsd needs some
> way to know that something changed in the file, and better to have too
> many client invalidations than not enough?  And I guess bumping
> i_version will keep nfsd sane for now?
> 
> If the answers are [yes, yes, yes] then:
> Reviewed-by: Darrick J. Wong <djwong@kernel.org>

Yes, yes, and yes. Can you guys shepherd this into mainline?

Thanks for the R-b!
-- 
Jeff Layton <jlayton@kernel.org>