On Mon, 18 Feb 2013 20:15:02 +0000 Al Viro <viro@ZenIV.linux.org.uk> wrote:

> On Mon, Feb 18, 2013 at 02:46:55PM -0500, Jeff Layton wrote:
> 
> > Ok, that helps. In that case, this patch might be a reasonable
> > forward-port of the one Neil sent earlier today. Note that this doesn't
> > really do anything for the umount problem, but it does seem to fix the
> > testcase for the problem I've been looking at.
> > 
> > Thoughts?
> 
> If we really go for "in this case revalidate should be weaker", we might
> as well introduce a separate method for it.
> 
> As it is, we have several callers of ->d_revalidate(); this one (in
> complete_walk(), only for FS_REVAL_DOT filesystems) and ones in
> lookup_dcache() and lookup_fast() (in both cases we have and want the
> name to match).  There are only two fs with FS_REVAL_DOT present - nfs
> and 9p.  *IF* we want to make ->d_revalidate() on NFS behave differently
> in complete_walk() case, it would argue for just splitting the method
> in two, replacing FS_REVAL_DOT with "dentry has this method" and probably
> taking a good look at what 9p needs in the same case.

Sounds good to me.

Reminds me that we used to have an i_op->revalidate() method for revalidating
just the inode (not the dentry).  It called nfs_revalidate_inode() for NFS.

We lost it over a decade ago:

commit cc41b90f8a9ad3cd85a39dd4fcc71f965a675b0e
Author: Alexander Viro <viro@math.psu.edu>
Date:   Tue May 21 21:12:46 2002 -0700

    [PATCH] kill ->i_op->revalidate()
    
    kill ->i_op->revalidate()



:-)

(this doesn't help my umount problem though)

NeilBrown