Re: [PATCH] nfsd: don't break lease while servicing a COMMIT call (try #2)

[Jeff Layton <jlayton@redhat.com>]

On Thu, 25 Mar 2010 18:07:48 -0400
"J. Bruce Fields" <bfields@fieldses.org> wrote:

> On Thu, Mar 25, 2010 at 02:16:28PM -0400, Trond Myklebust wrote:
> > On Thu, 2010-03-25 at 13:47 -0400, J. Bruce Fields wrote: 
> > > On Mon, Mar 22, 2010 at 04:33:40PM -0400, Jeff Layton wrote:
> > > > It looks like nfs_inode_return_delegation always calls nfs_msync_inode
> > > > on any valid delegation before returning it, regardless of the
> > > > delegation type.
> > > > 
> > > > RFC 3530 says this:
> > > > 
> > > >    If the client is granted a read delegation, it is assured that no
> > > >    other client has the ability to write to the file for the duration of
> > > >    the delegation.  If the client is granted a write delegation, the
> > > >    client is assured that no other client has read or write access to
> > > >    the file.
> > > > 
> > > > That doesn't seem to imply that we must flush writes before returning
> > > > either type of delegation. OTOH, maybe it makes sense to treat those as
> > > > cache consistency points since a delegreturn sort of implies that
> > > > another client wants to use the file.
> > > > 
> > > > I'm not quite sure how to interpret the spec here...
> > > 
> > > If there's that call could cause the client to wait for an actual write
> > > to succeed before returning the delegation, then something's wrong.
> > 
> > We're certainly expected to write back data before returning a write
> > delegation (see Section 9.4.4 of RFC 3530).
> > 
> > For the case of a read delegation, then the spec is silent because it
> > contains no discussion of the case where a server grants both an open
> > for write and a read delegation. If you want a normative statement on
> > what clients should do for that case, then I suggest a discussion on the
> > IETF list with a view to getting it into RFC3530-bis.
> 
> Yeah, that would be a good idea to get nailed down at some point.
> 
> (But the current server implementation doesn't allow write opens in this
> situation.  So I wonder why we're seeing any commit from the client at
> all?)
> 
> --b.

This problem was reported against a RHEL5 client and server. It's
possible that there's another bug that's allowing that somehow, but
I'll need to look over the capture file again to see if I can find it.

-- 
Jeff Layton <jlayton@redhat.com>