NFS oops on a 2.6.24.4 kernel server

[Quentin Godfroy]

On Wed, Apr 23, 2008 at 12:02:06PM -0400, J. Bruce Fields wrote:
> On Wed, Apr 23, 2008 at 05:37:46PM +0200, Quentin Godfroy wrote:
> > On Wed, Apr 23, 2008 at 05:24:01PM +0200, Quentin Godfroy wrote:
> > > On Wed, Apr 23, 2008 at 01:01:19PM +0200, Quentin Godfroy wrote:
> > [snip] 
> > > Is there some kind of race condition in
> > > fs/nfsd/nfs4callback.c:do_probe_callback? I see sometimes two nfs4_cb_probe
> > > for only one client. Would they possibly share the same struct nfs4_callback
> > > cb?
> > 
> > commit 63c86716ea34ad94d52e5b0abbda152574dc42b5
> > nfsd: move callback rpc_client creation into separate thread
> > and 
> > commit 46f8a64bae11f5c9b15b4401f6e9863281999b66
> > nfsd4: probe callback channel only once
> > 
> > should correct the problem. Sorry for the mess.
> 
> I was just about to recommend trying those....  If you manage to
> confirm that they do fix the problem you're seeing, that'd be useful
> information.  Thanks!

On a retropatched 2.6.24.5, it *looks* like the problem went away.

However, if the tcp connection is destroyed for lack of network activity,
a new probe can be done on reopening of the connection (I suppose this means
that a client disappears for the server if the connection is torn out).