Re: [PATCH v2] nfsd: Fix races between nfsd4_cb_release() and nfsd4_shutdown_callback()

[bfields@fieldses.org (J. Bruce Fields)]

On Fri, Oct 25, 2019 at 11:21:19AM -0400, J. Bruce Fields wrote:
> On Fri, Oct 25, 2019 at 02:55:45PM +0000, Trond Myklebust wrote:
> > On Fri, 2019-10-25 at 10:51 -0400, J. Bruce Fields wrote:
> > > On Wed, Oct 23, 2019 at 05:43:18PM -0400, Trond Myklebust wrote:
> > > > When we're destroying the client lease, and we call
> > > > nfsd4_shutdown_callback(), we must ensure that we do not return
> > > > before all outstanding callbacks have terminated and have
> > > > released their payloads.
> > > 
> > > This is great, thanks!  We've seen what I'm fairly sure is the same
> > > bug
> > > from Red Hat users.  I think my blind spot was an assumption that
> > > rpc tasks wouldn't outlive rpc_shutdown_client().
> > > 
> > > However, it's causing xfstests runs to hang, and I haven't worked out
> > > why yet.
> > > 
> > > I'll spend some time on it this afternoon and let you know what I
> > > figure
> > > out.
> > > 
> > 
> > Is that happening with v2 or with v1? With v1 there is definitely a
> > hang in __destroy_client() due to the refcount leak that I believe I
> > fixed in v2.
> 
> I thought I was running v2, let me double-check....

Yes, with v2 I'm getting a hang on generic/013.

I checked quickly and didn't see anything interesting in the logs,
otherwise I haven't done any digging.

--b.