Re: [PATCH v2 8/9] nfsd4: keep a reference count on client while in use

["J. Bruce Fields" <bfields@citi.umich.edu>]

On Wed, May 12, 2010 at 09:19:34AM +0300, Benny Halevy wrote:
> On May. 12, 2010, 7:26 +0300, Benny Halevy <bhalevy@panasas.com> wrote:
> > On May. 12, 2010, 5:40 +0300, " J. Bruce Fields" <bfields@citi.umich.edu> wrote:
> >> Something still doesn't look quite right: a sequence operation
> >> increments cl_count from 1 to 2; then, say, an exchangeid in another
> >> thread expires the client, dropping cl_count from 2 to 1; then the
> >> laundromat runs, sees cl_count 1, and decides it can expire the
> >> client.  Meanwhile, the original compound is still running.
> >>
> >> Am I missing something?
> > 
> > Yeah.  exchange_id doesn't touch cl_refcount.  It may call expire_client
> > explicitly which will unhash the client but will not destroy it if cl_refcount > 0
> > the laundromat won't the client either after that since it's not on the lru list
> > any more (and even if it would, it's refcount is still great than zero so it would
> > have been ignored)
> 
> Sorry, I misspoken.  exchange_id may decrement cl_refcount via expire_client()
> but still the laundromat won't see it as expire_client will have already removed the
> client from client_lru.

So the only potential problems are with operations that reach the client
through the pointer in the session, so such operations may need to do
something special.  OK.

--b.