Re: [PATCH] nfsd: CB_RECALL can race with FREE_STATEID

["J. Bruce Fields" <bfields@fieldses.org>]

On Thu, Apr 18, 2019 at 04:50:24PM -0400, Scott Mayhew wrote:
> On Thu, 18 Apr 2019, J. Bruce Fields wrote:
> 
> > On Thu, Apr 18, 2019 at 09:24:00AM -0400, Scott Mayhew wrote:
> > > While trying to track down some issues involving large numbers of
> > > delegations being recalled/revoked, I caught the server setting
> > > SEQ4_STATUS_CB_PATH_DOWN while the client was actively responding to
> > > CB_RECALLs.  It turns out that the client had already done a
> > > TEST_STATEID and FREE_STATEID for a delegation being recalled by the
> > > time it received the CB_RECALL.
> > 
> > That's interesting, thanks!
> > 
> > This exception seems awfully narrow, though.
> > 
> > If we get back any NFS-level error at all, then I think the callback
> > channel is working (am I wrong?)
> 
> Correct, if the client replies with either NFS4ERR_DELAY or
> NFS4ERR_BAD_STATEID, the server will retry 1 time (see dl_retries).
> After that, we fall thru and nfsd4_cb_recall_done() returns -1 which
> causes the SEQ4_STATUS_CB_PATH_DOWN flag to be set.
> 
> > and telling the client to set up a new
> > one is probably not going to help.  The best we can do is probably just
> > give up
> 
> That's what the patch is essentially doing.  Or are you saying don't
> even bother with the checks but still return 1 so we don't set the
> SEQ4_STATUS_CB_PATH_DOWN flag?

Right, I don't see any point returning -1 (which ends up setting
CB_PATH_DOWN) in any case where we get an nfs-level error.  If the
client got so far as returning an error, then the callback path is
working.

I'm not sure exactly what errors *should* result in CB_PATH_DOWN,
though.  ETIMEDOUT, ENOTCONN, EIO?  And maybe we should be checking for
those in nfsd4_cb_done, and do away with the convention that -1 means
CB_PATH_DOWN.  I don't think there's a reason individual callback ops
would need different rules for when to mark the callback channel down.

--b.

> 
> > and let the client deal with the ensuing
> > RECALLABLE_STATE_REVOKED flag.
> 
> The client's already dealing with the RECALLABLE_STATE_REVOKED flag,
> that's why it sent a TEST_STATEID and FREE_STATEID before it got this
> particular CB_RECALL.  The idea behind the patch is to not give the
> state manager on the client additional work by setting CB_PATH_DOWN when
> the callback channel is clearly working...
> 
> -Scott
> > 
> > --b.
> > 
> > > 
> > > Signed-off-by: Scott Mayhew <smayhew@redhat.com>
> > > ---
> > >  fs/nfsd/nfs4state.c | 8 ++++++++
> > >  1 file changed, 8 insertions(+)
> > > 
> > > diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
> > > index 6a45fb00c5fc..e88e429133a8 100644
> > > --- a/fs/nfsd/nfs4state.c
> > > +++ b/fs/nfsd/nfs4state.c
> > > @@ -3958,6 +3958,14 @@ static int nfsd4_cb_recall_done(struct nfsd4_callback *cb,
> > >  			rpc_delay(task, 2 * HZ);
> > >  			return 0;
> > >  		}
> > > +		/*
> > > +		 * Race: client may have done a FREE_STATEID before
> > > +		 * receiving the CB_RECALL.
> > > +		 */
> > > +		if (dp->dl_stid.sc_type == NFS4_REVOKED_DELEG_STID &&
> > > +				refcount_read(&dp->dl_stid.sc_count) == 1 &&
> > > +				list_empty(&dp->dl_recall_lru))
> > > +			return 1;
> > >  		/*FALLTHRU*/
> > >  	default:
> > >  		return -1;
> > > -- 
> > > 2.17.2