Re: [PATCH 1/2] nfsd: use new wake_up_var interfaces.

[Jeff Layton <jlayton@kernel.org>]

On Fri, 2024-12-06 at 17:10 +1100, NeilBrown wrote:
> On Fri, 06 Dec 2024, Jeff Layton wrote:
> > On Fri, 2024-12-06 at 13:55 +1100, NeilBrown wrote:
> > > The wake_up_var interface is fragile as barriers are sometimes needed.
> > > There are now new interfaces so that most wake-ups can use an interface
> > > that is guaranteed to have all barriers needed.
> > > 
> > > This patch changes the wake up on cl_cb_inflight to use
> > > atomic_dec_and_wake_up().
> > > 
> > > It also changes the wake up on rp_locked to use store_release_wake_up().
> > > This involves changing rp_locked from atomic_t to int.
> > > 
> > > Signed-off-by: NeilBrown <neilb@suse.de>
> > > ---
> > >  fs/nfsd/nfs4callback.c |  3 +--
> > >  fs/nfsd/nfs4state.c    | 16 ++++++----------
> > >  fs/nfsd/state.h        |  2 +-
> > >  3 files changed, 8 insertions(+), 13 deletions(-)
> > > 
> > > diff --git a/fs/nfsd/nfs4callback.c b/fs/nfsd/nfs4callback.c
> > > index 3877b53e429f..a8dc9de2f7fb 100644
> > > --- a/fs/nfsd/nfs4callback.c
> > > +++ b/fs/nfsd/nfs4callback.c
> > > @@ -1036,8 +1036,7 @@ static void nfsd41_cb_inflight_begin(struct nfs4_client *clp)
> > >  static void nfsd41_cb_inflight_end(struct nfs4_client *clp)
> > >  {
> > >  
> > > -	if (atomic_dec_and_test(&clp->cl_cb_inflight))
> > > -		wake_up_var(&clp->cl_cb_inflight);
> > > +	atomic_dec_and_wake_up(&clp->cl_cb_inflight);
> > >  }
> > >  
> > >  static void nfsd41_cb_inflight_wait_complete(struct nfs4_client *clp)
> > > diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
> > > index 741b9449f727..9fbf7c8f0a3e 100644
> > > --- a/fs/nfsd/nfs4state.c
> > > +++ b/fs/nfsd/nfs4state.c
> > > @@ -4739,7 +4739,7 @@ static void init_nfs4_replay(struct nfs4_replay *rp)
> > >  	rp->rp_status = nfserr_serverfault;
> > >  	rp->rp_buflen = 0;
> > >  	rp->rp_buf = rp->rp_ibuf;
> > > -	atomic_set(&rp->rp_locked, RP_UNLOCKED);
> > > +	rp->rp_locked = RP_UNLOCKED;
> > >  }
> > >  
> > >  static int nfsd4_cstate_assign_replay(struct nfsd4_compound_state *cstate,
> > > @@ -4747,9 +4747,9 @@ static int nfsd4_cstate_assign_replay(struct nfsd4_compound_state *cstate,
> > >  {
> > >  	if (!nfsd4_has_session(cstate)) {
> > >  		wait_var_event(&so->so_replay.rp_locked,
> > > -			       atomic_cmpxchg(&so->so_replay.rp_locked,
> > > -					      RP_UNLOCKED, RP_LOCKED) != RP_LOCKED);
> > > -		if (atomic_read(&so->so_replay.rp_locked) == RP_UNHASHED)
> > > +			       cmpxchg(&so->so_replay.rp_locked,
> > > +				       RP_UNLOCKED, RP_LOCKED) != RP_LOCKED);
> > 
> > nit: try_cmpxchg() generates more efficient assembly. Can we switch to
> > that here too?
> 
> Does it?  try_cmpxchg() makes loops smaller (as described in
> atomic_t.txt).  I think it wins when the "old" value has to be updated
> each time around the loop.  In this case the "old" value is always the
> same.
> 
> 

In most cases, it does, because we have to return "old" in the case of
the traditional cmpxchg() operation. From atomic_t.txt:

  int atomic_cmpxchg(atomic_t *ptr, int old, int new)
  {
    (void)atomic_try_cmpxchg(ptr, &old, new);
    return old;
  }

That said, in this case it my not be a win. You need something to
return that value anyway, so it can properly act as a wait_var_event()
condition.

> > 
> > > +		if (so->so_replay.rp_locked == RP_UNHASHED)
> > >  			return -EAGAIN;
> > >  		cstate->replay_owner = nfs4_get_stateowner(so);
> > >  	}
> > > @@ -4762,9 +4762,7 @@ void nfsd4_cstate_clear_replay(struct nfsd4_compound_state *cstate)
> > >  
> > >  	if (so != NULL) {
> > >  		cstate->replay_owner = NULL;
> > > -		atomic_set(&so->so_replay.rp_locked, RP_UNLOCKED);
> > > -		smp_mb__after_atomic();
> > > -		wake_up_var(&so->so_replay.rp_locked);
> > > +		store_release_wake_up(&so->so_replay.rp_locked, RP_UNLOCKED);
> > >  		nfs4_put_stateowner(so);
> > >  	}
> > >  }
> > > @@ -5069,9 +5067,7 @@ move_to_close_lru(struct nfs4_ol_stateid *s, struct net *net)
> > >  	 * Some threads with a reference might be waiting for rp_locked,
> > >  	 * so tell them to stop waiting.
> > >  	 */
> > > -	atomic_set(&oo->oo_owner.so_replay.rp_locked, RP_UNHASHED);
> > > -	smp_mb__after_atomic();
> > > -	wake_up_var(&oo->oo_owner.so_replay.rp_locked);
> > > +	store_release_wake_up(&oo->oo_owner.so_replay.rp_locked, RP_UNHASHED);
> > >  	wait_event(close_wq, refcount_read(&s->st_stid.sc_count) == 2);
> > >  
> > >  	release_all_access(s);
> > > diff --git a/fs/nfsd/state.h b/fs/nfsd/state.h
> > > index e16bb3717fb9..ba30b2335b66 100644
> > > --- a/fs/nfsd/state.h
> > > +++ b/fs/nfsd/state.h
> > > @@ -505,7 +505,7 @@ struct nfs4_replay {
> > >  	unsigned int		rp_buflen;
> > >  	char			*rp_buf;
> > >  	struct knfsd_fh		rp_openfh;
> > > -	atomic_t		rp_locked;
> > > +	int			rp_locked;
> > >  	char			rp_ibuf[NFSD4_REPLAY_ISIZE];
> > >  };
> > >  
> > 
> > Looks good otherwise.
> > 
> > Reviewed-by: Jeff Layton <jlayton@kernel.org>
> > 
> 

-- 
Jeff Layton <jlayton@kernel.org>