Re: [PATCH v4 06/30] xprtrdma: Don't wake pending tasks until disconnect is done

[Trond Myklebust <trondmy@hammerspace.com>]

On Mon, 2018-12-17 at 14:00 -0500, Chuck Lever wrote:
> > On Dec 17, 2018, at 1:55 PM, Trond Myklebust <
> > trondmy@hammerspace.com> wrote:
> > 
> > On Mon, 2018-12-17 at 13:37 -0500, Chuck Lever wrote:
> > > > On Dec 17, 2018, at 12:28 PM, Trond Myklebust <
> > > > trondmy@hammerspace.com> wrote:
> > > > 
> > > > On Mon, 2018-12-17 at 11:39 -0500, Chuck Lever wrote:
> > > > > Transport disconnect processing does a "wake pending tasks"
> > > > > at
> > > > > various points.
> > > > > 
> > > > > Suppose an RPC Reply is being processed. The RPC task that
> > > > > Reply
> > > > > goes with is waiting on the pending queue. If a disconnect
> > > > > wake-
> > > > > up
> > > > > happens before reply processing is done, that reply, even if
> > > > > it
> > > > > is
> > > > > good, is thrown away, and the RPC has to be sent again.
> > > > > 
> > > > > This window apparently does not exist for socket transports
> > > > > because
> > > > > there is a lock held while a reply is being received which
> > > > > prevents
> > > > > the wake-up call until after reply processing is done.
> > > > > 
> > > > > To resolve this, all RPC replies being processed on an RPC-
> > > > > over-
> > > > > RDMA
> > > > > transport have to complete before pending tasks are awoken
> > > > > due to
> > > > > a
> > > > > transport disconnect.
> > > > > 
> > > > > Callers that already hold the transport write lock may invoke
> > > > > ->ops->close directly. Others use a generic helper that
> > > > > schedules
> > > > > a close when the write lock can be taken safely.
> > > > > 
> > > > > Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
> > > > > ---
> > > > > include/linux/sunrpc/xprt.h                |    1 +
> > > > > net/sunrpc/xprt.c                          |   19
> > > > > +++++++++++++++++++
> > > > > net/sunrpc/xprtrdma/backchannel.c          |   13 +++++++--
> > > > > ----
> > > > > net/sunrpc/xprtrdma/svc_rdma_backchannel.c |    8 +++++---
> > > > > net/sunrpc/xprtrdma/transport.c            |   16 ++++++++++-
> > > > > ----
> > > > > -
> > > > > net/sunrpc/xprtrdma/verbs.c                |    5 ++---
> > > > > 6 files changed, 44 insertions(+), 18 deletions(-)
> > > > > 
> > > > > diff --git a/include/linux/sunrpc/xprt.h
> > > > > b/include/linux/sunrpc/xprt.h
> > > > > index a4ab4f8..ee94ed0 100644
> > > > > --- a/include/linux/sunrpc/xprt.h
> > > > > +++ b/include/linux/sunrpc/xprt.h
> > > > > @@ -401,6 +401,7 @@ static inline __be32
> > > > > *xprt_skip_transport_header(struct rpc_xprt *xprt, __be32 *
> > > > > bool			xprt_request_get_cong(struct rpc_xprt
> > > > > *xprt,
> > > > > struct rpc_rqst *req);
> > > > > void			xprt_disconnect_done(struct rpc_xprt
> > > > > *xprt);
> > > > > void			xprt_force_disconnect(struct rpc_xprt
> > > > > *xprt);
> > > > > +void			xprt_disconnect_nowake(struct rpc_xprt
> > > > > *xprt);
> > > > > void			xprt_conditional_disconnect(struct
> > > > > rpc_xprt
> > > > > *xprt, unsigned int cookie);
> > > > > 
> > > > > bool			xprt_lock_connect(struct rpc_xprt *,
> > > > > struct
> > > > > rpc_task *, void *);
> > > > > diff --git a/net/sunrpc/xprt.c b/net/sunrpc/xprt.c
> > > > > index ce92700..afe412e 100644
> > > > > --- a/net/sunrpc/xprt.c
> > > > > +++ b/net/sunrpc/xprt.c
> > > > > @@ -685,6 +685,25 @@ void xprt_force_disconnect(struct
> > > > > rpc_xprt
> > > > > *xprt)
> > > > > }
> > > > > EXPORT_SYMBOL_GPL(xprt_force_disconnect);
> > > > > 
> > > > > +/**
> > > > > + * xprt_disconnect_nowake - force a call to xprt->ops->close
> > > > > + * @xprt: transport to disconnect
> > > > > + *
> > > > > + * The caller must ensure that xprt_wake_pending_tasks() is
> > > > > + * called later.
> > > > > + */
> > > > > +void xprt_disconnect_nowake(struct rpc_xprt *xprt)
> > > > > +{
> > > > > +       /* Don't race with the test_bit() in
> > > > > xprt_clear_locked()
> > > > > */
> > > > > +       spin_lock_bh(&xprt->transport_lock);
> > > > > +       set_bit(XPRT_CLOSE_WAIT, &xprt->state);
> > > > > +       /* Try to schedule an autoclose RPC call */
> > > > > +       if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
> > > > > +               queue_work(xprtiod_workqueue, &xprt-
> > > > > > task_cleanup);
> > > > > +       spin_unlock_bh(&xprt->transport_lock);
> > > > > +}
> > > > > +EXPORT_SYMBOL_GPL(xprt_disconnect_nowake);
> > > > > +
> > > > 
> > > > We shouldn't need both xprt_disconnect_nowake() and
> > > > xprt_force_disconnect() to be exported given that you can build
> > > > the
> > > > latter from the former + xprt_wake_pending_tasks() (which is
> > > > also
> > > > already exported).
> > > 
> > > Thanks for your review!
> > > 
> > > I can get rid of xprt_disconnect_nowake. There are some
> > > variations,
> > > depending on why wake_pending_tasks is protected by xprt-
> > > > transport_lock.
> > 
> > I'm having some second thoughts about the patch that Scott sent out
> > last week to fix the issue that Dave and he were seeing. I think
> > that
> > what we really need to do to fix his issue is to call
> > xprt_disconnect_done() after we've released the TCP socket.
> > 
> > Given that realisation, I think that we can fix up
> > xprt_force_disconnect() to only wake up the task that holds the
> > XPRT_LOCKED instead of doing a thundering herd wakeup like we do
> > today.
> > That should (I think) obviate the need for a separate
> > xprt_disconnect_nowake().
> 
> For RPC-over-RDMA, there really is a dangerous race between the
> waking
> task(s) and work being done by the deferred RPC completion handler.
> IMO
> the only safe thing RPC-over-RDMA can do is not wake anything until
> the
> deferred queue is well and truly drained.

The deferred RPC completion handler (and hence the close) cannot
execute if another task is holding XPRT_LOCKED, so we do need to wake
up that task (and only that one).

Note that in the new code, the only reason why a task would be holding
XPRT_LOCKED while sleeping is because

   1. It is waiting for a connection attempt to complete following a call
      to xprt_connect().
   2. It is waiting for a write_space event following an attempt to
      transmit.



> As you observed when we last spoke, socket transports are already
> protected from this race by the socket lock.... RPC-over-RDMA is
> going
> to have to be more careful.
> 
> 
> > A patch is forthcoming later today. I'll make sure you are Cced so
> > you
> > can comment.
> > 
> > -- 
> > Trond Myklebust
> > Linux NFS client maintainer, Hammerspace
> > trond.myklebust@hammerspace.com
> 
> --
> Chuck Lever
> 
> 
> 
-- 
Trond Myklebust
CTO, Hammerspace Inc
4300 El Camino Real, Suite 105
Los Altos, CA 94022
www.hammer.space