Re: Fw: Deadlock regression in v2.6.31.6

[Trond Myklebust <trond.myklebust@fys.uio.no>]

On Fri, 2009-11-27 at 01:14 +0100, Stephen R. van den Berg wrote: 
> On Fri, Nov 27, 2009 at 01:07, Stephen R. van den Berg <srb-PCMv+cxZuL0@public.gmane.org> wrote:
> > RPC:       worker connecting xprt cfa94400 to address: addr=1.2.3.151
> > port=2049 proto=tcp
> > RPC:       cfa94400 connect status 99 connected 0 sock state 7
> 
> errno 99 means EADDRNOTAVAIL.  In userspace this normally is solved by
> using the REUSEADDR sockopt.  In xprtsock.c we try something like:
> 
>                 /* We're probably in TIME_WAIT. Get rid of existing socket,
>                  * and retry
>                  */
>                 set_bit(XPRT_CONNECTION_CLOSE, &xprt->state);
>                 xprt_force_disconnect(xprt);
> 
> I'd guess that this needs to be fixed, or the REUSEADDR sockopt needs to be set.

Does the following patch fix matters?

Trond

--------------------------------------------------------------------------------------------------------- 
SUNRPC: Ensure that we honour autoclose before attempting to reconnect

From: Trond Myklebust <Trond.Myklebust@netapp.com>

If the XPRT_CLOSE_WAIT flag is set, we need to ensure that we call
xprt->ops->close() while holding xprt_lock_write() before we can
start reconnecting.

Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
---

 net/sunrpc/xprt.c |    4 ++++
 1 files changed, 4 insertions(+), 0 deletions(-)


diff --git a/net/sunrpc/xprt.c b/net/sunrpc/xprt.c
index fd46d42..469de29 100644
--- a/net/sunrpc/xprt.c
+++ b/net/sunrpc/xprt.c
@@ -700,6 +700,10 @@ void xprt_connect(struct rpc_task *task)
 	}
 	if (!xprt_lock_write(xprt, task))
 		return;
+
+	if (test_and_clear_bit(XPRT_CLOSE_WAIT, &xprt->state))
+		xprt->ops->close(xprt);
+
 	if (xprt_connected(xprt))
 		xprt_release_write(xprt, task);
 	else {