xprtrdma: Move credit update to RPC reply handler

Reduce work in the receive CQ handler, which can be run at hardware
interrupt level, by moving the RPC/RDMA credit update logic to the
RPC reply handler.

This has some additional benefits: More header sanity checking is
done before trusting the incoming credit value, and the receive CQ
handler no longer touches the RPC/RDMA header (the CPU stalls while
waiting for the header contents to be brought into the cache).

This further extends work begun by commit e7ce710a8802 ("xprtrdma:
Avoid deadlock when credit window is reset").

Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
diff --git a/net/sunrpc/xprtrdma/rpc_rdma.c b/net/sunrpc/xprtrdma/rpc_rdma.c
index dcf5ebc..d731010 100644
--- a/net/sunrpc/xprtrdma/rpc_rdma.c
+++ b/net/sunrpc/xprtrdma/rpc_rdma.c
@@ -736,7 +736,7 @@
 	struct rpc_xprt *xprt = rep->rr_xprt;
 	struct rpcrdma_xprt *r_xprt = rpcx_to_rdmax(xprt);
 	__be32 *iptr;
-	int rdmalen, status;
+	int credits, rdmalen, status;
 	unsigned long cwnd;
 
 	/* Check status. If bad, signal disconnect and return rep to pool */
@@ -871,8 +871,14 @@
 		break;
 	}
 
+	credits = be32_to_cpu(headerp->rm_credit);
+	if (credits == 0)
+		credits = 1;	/* don't deadlock */
+	else if (credits > r_xprt->rx_buf.rb_max_requests)
+		credits = r_xprt->rx_buf.rb_max_requests;
+
 	cwnd = xprt->cwnd;
-	xprt->cwnd = atomic_read(&r_xprt->rx_buf.rb_credits) << RPC_CWNDSHIFT;
+	xprt->cwnd = credits << RPC_CWNDSHIFT;
 	if (xprt->cwnd > cwnd)
 		xprt_release_rqst_cong(rqst->rq_task);