xprtrdma: Serialize credit accounting again

Commit fe97b47c ("xprtrdma: Use workqueue to process RPC/RDMA
replies") replaced the reply tasklet with a workqueue that allows
RPC replies to be processed in parallel. Thus the credit values in
RPC-over-RDMA replies can be applied in a different order than in
which the server sent them.

To fix this, revert commit eba8ff66 ("xprtrdma: Move credit
update to RPC reply handler"). Reverting is done by hand to
accommodate code changes that have occurred since then.

Fixes: fe97b47c ("xprtrdma: Use workqueue to process . . .")
Signed-off-by: Chuck Lever <chuck.lever@oracle.com>
Reviewed-by: Sagi Grimberg <sagig@mellanox.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>

net/sunrpc/xprtrdma/rpc_rdma.c

@@ -797,7 +797,6 @@ rpcrdma_reply_handler(struct rpcrdma_rep *rep)
 	__be32 *iptr;
 	int rdmalen, status, rmerr;
 	unsigned long cwnd;
-	u32 credits;
 
 	dprintk("RPC:       %s: incoming rep %p\n", __func__, rep);
 
@@ -928,15 +927,9 @@ rpcrdma_reply_handler(struct rpcrdma_rep *rep)
 	if (req->rl_nchunks)
 		r_xprt->rx_ia.ri_ops->ro_unmap_sync(r_xprt, req);
 
-	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;
-
 	spin_lock_bh(&xprt->transport_lock);
 	cwnd = xprt->cwnd;
-	xprt->cwnd = credits << RPC_CWNDSHIFT;
+	xprt->cwnd = atomic_read(&r_xprt->rx_buf.rb_credits) << RPC_CWNDSHIFT;
 	if (xprt->cwnd > cwnd)
 		xprt_release_rqst_cong(rqst->rq_task);
 

net/sunrpc/xprtrdma/verbs.c

@@ -190,6 +190,28 @@ rpcrdma_receive_worker(struct work_struct *work)
 	rpcrdma_reply_handler(rep);
 }
 
+/* Perform basic sanity checking to avoid using garbage
+ * to update the credit grant value.
+ */
+static void
+rpcrdma_update_granted_credits(struct rpcrdma_rep *rep)
+{
+	struct rpcrdma_msg *rmsgp = rdmab_to_msg(rep->rr_rdmabuf);
+	struct rpcrdma_buffer *buffer = &rep->rr_rxprt->rx_buf;
+	u32 credits;
+
+	if (rep->rr_len < RPCRDMA_HDRLEN_ERR)
+		return;
+
+	credits = be32_to_cpu(rmsgp->rm_credit);
+	if (credits == 0)
+		credits = 1;	/* don't deadlock */
+	else if (credits > buffer->rb_max_requests)
+		credits = buffer->rb_max_requests;
+
+	atomic_set(&buffer->rb_credits, credits);
+}
+
 static void
 rpcrdma_recvcq_process_wc(struct ib_wc *wc)
 {
@@ -211,7 +233,8 @@ rpcrdma_recvcq_process_wc(struct ib_wc *wc)
 	ib_dma_sync_single_for_cpu(rep->rr_device,
 				   rdmab_addr(rep->rr_rdmabuf),
 				   rep->rr_len, DMA_FROM_DEVICE);
-	prefetch(rdmab_to_msg(rep->rr_rdmabuf));
+
+	rpcrdma_update_granted_credits(rep);
 
 out_schedule:
 	queue_work(rpcrdma_receive_wq, &rep->rr_work);
@@ -330,6 +353,7 @@ rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
 connected:
 		dprintk("RPC:       %s: %sconnected\n",
 					__func__, connstate > 0 ? "" : "dis");
+		atomic_set(&xprt->rx_buf.rb_credits, 1);
 		ep->rep_connected = connstate;
 		rpcrdma_conn_func(ep);
 		wake_up_all(&ep->rep_connect_wait);
@@ -943,6 +967,7 @@ rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
 	buf->rb_max_requests = r_xprt->rx_data.max_requests;
 	buf->rb_bc_srv_max_requests = 0;
 	spin_lock_init(&buf->rb_lock);
+	atomic_set(&buf->rb_credits, 1);
 
 	rc = ia->ri_ops->ro_init(r_xprt);
 	if (rc)

net/sunrpc/xprtrdma/xprt_rdma.h

@@ -311,6 +311,7 @@ struct rpcrdma_buffer {
 	struct list_head	rb_send_bufs;
 	struct list_head	rb_recv_bufs;
 	u32			rb_max_requests;
+	atomic_t		rb_credits;	/* most recent credit grant */
 
 	u32			rb_bc_srv_max_requests;
 	spinlock_t		rb_reqslock;	/* protect rb_allreqs */