RDMA/cm: Use DLID from inbound/outbound PathRecords as the datapath DLID

In inter-subnet cases, when inbound/outbound PRs are available,
outbound_PR.dlid is used as the requestor's datapath DLID and
inbound_PR.dlid is used as the responder's DLID. The inbound_PR.dlid
is passed to responder side with the "ConnectReq.Primary_Local_Port_LID"
field. With this solution the PERMISSIVE_LID is no longer used in
Primary Local LID field.

Signed-off-by: Mark Zhang <[email protected]>
Reviewed-by: Mark Bloch <[email protected]>
Link: https://lore.kernel.org/r/b3f6cac685bce9dde37c610be82e2c19d9e51d9e.1662631201.git.leonro@nvidia.com
Signed-off-by: Leon Romanovsky <[email protected]>

drivers/infiniband/core/cm.c

@@ -175,6 +175,7 @@ struct cm_device {
 struct cm_av {
 	struct cm_port *port;
 	struct rdma_ah_attr ah_attr;
+	u16 dlid_datapath;
 	u16 pkey_index;
 	u8 timeout;
 };
@@ -1304,6 +1305,7 @@ static void cm_format_req(struct cm_req_msg *req_msg,
 	struct sa_path_rec *pri_path = param->primary_path;
 	struct sa_path_rec *alt_path = param->alternate_path;
 	bool pri_ext = false;
+	__be16 lid;
 
 	if (pri_path->rec_type == SA_PATH_REC_TYPE_OPA)
 		pri_ext = opa_is_extended_lid(pri_path->opa.dlid,
@@ -1363,9 +1365,16 @@ static void cm_format_req(struct cm_req_msg *req_msg,
 					      htons(ntohl(sa_path_get_dlid(
 						      pri_path)))));
 	} else {
+
+		if (param->primary_path_inbound) {
+			lid = param->primary_path_inbound->ib.dlid;
+			IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg,
+				be16_to_cpu(lid));
+		} else
+			IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg,
+				be16_to_cpu(IB_LID_PERMISSIVE));
+
 		/* Work-around until there's a way to obtain remote LID info */
-		IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg,
-			be16_to_cpu(IB_LID_PERMISSIVE));
 		IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg,
 			be16_to_cpu(IB_LID_PERMISSIVE));
 	}
@@ -1520,6 +1529,10 @@ int ib_send_cm_req(struct ib_cm_id *cm_id,
 	spin_lock_irqsave(&cm_id_priv->lock, flags);
 
 	cm_move_av_from_path(&cm_id_priv->av, &av);
+	if (param->primary_path_outbound)
+		cm_id_priv->av.dlid_datapath =
+			be16_to_cpu(param->primary_path_outbound->ib.dlid);
+
 	if (param->alternate_path)
 		cm_move_av_from_path(&cm_id_priv->alt_av, &alt_av);
 
@@ -2154,6 +2167,10 @@ static int cm_req_handler(struct cm_work *work)
 				       NULL, 0);
 		goto rejected;
 	}
+	if (cm_id_priv->av.ah_attr.type == RDMA_AH_ATTR_TYPE_IB)
+		cm_id_priv->av.dlid_datapath =
+			IBA_GET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg);
+
 	if (cm_req_has_alt_path(req_msg)) {
 		ret = cm_init_av_by_path(&work->path[1], NULL,
 					 &cm_id_priv->alt_av);
@@ -4113,6 +4130,10 @@ static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
 		*qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
 				IB_QP_DEST_QPN | IB_QP_RQ_PSN;
 		qp_attr->ah_attr = cm_id_priv->av.ah_attr;
+		if ((qp_attr->ah_attr.type == RDMA_AH_ATTR_TYPE_IB) &&
+		    cm_id_priv->av.dlid_datapath &&
+		    (cm_id_priv->av.dlid_datapath != 0xffff))
+			qp_attr->ah_attr.ib.dlid = cm_id_priv->av.dlid_datapath;
 		qp_attr->path_mtu = cm_id_priv->path_mtu;
 		qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
 		qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);

drivers/infiniband/core/cma.c

@@ -4313,6 +4313,8 @@ static int cma_connect_ib(struct rdma_id_private *id_priv,
 	}
 
 	req.primary_path = &route->path_rec[0];
+	req.primary_path_inbound = route->path_rec_inbound;
+	req.primary_path_outbound = route->path_rec_outbound;
 	if (route->num_pri_alt_paths == 2)
 		req.alternate_path = &route->path_rec[1];
 

include/rdma/ib_cm.h

@@ -348,6 +348,8 @@ struct ib_cm_id *ib_cm_insert_listen(struct ib_device *device,
 
 struct ib_cm_req_param {
 	struct sa_path_rec	*primary_path;
+	struct sa_path_rec	*primary_path_inbound;
+	struct sa_path_rec	*primary_path_outbound;
 	struct sa_path_rec	*alternate_path;
 	const struct ib_gid_attr *ppath_sgid_attr;
 	__be64			service_id;