prov/verbs: CM req header is part of private data with AF_INET

With AF_INET, the CM request header is part of the private
data and it is defined in the first 36 bytes. To comply with
the CM REQ private data format, user private data must be
stored starting at byte 36. Worse, any data saved at the
beginning of the private data buffer would be corrupted as
the first byte would be overwritten with 0.

References:
- https://www.spinics.net/lists/linux-rdma/msg22381.html
- IBTA Architecture Specification Vol 1. Annex A11: RDMA
  IP CM Service.

Signed-off-by: Sylvain Didelot <sdidelot@ddn.com>

prov/verbs/src/fi_verbs.h

@@ -114,6 +114,7 @@
 
 #define VERBS_NO_COMP_FLAG	((uint64_t)-1)
 
+#define VRB_CM_REQ_HDR_SIZE (36)
 #define VRB_CM_DATA_SIZE	(56)
 #define VERBS_CM_DATA_SIZE	(VRB_CM_DATA_SIZE -		\
 				 sizeof(struct vrb_cm_data_hdr))
@@ -594,6 +595,7 @@ struct vrb_ep {
 	size_t				rx_cq_size;
 	struct rdma_conn_param		conn_param;
 	struct vrb_cm_data_hdr		*cm_hdr;
+	void				*cm_priv_data;
 };
 
 

prov/verbs/src/verbs_cm.c

@@ -125,11 +125,10 @@ vrb_msg_ep_prepare_cm_data(const void *param, size_t param_size,
 
 static inline void
 vrb_ep_prepare_rdma_cm_param(struct rdma_conn_param *conn_param,
-				struct vrb_cm_data_hdr *cm_hdr,
-				size_t cm_hdr_data_size)
+				void *priv_data, size_t priv_data_size)
 {
-	conn_param->private_data = cm_hdr;
-	conn_param->private_data_len = (uint8_t)cm_hdr_data_size;
+	conn_param->private_data = priv_data;
+	conn_param->private_data_len = (uint8_t)priv_data_size;
 	conn_param->responder_resources = RDMA_MAX_RESP_RES;
 	conn_param->initiator_depth = RDMA_MAX_INIT_DEPTH;
 	conn_param->flow_control = 1;
@@ -142,6 +141,9 @@ vrb_msg_ep_connect(struct fid_ep *ep_fid, const void *addr,
 {
 	struct vrb_ep *ep =
 		container_of(ep_fid, struct vrb_ep, util_ep.ep_fid);
+	size_t priv_data_len;
+	struct vrb_cm_data_hdr *cm_hdr;
+	off_t cm_hdr_offset = 0;
 	int ret;
 
 	if (OFI_UNLIKELY(paramlen > VERBS_CM_DATA_SIZE))
@@ -153,13 +155,19 @@ vrb_msg_ep_connect(struct fid_ep *ep_fid, const void *addr,
 			return ret;
 	}
 
-	ep->cm_hdr = malloc(sizeof(*(ep->cm_hdr)) + paramlen);
-	if (!ep->cm_hdr)
+	if (ep->id->route.addr.src_addr.sa_family == AF_IB)
+		cm_hdr_offset = VRB_CM_REQ_HDR_SIZE;
+
+	priv_data_len = cm_hdr_offset + sizeof(*cm_hdr) + paramlen;
+	ep->cm_priv_data = malloc(priv_data_len);
+	if (!ep->cm_priv_data)
 		return -FI_ENOMEM;
 
-	vrb_msg_ep_prepare_cm_data(param, paramlen, ep->cm_hdr);
-	vrb_ep_prepare_rdma_cm_param(&ep->conn_param, ep->cm_hdr,
-					sizeof(*(ep->cm_hdr)) + paramlen);
+	cm_hdr = (void*)((char *)ep->cm_priv_data + cm_hdr_offset);
+
+	vrb_msg_ep_prepare_cm_data(param, paramlen, cm_hdr);
+	vrb_ep_prepare_rdma_cm_param(&ep->conn_param, ep->cm_priv_data,
+					priv_data_len);
 	ep->conn_param.retry_count = 15;
 
 	if (ep->srq_ep)
@@ -170,8 +178,8 @@ vrb_msg_ep_connect(struct fid_ep *ep_fid, const void *addr,
 		FI_WARN(&vrb_prov, FI_LOG_EP_CTRL,
 			"rdma_resolve_route failed: %s (%d)\n",
 			strerror(-ret), -ret);
-		free(ep->cm_hdr);
-		ep->cm_hdr = NULL;
+		free(ep->cm_priv_data);
+		ep->cm_priv_data = NULL;
 		return ret;
 	}
 	return 0;

prov/verbs/src/verbs_ep.c

@@ -355,7 +355,7 @@ static int vrb_close_free_ep(struct vrb_ep *ep)
 
 	free(ep->util_ep.ep_fid.msg);
 	ep->util_ep.ep_fid.msg = NULL;
-	free(ep->cm_hdr);
+	free(ep->cm_priv_data);
 
 	if (ep->util_ep.rx_cq) {
 		cq = container_of(ep->util_ep.rx_cq, struct vrb_cq, util_cq);

prov/verbs/src/verbs_eq.c

@@ -282,6 +282,15 @@ static inline int vrb_eq_copy_event_data(struct fi_eq_cm_entry *entry,
 	return datalen;
 }
 
+static void vrb_eq_skip_cm_req_hdr(const void **priv_data,
+				       size_t *priv_data_len)
+{
+	if (*priv_data_len > VRB_CM_REQ_HDR_SIZE) {
+		*priv_data = (void*)((char *)*priv_data + VRB_CM_REQ_HDR_SIZE);
+		*priv_data_len -= VRB_CM_REQ_HDR_SIZE;
+	}
+}
+
 static void vrb_eq_skip_xrc_cm_data(const void **priv_data,
 				       size_t *priv_data_len)
 {
@@ -894,6 +903,8 @@ vrb_eq_cm_process_event(struct vrb_eq *eq,
 			}
 			if (*event == FI_CONNECTED)
 				goto ack;
+		} else if (cma_event->id->route.addr.src_addr.sa_family == AF_IB) {
+			vrb_eq_skip_cm_req_hdr(&priv_data, &priv_datalen);
 		}
 		break;
 	case RDMA_CM_EVENT_CONNECT_RESPONSE: