1 files changed, 172 insertions, 9 deletions

diff --git a/src/drivers/net/ipoib.c b/src/drivers/net/ipoib.c
index 2d351e4b..a848b711 100644
--- a/src/drivers/net/ipoib.c
+++ b/src/drivers/net/ipoib.c
@@ -44,15 +44,24 @@ extern struct ib_address_vector hack_ipoib_bcast_av;
 /** IPoIB MTU */
 #define IPOIB_MTU 2048
 
-/** Number of IPoIB send work queue entries */
+/** Number of IPoIB data send work queue entries */
 #define IPOIB_DATA_NUM_SEND_WQES 4
 
-/** Number of IPoIB receive work queue entries */
-#define IPOIB_DATA_NUM_RECV_WQES 8
+/** Number of IPoIB data receive work queue entries */
+#define IPOIB_DATA_NUM_RECV_WQES 4
 
-/** Number of IPoIB completion entries */
+/** Number of IPoIB data completion entries */
 #define IPOIB_DATA_NUM_CQES 8
 
+/** Number of IPoIB metadata send work queue entries */
+#define IPOIB_META_NUM_SEND_WQES 4
+
+/** Number of IPoIB metadata receive work queue entries */
+#define IPOIB_META_NUM_RECV_WQES 4
+
+/** Number of IPoIB metadata completion entries */
+#define IPOIB_META_NUM_CQES 8
+
 /** An IPoIB queue set */
 struct ipoib_queue_set {
 	/** Completion queue */
@@ -84,10 +93,15 @@ struct ipoib_device {
  ****************************************************************************
  */
 
+/** Broadcast QPN used in IPoIB MAC addresses
+ *
+ * This is a guaranteed invalid real QPN
+ */
+#define IPOIB_BROADCAST_QPN 0xffffffffUL
+
 /** Broadcast IPoIB address */
 static struct ipoib_mac ipoib_broadcast = {
-	.gid = { { 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
-		   0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff } },
+	.qpn = ntohl ( IPOIB_BROADCAST_QPN ),
 };
 
 /**
@@ -245,6 +259,73 @@ static int ipoib_create_qset ( struct ipoib_device *ipoib,
 }
 
 /**
+ * Transmit path record request
+ *
+ * @v ipoib		IPoIB device
+ * @v gid		Destination GID
+ * @ret rc		Return status code
+ */
+static int ipoib_get_path_record ( struct ipoib_device *ipoib,
+				   struct ib_gid *gid ) {
+	struct ib_device *ibdev = ipoib->ibdev;
+	struct io_buffer *iobuf;
+	struct ib_mad_path_record *path_record;
+	struct ib_address_vector av;
+ 	static uint32_t tid = 0;
+	int rc;
+
+	DBG ( "get_path_record():\n" );
+	int get_path_record(struct ib_gid *dgid, uint16_t *dlid_p,
+			    uint8_t *sl_p, uint8_t *rate_p);
+	uint16_t tmp_dlid;
+	uint8_t tmp_sl;
+	uint8_t tmp_rate;
+	get_path_record ( gid, &tmp_dlid, &tmp_sl, &tmp_rate );
+
+	DBG ( "ipoib_get_path_record():\n" );
+
+	/* Allocate I/O buffer */
+	iobuf = alloc_iob ( sizeof ( *path_record ) );
+	if ( ! iobuf )
+		return -ENOMEM;
+	iob_put ( iobuf, sizeof ( *path_record ) );
+	path_record = iobuf->data;
+	memset ( path_record, 0, sizeof ( *path_record ) );
+
+	/* Construct path record request */
+	path_record->mad_hdr.base_version = IB_MGMT_BASE_VERSION;
+	path_record->mad_hdr.mgmt_class = IB_MGMT_CLASS_SUBN_ADM;
+	path_record->mad_hdr.class_version = 2;
+	path_record->mad_hdr.method = IB_MGMT_METHOD_GET;
+	path_record->mad_hdr.attr_id = htons ( IB_SA_ATTR_PATH_REC );
+	path_record->mad_hdr.tid = tid++;
+	path_record->sa_hdr.comp_mask[1] =
+		htonl ( IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID );
+	memcpy ( &path_record->dgid, gid, sizeof ( path_record->dgid ) );
+	memcpy ( &path_record->sgid, &ibdev->port_gid,
+		 sizeof ( path_record->sgid ) );
+
+	DBG_HD ( path_record, sizeof ( *path_record ) );
+
+	/* Construct address vector */
+	memset ( &av, 0, sizeof ( av ) );
+	av.dlid = ibdev->sm_lid;
+	av.dest_qp = IB_SA_QPN;
+	av.qkey = IB_SA_QKEY;
+
+	/* Post send request */
+	if ( ( rc = ib_post_send ( ibdev, ipoib->meta.qp, &av,
+				   iobuf ) ) != 0 ) {
+		DBGC ( ipoib, "IPoIB %p could not send get path record: %s\n",
+		       ipoib, strerror ( rc ) );
+		free_iob ( iobuf );
+		return rc;
+	}
+
+	return 0;
+}
+
+/**
  * Transmit packet via IPoIB network device
  *
  * @v netdev		Network device
@@ -256,19 +337,29 @@ static int ipoib_transmit ( struct net_device *netdev,
 	struct ipoib_device *ipoib = netdev->priv;
 	struct ib_device *ibdev = ipoib->ibdev;
 	struct ipoib_pseudo_hdr *ipoib_pshdr = iobuf->data;
+	int rc;
 
 	if ( iob_len ( iobuf ) < sizeof ( *ipoib_pshdr ) ) {
 		DBGC ( ipoib, "IPoIB %p buffer too short\n", ipoib );
 		return -EINVAL;
 	}
 
+	DBG ( "TX pseudo-header:\n" );
+	DBG_HD ( ipoib_pshdr, sizeof ( *ipoib_pshdr ) );
+	if ( ipoib_pshdr->peer.qpn != htonl ( IPOIB_BROADCAST_QPN ) ) {
+		DBG ( "Get path record\n" );
+		rc = ipoib_get_path_record ( ipoib, &ipoib_pshdr->peer.gid );
+		free_iob ( iobuf );
+		return 0;
+	}
+
 	iob_pull ( iobuf, ( sizeof ( *ipoib_pshdr ) ) );
 	return ib_post_send ( ibdev, ipoib->data.qp,
 			      &hack_ipoib_bcast_av, iobuf );
 }
 
 /**
- * Handle IPoIB send completion
+ * Handle IPoIB data send completion
  *
  * @v ibdev		Infiniband device
  * @v qp		Queue pair
@@ -286,7 +377,7 @@ static void ipoib_data_complete_send ( struct ib_device *ibdev __unused,
 }
 
 /**
- * Handle IPoIB receive completion
+ * Handle IPoIB data receive completion
  *
  * @v ibdev		Infiniband device
  * @v qp		Queue pair
@@ -316,6 +407,61 @@ static void ipoib_data_complete_recv ( struct ib_device *ibdev __unused,
 }
 
 /**
+ * Handle IPoIB metadata send completion
+ *
+ * @v ibdev		Infiniband device
+ * @v qp		Queue pair
+ * @v completion	Completion
+ * @v iobuf		I/O buffer
+ */
+static void ipoib_meta_complete_send ( struct ib_device *ibdev __unused,
+				       struct ib_queue_pair *qp,
+				       struct ib_completion *completion,
+				       struct io_buffer *iobuf ) {
+	struct net_device *netdev = qp->owner_priv;
+	struct ipoib_device *ipoib = netdev->priv;
+
+	DBG ( "Woohoo! METADATA TX completion\n" );
+
+
+	if ( completion->syndrome ) {
+		DBGC ( ipoib, "IPoIB %p metadata TX completion error %x\n",
+		       ipoib, completion->syndrome );
+	}
+	free_iob ( iobuf );
+}
+
+/**
+ * Handle IPoIB metadata receive completion
+ *
+ * @v ibdev		Infiniband device
+ * @v qp		Queue pair
+ * @v completion	Completion
+ * @v iobuf		I/O buffer
+ */
+static void ipoib_meta_complete_recv ( struct ib_device *ibdev __unused,
+				       struct ib_queue_pair *qp,
+				       struct ib_completion *completion,
+				       struct io_buffer *iobuf ) {
+	struct net_device *netdev = qp->owner_priv;
+	struct ipoib_device *ipoib = netdev->priv;
+
+	DBG ( "***************** META TX!!!!!! ********\n" );
+
+	if ( completion->syndrome ) {
+		DBGC ( ipoib, "IPoIB %p metadata RX completion error %x\n",
+		       ipoib, completion->syndrome );
+	} else {
+		iob_put ( iobuf, completion->len );
+		DBG ( "Metadata RX:\n" );
+		DBG_HD ( iobuf->data, iob_len ( iobuf ) );
+	}
+
+	ipoib->meta.recv_fill--;
+	free_iob ( iobuf );
+}
+
+/**
  * Refill IPoIB receive ring
  *
  * @v ipoib		IPoIB device
@@ -349,6 +495,9 @@ static void ipoib_poll ( struct net_device *netdev ) {
 
 	ib_poll_cq ( ibdev, ipoib->data.cq, ipoib_data_complete_send,
 		     ipoib_data_complete_recv );
+	ib_poll_cq ( ibdev, ipoib->meta.cq, ipoib_meta_complete_send,
+		     ipoib_meta_complete_recv );
+	ipoib_refill_recv ( ipoib, &ipoib->meta );
 	ipoib_refill_recv ( ipoib, &ipoib->data );
 }
 
@@ -382,7 +531,8 @@ static int ipoib_open ( struct net_device *netdev ) {
 		return rc;
 	}
 
-	/* Fill receive ring */
+	/* Fill receive rings */
+	ipoib_refill_recv ( ipoib, &ipoib->meta );
 	ipoib_refill_recv ( ipoib, &ipoib->data );
 
 	return 0;
@@ -436,6 +586,17 @@ int ipoib_probe ( struct ib_device *ibdev ) {
 	ipoib->netdev = netdev;
 	ipoib->ibdev = ibdev;
 
+	/* Allocate metadata queue set */
+	if ( ( rc = ipoib_create_qset ( ipoib, &ipoib->meta,
+					IPOIB_META_NUM_CQES,
+					IPOIB_META_NUM_SEND_WQES,
+					IPOIB_META_NUM_RECV_WQES,
+					IB_SA_QKEY ) ) != 0 ) {
+		DBGC ( ipoib, "IPoIB %p could not allocate metadata QP: %s\n",
+		       ipoib, strerror ( rc ) );
+		goto err_create_meta_qset;
+	}
+
 	/* Allocate data queue set */
 	if ( ( rc = ipoib_create_qset ( ipoib, &ipoib->data,
 					IPOIB_DATA_NUM_CQES,
@@ -461,6 +622,8 @@ int ipoib_probe ( struct ib_device *ibdev ) {
  err_register_netdev:
 	ipoib_destroy_qset ( ipoib, &ipoib->data );
  err_create_data_qset:
+	ipoib_destroy_qset ( ipoib, &ipoib->meta );
+ err_create_meta_qset:
 	netdev_nullify ( netdev );
 	netdev_put ( netdev );
 	return rc;