Some quick comments that got cut out of the original mail.

+
+/*
+ * map single buffer
+ */
+static int bnx2i_map_single_buf(struct bnx2i_hba *hba,
+				       struct bnx2i_cmd *cmd)
+{
+	struct scsi_cmnd *sc = cmd->scsi_cmd;
+	struct iscsi_bd *bd = cmd->bd_tbl->bd_tbl;
+	int byte_count;
+	int bd_count;
+	u64 addr;
+
+	byte_count = sc->request_bufflen;
+	sc->SCp.dma_handle =
+		pci_map_single(hba->pci_dev, sc->request_buffer,
+			       sc->request_bufflen, sc->sc_data_direction);
+	addr = sc->SCp.dma_handle;
+
+	if (byte_count > MAX_BD_LENGTH) {
+		bd_count = bnx2i_split_bd(cmd, addr, byte_count, 0);
+	} else {
+		bd_count = 1;
+		bd[0].buffer_addr_lo = addr & 0xffffffff;
+		bd[0].buffer_addr_hi = addr >> 32;
+		bd[0].buffer_length = sc->request_bufflen;
+		bd[0].flags = ISCSI_BD_FIRST_IN_BD_CHAIN |
+			      ISCSI_BD_LAST_IN_BD_CHAIN;
+	}
+	bd[bd_count - 1].flags |= ISCSI_BD_LAST_IN_BD_CHAIN;
+
+	return bd_count;
+}

I think you should always be getting use_sg greater than zero now, so 
the map single path is not needed.

+
+
+/*
+ * map SG list
+ */
+static int bnx2i_map_sg(struct bnx2i_hba *hba, struct bnx2i_cmd *cmd)
+{
+	struct scsi_cmnd *sc = cmd->scsi_cmd;
+	struct iscsi_bd *bd = cmd->bd_tbl->bd_tbl;
+	struct scatterlist *sg;
+	int byte_count = 0;
+	int sg_frags;
+	int bd_count = 0;
+	int sg_count;
+	int sg_len;
+	u64 addr;
+	int i;
+
+	sg = sc->request_buffer;
+	sg_count = pci_map_sg(hba->pci_dev, sg, sc->use_sg,
+			      sc->sc_data_direction);
+
+	for (i = 0; i < sg_count; i++) {
+		sg_len = sg_dma_len(sg);
+		addr = sg_dma_address(sg);
+		if (sg_len > MAX_BD_LENGTH)
+			sg_frags = bnx2i_split_bd(cmd, addr, sg_len,
+						  bd_count);

If you call blk_queue_max_segment_size() in the slave_configure callout 
you can limit the size of the segments that the block layer builds so 
they are smaller than MAX_BD_LENGTH. However, I am not sure how useful 
that is. I think DMA-API.txt states that the mapping code is ok to 
merged mutliple sglists entries into one so I think that means that we 
can still end up with an entry that is larger than MAX_BD_LENGTH. Not 
sure if there is way to tell the pci/dma map_sg code to limit this too.