On Mon, 30 Jan 2017 20:49:35 -0500, Michael Chan wrote:

+static int bnxt_xmit_xdp(struct bnxt *bp, struct bnxt_napi *bnapi,
+			 struct page *page, dma_addr_t mapping, u32 offset,
+			 u32 len)
+{
+	struct bnxt_tx_ring_info *txr = bnapi->tx_ring;
+	struct bnxt_sw_tx_bd *tx_buf;
+	struct tx_bd_ext *txbd1;
+	struct tx_bd *txbd;
+	u32 flags;
+	u16 prod;
+
+	if (bnxt_tx_avail(bp, txr) < 2)
+		return -ENOSPC;
+
+	prod = txr->tx_prod;
+	txbd = &txr->tx_desc_ring[TX_RING(prod)][TX_IDX(prod)];
+
+	tx_buf = &txr->tx_buf_ring[prod];
+	tx_buf->page = page;
+	dma_unmap_addr_set(tx_buf, mapping, mapping);
+	flags = (len << TX_BD_LEN_SHIFT) | TX_BD_TYPE_LONG_TX_BD |
+		(2 << TX_BD_FLAGS_BD_CNT_SHIFT) | TX_BD_FLAGS_PACKET_END |
+		bnxt_lhint_arr[len >> 9];
+	txbd->tx_bd_len_flags_type = cpu_to_le32(flags);
+	txbd->tx_bd_opaque = prod;
+	txbd->tx_bd_haddr = cpu_to_le64(mapping + offset);
+
+	prod = NEXT_TX(prod);
+	txbd1 = (struct tx_bd_ext *)
+		&txr->tx_desc_ring[TX_RING(prod)][TX_IDX(prod)];
+
+	txbd1->tx_bd_hsize_lflags = cpu_to_le32(0);
+	txbd1->tx_bd_mss = cpu_to_le32(0);
+	txbd1->tx_bd_cfa_action = cpu_to_le32(0);
+	txbd1->tx_bd_cfa_meta = cpu_to_le32(0);
+
+	prod = NEXT_TX(prod);
+	txr->tx_prod = prod;
+	return 0;
+}

Are you not lacking DMA syncs in general?  You map the buffers
bidirectionally, but I failed to find any dma_syncs.  I would expect
one before you run xdp and one before you TX because packet could have
been modified.  What am I missing?