[dpdk-dev] [PATCH v2] net/i40e: fix TSO pkt exceeds allowed buf size issue

[Xiaoyun Li <xiaoyun.li@intel.com>]

Hardware limits that max buffer size per tx descriptor should be
(16K-1)B. So when TSO enabled, the mbuf data size may exceed the
limit and cause malicious behaviour to the NIC. This patch fixes
this issue by using more tx descs for this kind of large buffer.

Fixes: 4861cde46116 ("i40e: new poll mode driver")
Cc: stable@dpdk.org

Signed-off-by: Xiaoyun Li <xiaoyun.li@intel.com>
---
v2:
 * Each pkt can have several segments so the needed tx descs should sum
 * all segments up.
---
 drivers/net/i40e/i40e_rxtx.c | 44 +++++++++++++++++++++++++++++++++++-
 1 file changed, 43 insertions(+), 1 deletion(-)

diff --git a/drivers/net/i40e/i40e_rxtx.c b/drivers/net/i40e/i40e_rxtx.c
index 17dc8c78f..ce95d8c20 100644
--- a/drivers/net/i40e/i40e_rxtx.c
+++ b/drivers/net/i40e/i40e_rxtx.c
@@ -989,6 +989,23 @@ i40e_set_tso_ctx(struct rte_mbuf *mbuf, union i40e_tx_offload tx_offload)
 	return ctx_desc;
 }
 
+/* HW requires that Tx buffer size ranges from 1B up to (16K-1)B. */
+#define I40E_MAX_DATA_PER_TXD	(16 * 1024 - 1)
+/* Calculate the number of TX descriptors needed for each pkt */
+static inline uint16_t
+i40e_calc_pkt_desc(struct rte_mbuf *tx_pkt)
+{
+	struct rte_mbuf *txd = tx_pkt;
+	uint16_t count = 0;
+
+	while (txd != NULL) {
+		count += DIV_ROUND_UP(txd->data_len, I40E_MAX_DATA_PER_TXD);
+		txd = txd->next;
+	}
+
+	return count;
+}
+
 uint16_t
 i40e_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 {
@@ -1046,8 +1063,15 @@ i40e_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 		 * The number of descriptors that must be allocated for
 		 * a packet equals to the number of the segments of that
 		 * packet plus 1 context descriptor if needed.
+		 * Recalculate the needed tx descs when TSO enabled in case
+		 * the mbuf data size exceeds max data size that hw allows
+		 * per tx desc.
 		 */
-		nb_used = (uint16_t)(tx_pkt->nb_segs + nb_ctx);
+		if (ol_flags & PKT_TX_TCP_SEG)
+			nb_used = (uint16_t)(i40e_calc_pkt_desc(tx_pkt) +
+					     nb_ctx);
+		else
+			nb_used = (uint16_t)(tx_pkt->nb_segs + nb_ctx);
 		tx_last = (uint16_t)(tx_id + nb_used - 1);
 
 		/* Circular ring */
@@ -1160,6 +1184,24 @@ i40e_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 			slen = m_seg->data_len;
 			buf_dma_addr = rte_mbuf_data_iova(m_seg);
 
+			while ((ol_flags & PKT_TX_TCP_SEG) &&
+				unlikely(slen > I40E_MAX_DATA_PER_TXD)) {
+				txd->buffer_addr =
+					rte_cpu_to_le_64(buf_dma_addr);
+				txd->cmd_type_offset_bsz =
+					i40e_build_ctob(td_cmd,
+					td_offset, I40E_MAX_DATA_PER_TXD,
+					td_tag);
+
+				buf_dma_addr += I40E_MAX_DATA_PER_TXD;
+				slen -= I40E_MAX_DATA_PER_TXD;
+
+				txe->last_id = tx_last;
+				tx_id = txe->next_id;
+				txe = txn;
+				txd = &txr[tx_id];
+				txn = &sw_ring[txe->next_id];
+			}
 			PMD_TX_LOG(DEBUG, "mbuf: %p, TDD[%u]:\n"
 				"buf_dma_addr: %#"PRIx64";\n"
 				"td_cmd: %#x;\n"
-- 
2.17.1