[PATCH 1/7] net/gve: send whole packet when mbuf is large

[Joshua Washington]

Before this patch, only one descriptor would be written per mbuf in a
packet. In cases like TSO, it is possible for a single mbuf to have more
bytes than GVE_MAX_TX_BUF_SIZE_DQO. As such, instead of simply
truncating the data down to this size, the driver should actually write
descriptors for the rest of the buffers in the mbuf segment.

To that effect, the number of descriptors needed to send a packet must
be corrected to account for the potential additional descriptors.

Fixes: 4022f9999f56 ("net/gve: support basic Tx data path for DQO")
Cc: junfeng.guo at intel.com
Cc: stable at dpdk.org

Signed-off-by: Joshua Washington <joshwash at google.com>
Reviewed-by: Ankit Garg <nktgrg at google.com>
---
 .mailmap                     |  1 +
 drivers/net/gve/gve_tx_dqo.c | 54 ++++++++++++++++++++++++------------
 2 files changed, 38 insertions(+), 17 deletions(-)

diff --git a/.mailmap b/.mailmap
index 1ea4f9446d..758878bd8b 100644
--- a/.mailmap
+++ b/.mailmap
@@ -124,6 +124,7 @@ Andy Green <andy at warmcat.com>
 Andy Moreton <andy.moreton at amd.com> <amoreton at xilinx.com> <amoreton at solarflare.com>
 Andy Pei <andy.pei at intel.com>
 Anirudh Venkataramanan <anirudh.venkataramanan at intel.com>
+Ankit Garg <nktgrg at google.com>
 Ankur Dwivedi <adwivedi at marvell.com> <ankur.dwivedi at caviumnetworks.com> <ankur.dwivedi at cavium.com>
 Anna Lukin <annal at silicom.co.il>
 Anoob Joseph <anoobj at marvell.com> <anoob.joseph at caviumnetworks.com>
diff --git a/drivers/net/gve/gve_tx_dqo.c b/drivers/net/gve/gve_tx_dqo.c
index 6984f92443..6227fa73b0 100644
--- a/drivers/net/gve/gve_tx_dqo.c
+++ b/drivers/net/gve/gve_tx_dqo.c
@@ -74,6 +74,19 @@ gve_tx_clean_dqo(struct gve_tx_queue *txq)
 	txq->complq_tail = next;
 }
 
+static uint16_t
+gve_tx_pkt_nb_data_descs(struct rte_mbuf *tx_pkt)
+{
+	int nb_descs = 0;
+
+	while (tx_pkt) {
+		nb_descs += (GVE_TX_MAX_BUF_SIZE_DQO - 1 + tx_pkt->data_len) /
+			GVE_TX_MAX_BUF_SIZE_DQO;
+		tx_pkt = tx_pkt->next;
+	}
+	return nb_descs;
+}
+
 static inline void
 gve_tx_fill_seg_desc_dqo(volatile union gve_tx_desc_dqo *desc, struct rte_mbuf *tx_pkt)
 {
@@ -97,7 +110,7 @@ gve_tx_burst_dqo(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 	uint16_t nb_to_clean;
 	uint16_t nb_tx = 0;
 	uint64_t ol_flags;
-	uint16_t nb_used;
+	uint16_t nb_descs;
 	uint16_t tx_id;
 	uint16_t sw_id;
 	uint64_t bytes;
@@ -124,14 +137,14 @@ gve_tx_burst_dqo(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 		}
 
 		ol_flags = tx_pkt->ol_flags;
-		nb_used = tx_pkt->nb_segs;
 		first_sw_id = sw_id;
 
 		tso = !!(ol_flags & RTE_MBUF_F_TX_TCP_SEG);
 		csum = !!(ol_flags & GVE_TX_CKSUM_OFFLOAD_MASK_DQO);
 
-		nb_used += tso;
-		if (txq->nb_free < nb_used)
+		nb_descs = gve_tx_pkt_nb_data_descs(tx_pkt);
+		nb_descs += tso;
+		if (txq->nb_free < nb_descs)
 			break;
 
 		if (tso) {
@@ -144,21 +157,28 @@ gve_tx_burst_dqo(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 			if (sw_ring[sw_id] != NULL)
 				PMD_DRV_LOG(DEBUG, "Overwriting an entry in sw_ring");
 
-			txd = &txr[tx_id];
 			sw_ring[sw_id] = tx_pkt;
 
-			/* fill Tx descriptor */
-			txd->pkt.buf_addr = rte_cpu_to_le_64(rte_mbuf_data_iova(tx_pkt));
-			txd->pkt.dtype = GVE_TX_PKT_DESC_DTYPE_DQO;
-			txd->pkt.compl_tag = rte_cpu_to_le_16(first_sw_id);
-			txd->pkt.buf_size = RTE_MIN(tx_pkt->data_len, GVE_TX_MAX_BUF_SIZE_DQO);
-			txd->pkt.end_of_packet = 0;
-			txd->pkt.checksum_offload_enable = csum;
+			/* fill Tx descriptors */
+			int mbuf_offset = 0;
+			while (mbuf_offset < tx_pkt->data_len) {
+				uint64_t buf_addr = rte_mbuf_data_iova(tx_pkt) +
+					mbuf_offset;
+
+				txd = &txr[tx_id];
+				txd->pkt.buf_addr = rte_cpu_to_le_64(buf_addr);
+				txd->pkt.compl_tag = rte_cpu_to_le_16(first_sw_id);
+				txd->pkt.dtype = GVE_TX_PKT_DESC_DTYPE_DQO;
+				txd->pkt.buf_size = RTE_MIN(tx_pkt->data_len - mbuf_offset,
+							    GVE_TX_MAX_BUF_SIZE_DQO);
+				txd->pkt.end_of_packet = 0;
+				txd->pkt.checksum_offload_enable = csum;
+
+				mbuf_offset += txd->pkt.buf_size;
+				tx_id = (tx_id + 1) & mask;
+			}
 
-			/* size of desc_ring and sw_ring could be different */
-			tx_id = (tx_id + 1) & mask;
 			sw_id = (sw_id + 1) & sw_mask;
-
 			bytes += tx_pkt->data_len;
 			tx_pkt = tx_pkt->next;
 		} while (tx_pkt);
@@ -166,8 +186,8 @@ gve_tx_burst_dqo(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 		/* fill the last descriptor with End of Packet (EOP) bit */
 		txd->pkt.end_of_packet = 1;
 
-		txq->nb_free -= nb_used;
-		txq->nb_used += nb_used;
+		txq->nb_free -= nb_descs;
+		txq->nb_used += nb_descs;
 	}
 
 	/* update the tail pointer if any packets were processed */
-- 
2.50.0.727.gbf7dc18ff4-goog