[dpdk-dev] [PATCH 0/6] bnxt patchset for Tx performance optimization

[dpdk-dev] [PATCH 0/6] bnxt patchset for Tx performance optimization

[Ajit Khaparde]

This patchset brings in few fixes and some changes tuning
the Tx datapath performance in bnxt PMD.
Please apply.

Ajit Khaparde (4):
  net/bnxt: fix an issue seen with TSO
  net/bnxt: add support to update ipid
  net/bnxt: check for some error conditions in Tx path
  net/bnxt: support bulk free of Tx mbufs

Sriharsha Basavapatna (2):
  net/bnxt: fix a couple of issues with Tx batching
  net/bnxt: optimize Tx batching code

 drivers/net/bnxt/bnxt_txq.c |  11 +++
 drivers/net/bnxt/bnxt_txq.h |   2 +-
 drivers/net/bnxt/bnxt_txr.c | 160 ++++++++++++++++++++++++------------
 drivers/net/bnxt/bnxt_txr.h |   2 +
 4 files changed, 123 insertions(+), 52 deletions(-)

-- 
2.20.1 (Apple Git-117)

[dpdk-dev] [PATCH 0/6] bnxt patchset for Tx performance optimization

[Ajit Khaparde]

This patchset brings in few fixes and some changes tuning
the Tx datapath performance in bnxt PMD.
Please apply.

Ajit Khaparde (4):
  net/bnxt: fix an issue seen with TSO
  net/bnxt: add support to update ipid
  net/bnxt: check for some error conditions in Tx path
  net/bnxt: support bulk free of Tx mbufs

Sriharsha Basavapatna (2):
  net/bnxt: fix a couple of issues with Tx batching
  net/bnxt: optimize Tx batching code

 drivers/net/bnxt/bnxt_txq.c |  11 +++
 drivers/net/bnxt/bnxt_txq.h |   2 +-
 drivers/net/bnxt/bnxt_txr.c | 160 ++++++++++++++++++++++++------------
 drivers/net/bnxt/bnxt_txr.h |   2 +
 4 files changed, 123 insertions(+), 52 deletions(-)

-- 
2.20.1 (Apple Git-117)

[dpdk-dev] [PATCH 1/6] net/bnxt: fix an issue seen with TSO

[Ajit Khaparde]

We wrongly update lflags in the Tx descriptor; avoid it. Also, instead of
calculating the last producer index to see if mbuf segments are chained,
check if the pointer is NULL to iterate through the segment list.

Fixes: 6eb3cc2294fd ("net/bnxt: add initial Tx code")
Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
Reviewed-by: Somnath Kotur <somnath.kotur@broadcom.com>
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
---
 drivers/net/bnxt/bnxt_txr.c | 15 ++++++++-------
 1 file changed, 8 insertions(+), 7 deletions(-)

diff --git a/drivers/net/bnxt/bnxt_txr.c b/drivers/net/bnxt/bnxt_txr.c
index 39be7bdfa..6f55f3c53 100644
--- a/drivers/net/bnxt/bnxt_txr.c
+++ b/drivers/net/bnxt/bnxt_txr.c
@@ -123,7 +123,6 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 	struct tx_bd_long_hi *txbd1 = NULL;
 	uint32_t vlan_tag_flags, cfa_action;
 	bool long_bd = false;
-	uint16_t last_prod = 0;
 	struct rte_mbuf *m_seg;
 	struct bnxt_sw_tx_bd *tx_buf;
 	static const uint32_t lhint_arr[4] = {
@@ -143,8 +142,6 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 	tx_buf = &txr->tx_buf_ring[txr->tx_prod];
 	tx_buf->mbuf = tx_pkt;
 	tx_buf->nr_bds = long_bd + tx_pkt->nb_segs;
-	last_prod = (txr->tx_prod + tx_buf->nr_bds - 1) &
-				txr->tx_ring_struct->ring_mask;
 
 	if (unlikely(bnxt_tx_avail(txr) < tx_buf->nr_bds))
 		return -ENOMEM;
@@ -193,11 +190,17 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 		txbd1->cfa_action = cfa_action;
 
 		if (tx_pkt->ol_flags & PKT_TX_TCP_SEG) {
+			uint16_t hdr_size;
+
 			/* TSO */
 			txbd1->lflags |= TX_BD_LONG_LFLAGS_LSO;
-			txbd1->hdr_size = tx_pkt->l2_len + tx_pkt->l3_len +
+			hdr_size = tx_pkt->l2_len + tx_pkt->l3_len +
 					tx_pkt->l4_len + tx_pkt->outer_l2_len +
 					tx_pkt->outer_l3_len;
+			/* The hdr_size is multiple of 16bit units not 8bit.
+			 * Hence divide by 2.
+			 */
+			txbd1->hdr_size = hdr_size >> 1;
 			txbd1->mss = tx_pkt->tso_segsz;
 
 		} else if ((tx_pkt->ol_flags & PKT_TX_OIP_IIP_TCP_UDP_CKSUM) ==
@@ -282,7 +285,7 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 
 	m_seg = tx_pkt->next;
 	/* i is set at the end of the if(long_bd) block */
-	while (txr->tx_prod != last_prod) {
+	while (m_seg) {
 		txr->tx_prod = RING_NEXT(txr->tx_ring_struct, txr->tx_prod);
 		tx_buf = &txr->tx_buf_ring[txr->tx_prod];
 
@@ -295,8 +298,6 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 	}
 
 	txbd->flags_type |= TX_BD_LONG_FLAGS_PACKET_END;
-	if (txbd1)
-		txbd1->lflags = rte_cpu_to_le_32(txbd1->lflags);
 
 	txr->tx_prod = RING_NEXT(txr->tx_ring_struct, txr->tx_prod);
 
-- 
2.20.1 (Apple Git-117)

[dpdk-dev] [PATCH 1/6] net/bnxt: fix an issue seen with TSO

[Ajit Khaparde]

We wrongly update lflags in the Tx descriptor; avoid it. Also, instead of
calculating the last producer index to see if mbuf segments are chained,
check if the pointer is NULL to iterate through the segment list.

Fixes: 6eb3cc2294fd ("net/bnxt: add initial Tx code")
Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
Reviewed-by: Somnath Kotur <somnath.kotur@broadcom.com>
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
---
 drivers/net/bnxt/bnxt_txr.c | 15 ++++++++-------
 1 file changed, 8 insertions(+), 7 deletions(-)

diff --git a/drivers/net/bnxt/bnxt_txr.c b/drivers/net/bnxt/bnxt_txr.c
index 39be7bdfa..6f55f3c53 100644
--- a/drivers/net/bnxt/bnxt_txr.c
+++ b/drivers/net/bnxt/bnxt_txr.c
@@ -123,7 +123,6 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 	struct tx_bd_long_hi *txbd1 = NULL;
 	uint32_t vlan_tag_flags, cfa_action;
 	bool long_bd = false;
-	uint16_t last_prod = 0;
 	struct rte_mbuf *m_seg;
 	struct bnxt_sw_tx_bd *tx_buf;
 	static const uint32_t lhint_arr[4] = {
@@ -143,8 +142,6 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 	tx_buf = &txr->tx_buf_ring[txr->tx_prod];
 	tx_buf->mbuf = tx_pkt;
 	tx_buf->nr_bds = long_bd + tx_pkt->nb_segs;
-	last_prod = (txr->tx_prod + tx_buf->nr_bds - 1) &
-				txr->tx_ring_struct->ring_mask;
 
 	if (unlikely(bnxt_tx_avail(txr) < tx_buf->nr_bds))
 		return -ENOMEM;
@@ -193,11 +190,17 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 		txbd1->cfa_action = cfa_action;
 
 		if (tx_pkt->ol_flags & PKT_TX_TCP_SEG) {
+			uint16_t hdr_size;
+
 			/* TSO */
 			txbd1->lflags |= TX_BD_LONG_LFLAGS_LSO;
-			txbd1->hdr_size = tx_pkt->l2_len + tx_pkt->l3_len +
+			hdr_size = tx_pkt->l2_len + tx_pkt->l3_len +
 					tx_pkt->l4_len + tx_pkt->outer_l2_len +
 					tx_pkt->outer_l3_len;
+			/* The hdr_size is multiple of 16bit units not 8bit.
+			 * Hence divide by 2.
+			 */
+			txbd1->hdr_size = hdr_size >> 1;
 			txbd1->mss = tx_pkt->tso_segsz;
 
 		} else if ((tx_pkt->ol_flags & PKT_TX_OIP_IIP_TCP_UDP_CKSUM) ==
@@ -282,7 +285,7 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 
 	m_seg = tx_pkt->next;
 	/* i is set at the end of the if(long_bd) block */
-	while (txr->tx_prod != last_prod) {
+	while (m_seg) {
 		txr->tx_prod = RING_NEXT(txr->tx_ring_struct, txr->tx_prod);
 		tx_buf = &txr->tx_buf_ring[txr->tx_prod];
 
@@ -295,8 +298,6 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 	}
 
 	txbd->flags_type |= TX_BD_LONG_FLAGS_PACKET_END;
-	if (txbd1)
-		txbd1->lflags = rte_cpu_to_le_32(txbd1->lflags);
 
 	txr->tx_prod = RING_NEXT(txr->tx_ring_struct, txr->tx_prod);
 
-- 
2.20.1 (Apple Git-117)

[dpdk-dev] [PATCH 2/6] net/bnxt: add support to update ipid

[Ajit Khaparde]

Add support to update ipid during TSO.

Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
Reviewed-by: Somnath Kotur <somnath.kotur@broadcom.com>
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
---
 drivers/net/bnxt/bnxt_txr.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/net/bnxt/bnxt_txr.c b/drivers/net/bnxt/bnxt_txr.c
index 6f55f3c53..3a0d73af2 100644
--- a/drivers/net/bnxt/bnxt_txr.c
+++ b/drivers/net/bnxt/bnxt_txr.c
@@ -193,7 +193,8 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 			uint16_t hdr_size;
 
 			/* TSO */
-			txbd1->lflags |= TX_BD_LONG_LFLAGS_LSO;
+			txbd1->lflags |= TX_BD_LONG_LFLAGS_LSO |
+					 TX_BD_LONG_LFLAGS_T_IPID;
 			hdr_size = tx_pkt->l2_len + tx_pkt->l3_len +
 					tx_pkt->l4_len + tx_pkt->outer_l2_len +
 					tx_pkt->outer_l3_len;
-- 
2.20.1 (Apple Git-117)

[dpdk-dev] [PATCH 2/6] net/bnxt: add support to update ipid

[Ajit Khaparde]

Add support to update ipid during TSO.

Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
Reviewed-by: Somnath Kotur <somnath.kotur@broadcom.com>
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
---
 drivers/net/bnxt/bnxt_txr.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/net/bnxt/bnxt_txr.c b/drivers/net/bnxt/bnxt_txr.c
index 6f55f3c53..3a0d73af2 100644
--- a/drivers/net/bnxt/bnxt_txr.c
+++ b/drivers/net/bnxt/bnxt_txr.c
@@ -193,7 +193,8 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 			uint16_t hdr_size;
 
 			/* TSO */
-			txbd1->lflags |= TX_BD_LONG_LFLAGS_LSO;
+			txbd1->lflags |= TX_BD_LONG_LFLAGS_LSO |
+					 TX_BD_LONG_LFLAGS_T_IPID;
 			hdr_size = tx_pkt->l2_len + tx_pkt->l3_len +
 					tx_pkt->l4_len + tx_pkt->outer_l2_len +
 					tx_pkt->outer_l3_len;
-- 
2.20.1 (Apple Git-117)

[dpdk-dev] [PATCH 3/6] net/bnxt: check for some error conditions in Tx path

[Ajit Khaparde]

The HW can have limits on the minimum packet size it can support,
or the maximum number of segments it can support. Check for such
possibilities. Also check if we are going to have a 0 length buffer.

Fixes: 6eb3cc2294fd ("net/bnxt: add initial Tx code")
Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
Reviewed-by: Somnath Kotur <somnath.kotur@broadcom.com>
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
---
 drivers/net/bnxt/bnxt_txr.c | 31 ++++++++++++++++++++++++++++++-
 drivers/net/bnxt/bnxt_txr.h |  2 ++
 2 files changed, 32 insertions(+), 1 deletion(-)

diff --git a/drivers/net/bnxt/bnxt_txr.c b/drivers/net/bnxt/bnxt_txr.c
index 3a0d73af2..9684fb177 100644
--- a/drivers/net/bnxt/bnxt_txr.c
+++ b/drivers/net/bnxt/bnxt_txr.c
@@ -143,6 +143,33 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 	tx_buf->mbuf = tx_pkt;
 	tx_buf->nr_bds = long_bd + tx_pkt->nb_segs;
 
+	/* Check if number of Tx descriptors is above HW limit */
+	if (unlikely(tx_buf->nr_bds > BNXT_MAX_TSO_SEGS)) {
+		PMD_DRV_LOG(ERR,
+			    "Num descriptors %d exceeds HW limit\n",
+			    tx_buf->nr_bds);
+		return -ENOSPC;
+	}
+
+	/* If packet length is less than minimum packet size, pad it */
+	if (unlikely(rte_pktmbuf_pkt_len(tx_pkt) < BNXT_MIN_PKT_SIZE)) {
+		uint8_t pad = BNXT_MIN_PKT_SIZE - rte_pktmbuf_pkt_len(tx_pkt);
+		char *seg = rte_pktmbuf_append(tx_pkt, pad);
+
+		if (!seg) {
+			PMD_DRV_LOG(ERR,
+				    "Failed to pad mbuf by %d bytes\n",
+				    pad);
+			return -ENOMEM;
+		}
+
+		/* Note: data_len, pkt len are updated in rte_pktmbuf_append */
+		memset(seg, 0, pad);
+	}
+
+	/* Check non zero data_len */
+	RTE_VERIFY(tx_pkt->data_len);
+
 	if (unlikely(bnxt_tx_avail(txr) < tx_buf->nr_bds))
 		return -ENOMEM;
 
@@ -203,6 +230,7 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 			 */
 			txbd1->hdr_size = hdr_size >> 1;
 			txbd1->mss = tx_pkt->tso_segsz;
+			RTE_VERIFY(txbd1->mss);
 
 		} else if ((tx_pkt->ol_flags & PKT_TX_OIP_IIP_TCP_UDP_CKSUM) ==
 			   PKT_TX_OIP_IIP_TCP_UDP_CKSUM) {
@@ -285,8 +313,9 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 	}
 
 	m_seg = tx_pkt->next;
-	/* i is set at the end of the if(long_bd) block */
 	while (m_seg) {
+		/* Check non zero data_len */
+		RTE_VERIFY(m_seg->data_len);
 		txr->tx_prod = RING_NEXT(txr->tx_ring_struct, txr->tx_prod);
 		tx_buf = &txr->tx_buf_ring[txr->tx_prod];
 
diff --git a/drivers/net/bnxt/bnxt_txr.h b/drivers/net/bnxt/bnxt_txr.h
index 7f3c7cdb0..f802d5080 100644
--- a/drivers/net/bnxt/bnxt_txr.h
+++ b/drivers/net/bnxt/bnxt_txr.h
@@ -10,6 +10,8 @@
 
 #define MAX_TX_RINGS	16
 #define BNXT_TX_PUSH_THRESH 92
+#define BNXT_MAX_TSO_SEGS	32
+#define BNXT_MIN_PKT_SIZE	52
 
 #define B_TX_DB(db, prod)	rte_write32((DB_KEY_TX | (prod)), db)
 
-- 
2.20.1 (Apple Git-117)

[dpdk-dev] [PATCH 3/6] net/bnxt: check for some error conditions in Tx path

[Ajit Khaparde]

The HW can have limits on the minimum packet size it can support,
or the maximum number of segments it can support. Check for such
possibilities. Also check if we are going to have a 0 length buffer.

Fixes: 6eb3cc2294fd ("net/bnxt: add initial Tx code")
Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
Reviewed-by: Somnath Kotur <somnath.kotur@broadcom.com>
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
---
 drivers/net/bnxt/bnxt_txr.c | 31 ++++++++++++++++++++++++++++++-
 drivers/net/bnxt/bnxt_txr.h |  2 ++
 2 files changed, 32 insertions(+), 1 deletion(-)

diff --git a/drivers/net/bnxt/bnxt_txr.c b/drivers/net/bnxt/bnxt_txr.c
index 3a0d73af2..9684fb177 100644
--- a/drivers/net/bnxt/bnxt_txr.c
+++ b/drivers/net/bnxt/bnxt_txr.c
@@ -143,6 +143,33 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 	tx_buf->mbuf = tx_pkt;
 	tx_buf->nr_bds = long_bd + tx_pkt->nb_segs;
 
+	/* Check if number of Tx descriptors is above HW limit */
+	if (unlikely(tx_buf->nr_bds > BNXT_MAX_TSO_SEGS)) {
+		PMD_DRV_LOG(ERR,
+			    "Num descriptors %d exceeds HW limit\n",
+			    tx_buf->nr_bds);
+		return -ENOSPC;
+	}
+
+	/* If packet length is less than minimum packet size, pad it */
+	if (unlikely(rte_pktmbuf_pkt_len(tx_pkt) < BNXT_MIN_PKT_SIZE)) {
+		uint8_t pad = BNXT_MIN_PKT_SIZE - rte_pktmbuf_pkt_len(tx_pkt);
+		char *seg = rte_pktmbuf_append(tx_pkt, pad);
+
+		if (!seg) {
+			PMD_DRV_LOG(ERR,
+				    "Failed to pad mbuf by %d bytes\n",
+				    pad);
+			return -ENOMEM;
+		}
+
+		/* Note: data_len, pkt len are updated in rte_pktmbuf_append */
+		memset(seg, 0, pad);
+	}
+
+	/* Check non zero data_len */
+	RTE_VERIFY(tx_pkt->data_len);
+
 	if (unlikely(bnxt_tx_avail(txr) < tx_buf->nr_bds))
 		return -ENOMEM;
 
@@ -203,6 +230,7 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 			 */
 			txbd1->hdr_size = hdr_size >> 1;
 			txbd1->mss = tx_pkt->tso_segsz;
+			RTE_VERIFY(txbd1->mss);
 
 		} else if ((tx_pkt->ol_flags & PKT_TX_OIP_IIP_TCP_UDP_CKSUM) ==
 			   PKT_TX_OIP_IIP_TCP_UDP_CKSUM) {
@@ -285,8 +313,9 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 	}
 
 	m_seg = tx_pkt->next;
-	/* i is set at the end of the if(long_bd) block */
 	while (m_seg) {
+		/* Check non zero data_len */
+		RTE_VERIFY(m_seg->data_len);
 		txr->tx_prod = RING_NEXT(txr->tx_ring_struct, txr->tx_prod);
 		tx_buf = &txr->tx_buf_ring[txr->tx_prod];
 
diff --git a/drivers/net/bnxt/bnxt_txr.h b/drivers/net/bnxt/bnxt_txr.h
index 7f3c7cdb0..f802d5080 100644
--- a/drivers/net/bnxt/bnxt_txr.h
+++ b/drivers/net/bnxt/bnxt_txr.h
@@ -10,6 +10,8 @@
 
 #define MAX_TX_RINGS	16
 #define BNXT_TX_PUSH_THRESH 92
+#define BNXT_MAX_TSO_SEGS	32
+#define BNXT_MIN_PKT_SIZE	52
 
 #define B_TX_DB(db, prod)	rte_write32((DB_KEY_TX | (prod)), db)
 
-- 
2.20.1 (Apple Git-117)

[dpdk-dev] [PATCH 4/6] net/bnxt: fix a couple of issues with Tx batching

[Ajit Khaparde]

From: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>

This patch addresses the following issues with Tx batching:

1. Tx stall observed in some conditions:

The batching code doesn't request for a completion when only a partial
chain of packets is transmitted due to mbuf allocation errors. Because
of this, Tx consumer index is not updated correctly and it eventually
leads to qfull condition. Fix this by requesting a completion for the
last packet in the partial chain that is transmitted successfully.

2. Tx stall seen with Jumbo frames:

With jumbo frames, number of TxBDs is > 1. While setting up these
additional BDs in bnxt_start_xmit(), the flags field is being set using
the OR-assignment operator. We end up using a stale value of the flags
field (from a previous use of that descriptor). This results in an
invalid completion and eventually leads to tx stall. Fix this to just
assign the flags field with the right value.

Fixes: 5735eb241947 ("net/bnxt: support Tx batching")
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Reviewed-by: Ajit Kumar Khaparde <ajit.khaparde@broadcom.com>
Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
---
 drivers/net/bnxt/bnxt_txq.h |  1 -
 drivers/net/bnxt/bnxt_txr.c | 57 ++++++++++++++++++++++---------------
 2 files changed, 34 insertions(+), 24 deletions(-)

diff --git a/drivers/net/bnxt/bnxt_txq.h b/drivers/net/bnxt/bnxt_txq.h
index f2c712a75..720ca90cf 100644
--- a/drivers/net/bnxt/bnxt_txq.h
+++ b/drivers/net/bnxt/bnxt_txq.h
@@ -24,7 +24,6 @@ struct bnxt_tx_queue {
 	uint8_t			wthresh; /* Write-back threshold reg */
 	uint32_t		ctx_curr; /* Hardware context states */
 	uint8_t			tx_deferred_start; /* not in global dev start */
-	uint8_t			cmpl_next; /* Next BD to trigger a compl */
 
 	struct bnxt		*bp;
 	int			index;
diff --git a/drivers/net/bnxt/bnxt_txr.c b/drivers/net/bnxt/bnxt_txr.c
index 9684fb177..186934136 100644
--- a/drivers/net/bnxt/bnxt_txr.c
+++ b/drivers/net/bnxt/bnxt_txr.c
@@ -103,26 +103,33 @@ int bnxt_init_tx_ring_struct(struct bnxt_tx_queue *txq, unsigned int socket_id)
 	return 0;
 }
 
-static inline uint32_t bnxt_tx_avail(struct bnxt_tx_ring_info *txr)
+static inline uint32_t bnxt_tx_bds_in_hw(struct bnxt_tx_queue *txq)
+{
+	return ((txq->tx_ring->tx_prod - txq->tx_ring->tx_cons) &
+		txq->tx_ring->tx_ring_struct->ring_mask);
+}
+
+static inline uint32_t bnxt_tx_avail(struct bnxt_tx_queue *txq)
 {
 	/* Tell compiler to fetch tx indices from memory. */
 	rte_compiler_barrier();
 
-	return txr->tx_ring_struct->ring_size -
-		((txr->tx_prod - txr->tx_cons) &
-			txr->tx_ring_struct->ring_mask) - 1;
+	return ((txq->tx_ring->tx_ring_struct->ring_size -
+		 bnxt_tx_bds_in_hw(txq)) - 1);
 }
 
 static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 				struct bnxt_tx_queue *txq,
 				uint16_t *coal_pkts,
-				uint16_t *cmpl_next)
+				uint16_t *cmpl_next,
+				struct tx_bd_long **last_txbd)
 {
 	struct bnxt_tx_ring_info *txr = txq->tx_ring;
 	struct tx_bd_long *txbd;
 	struct tx_bd_long_hi *txbd1 = NULL;
 	uint32_t vlan_tag_flags, cfa_action;
 	bool long_bd = false;
+	unsigned short nr_bds = 0;
 	struct rte_mbuf *m_seg;
 	struct bnxt_sw_tx_bd *tx_buf;
 	static const uint32_t lhint_arr[4] = {
@@ -139,15 +146,14 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 				PKT_TX_TUNNEL_GENEVE))
 		long_bd = true;
 
-	tx_buf = &txr->tx_buf_ring[txr->tx_prod];
-	tx_buf->mbuf = tx_pkt;
-	tx_buf->nr_bds = long_bd + tx_pkt->nb_segs;
+	nr_bds = long_bd + tx_pkt->nb_segs;
+	if (unlikely(bnxt_tx_avail(txq) < nr_bds))
+		return -ENOMEM;
 
 	/* Check if number of Tx descriptors is above HW limit */
-	if (unlikely(tx_buf->nr_bds > BNXT_MAX_TSO_SEGS)) {
+	if (unlikely(nr_bds > BNXT_MAX_TSO_SEGS)) {
 		PMD_DRV_LOG(ERR,
-			    "Num descriptors %d exceeds HW limit\n",
-			    tx_buf->nr_bds);
+			    "Num descriptors %d exceeds HW limit\n", nr_bds);
 		return -ENOSPC;
 	}
 
@@ -170,12 +176,13 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 	/* Check non zero data_len */
 	RTE_VERIFY(tx_pkt->data_len);
 
-	if (unlikely(bnxt_tx_avail(txr) < tx_buf->nr_bds))
-		return -ENOMEM;
+	tx_buf = &txr->tx_buf_ring[txr->tx_prod];
+	tx_buf->mbuf = tx_pkt;
+	tx_buf->nr_bds = nr_bds;
 
 	txbd = &txr->tx_desc_ring[txr->tx_prod];
 	txbd->opaque = *coal_pkts;
-	txbd->flags_type = tx_buf->nr_bds << TX_BD_LONG_FLAGS_BD_CNT_SFT;
+	txbd->flags_type = nr_bds << TX_BD_LONG_FLAGS_BD_CNT_SFT;
 	txbd->flags_type |= TX_BD_SHORT_FLAGS_COAL_NOW;
 	if (!*cmpl_next) {
 		txbd->flags_type |= TX_BD_LONG_FLAGS_NO_CMPL;
@@ -189,6 +196,7 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 	else
 		txbd->flags_type |= lhint_arr[tx_pkt->pkt_len >> 9];
 	txbd->address = rte_cpu_to_le_64(rte_mbuf_data_iova(tx_buf->mbuf));
+	*last_txbd = txbd;
 
 	if (long_bd) {
 		txbd->flags_type |= TX_BD_LONG_TYPE_TX_BD_LONG;
@@ -321,7 +329,7 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 
 		txbd = &txr->tx_desc_ring[txr->tx_prod];
 		txbd->address = rte_cpu_to_le_64(rte_mbuf_data_iova(m_seg));
-		txbd->flags_type |= TX_BD_SHORT_TYPE_TX_BD_SHORT;
+		txbd->flags_type = TX_BD_SHORT_TYPE_TX_BD_SHORT;
 		txbd->len = m_seg->data_len;
 
 		m_seg = m_seg->next;
@@ -371,8 +379,7 @@ static int bnxt_handle_tx_cp(struct bnxt_tx_queue *txq)
 	uint32_t ring_mask = cp_ring_struct->ring_mask;
 	uint32_t opaque = 0;
 
-	if (((txq->tx_ring->tx_prod - txq->tx_ring->tx_cons) &
-		txq->tx_ring->tx_ring_struct->ring_mask) < txq->tx_free_thresh)
+	if (bnxt_tx_bds_in_hw(txq) < txq->tx_free_thresh)
 		return 0;
 
 	do {
@@ -411,7 +418,8 @@ uint16_t bnxt_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 	struct bnxt_tx_queue *txq = tx_queue;
 	uint16_t nb_tx_pkts = 0;
 	uint16_t coal_pkts = 0;
-	uint16_t cmpl_next = txq->cmpl_next;
+	uint16_t cmpl_next = 0;
+	struct tx_bd_long *last_txbd = NULL;
 
 	/* Handle TX completions */
 	bnxt_handle_tx_cp(txq);
@@ -422,20 +430,23 @@ uint16_t bnxt_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 		return 0;
 	}
 
-	txq->cmpl_next = 0;
 	/* Handle TX burst request */
 	for (nb_tx_pkts = 0; nb_tx_pkts < nb_pkts; nb_tx_pkts++) {
 		int rc;
 
-		/* Request a completion on first and last packet */
+		/* Request a completion on the last packet */
 		cmpl_next |= (nb_pkts == nb_tx_pkts + 1);
 		coal_pkts++;
 		rc = bnxt_start_xmit(tx_pkts[nb_tx_pkts], txq,
-				&coal_pkts, &cmpl_next);
+				     &coal_pkts, &cmpl_next, &last_txbd);
 
 		if (unlikely(rc)) {
-			/* Request a completion in next cycle */
-			txq->cmpl_next = 1;
+			/* Request a completion on the last successfully
+			 * enqueued packet
+			 */
+			if (last_txbd)
+				last_txbd->flags_type &=
+					~TX_BD_LONG_FLAGS_NO_CMPL;
 			break;
 		}
 	}
-- 
2.20.1 (Apple Git-117)

[dpdk-dev] [PATCH 4/6] net/bnxt: fix a couple of issues with Tx batching

[Ajit Khaparde]

From: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>

This patch addresses the following issues with Tx batching:

1. Tx stall observed in some conditions:

The batching code doesn't request for a completion when only a partial
chain of packets is transmitted due to mbuf allocation errors. Because
of this, Tx consumer index is not updated correctly and it eventually
leads to qfull condition. Fix this by requesting a completion for the
last packet in the partial chain that is transmitted successfully.

2. Tx stall seen with Jumbo frames:

With jumbo frames, number of TxBDs is > 1. While setting up these
additional BDs in bnxt_start_xmit(), the flags field is being set using
the OR-assignment operator. We end up using a stale value of the flags
field (from a previous use of that descriptor). This results in an
invalid completion and eventually leads to tx stall. Fix this to just
assign the flags field with the right value.

Fixes: 5735eb241947 ("net/bnxt: support Tx batching")
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Reviewed-by: Ajit Kumar Khaparde <ajit.khaparde@broadcom.com>
Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
---
 drivers/net/bnxt/bnxt_txq.h |  1 -
 drivers/net/bnxt/bnxt_txr.c | 57 ++++++++++++++++++++++---------------
 2 files changed, 34 insertions(+), 24 deletions(-)

diff --git a/drivers/net/bnxt/bnxt_txq.h b/drivers/net/bnxt/bnxt_txq.h
index f2c712a75..720ca90cf 100644
--- a/drivers/net/bnxt/bnxt_txq.h
+++ b/drivers/net/bnxt/bnxt_txq.h
@@ -24,7 +24,6 @@ struct bnxt_tx_queue {
 	uint8_t			wthresh; /* Write-back threshold reg */
 	uint32_t		ctx_curr; /* Hardware context states */
 	uint8_t			tx_deferred_start; /* not in global dev start */
-	uint8_t			cmpl_next; /* Next BD to trigger a compl */
 
 	struct bnxt		*bp;
 	int			index;
diff --git a/drivers/net/bnxt/bnxt_txr.c b/drivers/net/bnxt/bnxt_txr.c
index 9684fb177..186934136 100644
--- a/drivers/net/bnxt/bnxt_txr.c
+++ b/drivers/net/bnxt/bnxt_txr.c
@@ -103,26 +103,33 @@ int bnxt_init_tx_ring_struct(struct bnxt_tx_queue *txq, unsigned int socket_id)
 	return 0;
 }
 
-static inline uint32_t bnxt_tx_avail(struct bnxt_tx_ring_info *txr)
+static inline uint32_t bnxt_tx_bds_in_hw(struct bnxt_tx_queue *txq)
+{
+	return ((txq->tx_ring->tx_prod - txq->tx_ring->tx_cons) &
+		txq->tx_ring->tx_ring_struct->ring_mask);
+}
+
+static inline uint32_t bnxt_tx_avail(struct bnxt_tx_queue *txq)
 {
 	/* Tell compiler to fetch tx indices from memory. */
 	rte_compiler_barrier();
 
-	return txr->tx_ring_struct->ring_size -
-		((txr->tx_prod - txr->tx_cons) &
-			txr->tx_ring_struct->ring_mask) - 1;
+	return ((txq->tx_ring->tx_ring_struct->ring_size -
+		 bnxt_tx_bds_in_hw(txq)) - 1);
 }
 
 static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 				struct bnxt_tx_queue *txq,
 				uint16_t *coal_pkts,
-				uint16_t *cmpl_next)
+				uint16_t *cmpl_next,
+				struct tx_bd_long **last_txbd)
 {
 	struct bnxt_tx_ring_info *txr = txq->tx_ring;
 	struct tx_bd_long *txbd;
 	struct tx_bd_long_hi *txbd1 = NULL;
 	uint32_t vlan_tag_flags, cfa_action;
 	bool long_bd = false;
+	unsigned short nr_bds = 0;
 	struct rte_mbuf *m_seg;
 	struct bnxt_sw_tx_bd *tx_buf;
 	static const uint32_t lhint_arr[4] = {
@@ -139,15 +146,14 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 				PKT_TX_TUNNEL_GENEVE))
 		long_bd = true;
 
-	tx_buf = &txr->tx_buf_ring[txr->tx_prod];
-	tx_buf->mbuf = tx_pkt;
-	tx_buf->nr_bds = long_bd + tx_pkt->nb_segs;
+	nr_bds = long_bd + tx_pkt->nb_segs;
+	if (unlikely(bnxt_tx_avail(txq) < nr_bds))
+		return -ENOMEM;
 
 	/* Check if number of Tx descriptors is above HW limit */
-	if (unlikely(tx_buf->nr_bds > BNXT_MAX_TSO_SEGS)) {
+	if (unlikely(nr_bds > BNXT_MAX_TSO_SEGS)) {
 		PMD_DRV_LOG(ERR,
-			    "Num descriptors %d exceeds HW limit\n",
-			    tx_buf->nr_bds);
+			    "Num descriptors %d exceeds HW limit\n", nr_bds);
 		return -ENOSPC;
 	}
 
@@ -170,12 +176,13 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 	/* Check non zero data_len */
 	RTE_VERIFY(tx_pkt->data_len);
 
-	if (unlikely(bnxt_tx_avail(txr) < tx_buf->nr_bds))
-		return -ENOMEM;
+	tx_buf = &txr->tx_buf_ring[txr->tx_prod];
+	tx_buf->mbuf = tx_pkt;
+	tx_buf->nr_bds = nr_bds;
 
 	txbd = &txr->tx_desc_ring[txr->tx_prod];
 	txbd->opaque = *coal_pkts;
-	txbd->flags_type = tx_buf->nr_bds << TX_BD_LONG_FLAGS_BD_CNT_SFT;
+	txbd->flags_type = nr_bds << TX_BD_LONG_FLAGS_BD_CNT_SFT;
 	txbd->flags_type |= TX_BD_SHORT_FLAGS_COAL_NOW;
 	if (!*cmpl_next) {
 		txbd->flags_type |= TX_BD_LONG_FLAGS_NO_CMPL;
@@ -189,6 +196,7 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 	else
 		txbd->flags_type |= lhint_arr[tx_pkt->pkt_len >> 9];
 	txbd->address = rte_cpu_to_le_64(rte_mbuf_data_iova(tx_buf->mbuf));
+	*last_txbd = txbd;
 
 	if (long_bd) {
 		txbd->flags_type |= TX_BD_LONG_TYPE_TX_BD_LONG;
@@ -321,7 +329,7 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 
 		txbd = &txr->tx_desc_ring[txr->tx_prod];
 		txbd->address = rte_cpu_to_le_64(rte_mbuf_data_iova(m_seg));
-		txbd->flags_type |= TX_BD_SHORT_TYPE_TX_BD_SHORT;
+		txbd->flags_type = TX_BD_SHORT_TYPE_TX_BD_SHORT;
 		txbd->len = m_seg->data_len;
 
 		m_seg = m_seg->next;
@@ -371,8 +379,7 @@ static int bnxt_handle_tx_cp(struct bnxt_tx_queue *txq)
 	uint32_t ring_mask = cp_ring_struct->ring_mask;
 	uint32_t opaque = 0;
 
-	if (((txq->tx_ring->tx_prod - txq->tx_ring->tx_cons) &
-		txq->tx_ring->tx_ring_struct->ring_mask) < txq->tx_free_thresh)
+	if (bnxt_tx_bds_in_hw(txq) < txq->tx_free_thresh)
 		return 0;
 
 	do {
@@ -411,7 +418,8 @@ uint16_t bnxt_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 	struct bnxt_tx_queue *txq = tx_queue;
 	uint16_t nb_tx_pkts = 0;
 	uint16_t coal_pkts = 0;
-	uint16_t cmpl_next = txq->cmpl_next;
+	uint16_t cmpl_next = 0;
+	struct tx_bd_long *last_txbd = NULL;
 
 	/* Handle TX completions */
 	bnxt_handle_tx_cp(txq);
@@ -422,20 +430,23 @@ uint16_t bnxt_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 		return 0;
 	}
 
-	txq->cmpl_next = 0;
 	/* Handle TX burst request */
 	for (nb_tx_pkts = 0; nb_tx_pkts < nb_pkts; nb_tx_pkts++) {
 		int rc;
 
-		/* Request a completion on first and last packet */
+		/* Request a completion on the last packet */
 		cmpl_next |= (nb_pkts == nb_tx_pkts + 1);
 		coal_pkts++;
 		rc = bnxt_start_xmit(tx_pkts[nb_tx_pkts], txq,
-				&coal_pkts, &cmpl_next);
+				     &coal_pkts, &cmpl_next, &last_txbd);
 
 		if (unlikely(rc)) {
-			/* Request a completion in next cycle */
-			txq->cmpl_next = 1;
+			/* Request a completion on the last successfully
+			 * enqueued packet
+			 */
+			if (last_txbd)
+				last_txbd->flags_type &=
+					~TX_BD_LONG_FLAGS_NO_CMPL;
 			break;
 		}
 	}
-- 
2.20.1 (Apple Git-117)

[dpdk-dev] [PATCH 5/6] net/bnxt: optimize Tx batching code

[Ajit Khaparde]

From: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>

This patch adds a few optimizations to Tx batching. Instead of
specifying 'cmpl_next' to bnxt_start_xmit() as a hint to enable
completion for a given packet, request for completion on the last
successfully enqueued TxBD. This takes care of both success and
error cases. It eliminates 'cmpl_next' arg and related per-packet
condition checks around it.

Fixes: 5735eb241947 ("net/bnxt: support Tx batching")
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Reviewed-by: Ajit Kumar Khaparde <ajit.khaparde@broadcom.com>
Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
---
 drivers/net/bnxt/bnxt_txr.c | 32 +++++++++-----------------------
 1 file changed, 9 insertions(+), 23 deletions(-)

diff --git a/drivers/net/bnxt/bnxt_txr.c b/drivers/net/bnxt/bnxt_txr.c
index 186934136..b15778b39 100644
--- a/drivers/net/bnxt/bnxt_txr.c
+++ b/drivers/net/bnxt/bnxt_txr.c
@@ -121,7 +121,6 @@ static inline uint32_t bnxt_tx_avail(struct bnxt_tx_queue *txq)
 static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 				struct bnxt_tx_queue *txq,
 				uint16_t *coal_pkts,
-				uint16_t *cmpl_next,
 				struct tx_bd_long **last_txbd)
 {
 	struct bnxt_tx_ring_info *txr = txq->tx_ring;
@@ -184,12 +183,7 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 	txbd->opaque = *coal_pkts;
 	txbd->flags_type = nr_bds << TX_BD_LONG_FLAGS_BD_CNT_SFT;
 	txbd->flags_type |= TX_BD_SHORT_FLAGS_COAL_NOW;
-	if (!*cmpl_next) {
-		txbd->flags_type |= TX_BD_LONG_FLAGS_NO_CMPL;
-	} else {
-		*coal_pkts = 0;
-		*cmpl_next = false;
-	}
+	txbd->flags_type |= TX_BD_LONG_FLAGS_NO_CMPL;
 	txbd->len = tx_pkt->data_len;
 	if (tx_pkt->pkt_len >= 2014)
 		txbd->flags_type |= TX_BD_LONG_FLAGS_LHINT_GTE2K;
@@ -415,10 +409,10 @@ static int bnxt_handle_tx_cp(struct bnxt_tx_queue *txq)
 uint16_t bnxt_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 			       uint16_t nb_pkts)
 {
-	struct bnxt_tx_queue *txq = tx_queue;
+	int rc;
 	uint16_t nb_tx_pkts = 0;
 	uint16_t coal_pkts = 0;
-	uint16_t cmpl_next = 0;
+	struct bnxt_tx_queue *txq = tx_queue;
 	struct tx_bd_long *last_txbd = NULL;
 
 	/* Handle TX completions */
@@ -432,27 +426,19 @@ uint16_t bnxt_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 
 	/* Handle TX burst request */
 	for (nb_tx_pkts = 0; nb_tx_pkts < nb_pkts; nb_tx_pkts++) {
-		int rc;
-
-		/* Request a completion on the last packet */
-		cmpl_next |= (nb_pkts == nb_tx_pkts + 1);
 		coal_pkts++;
 		rc = bnxt_start_xmit(tx_pkts[nb_tx_pkts], txq,
-				     &coal_pkts, &cmpl_next, &last_txbd);
+				     &coal_pkts, &last_txbd);
 
-		if (unlikely(rc)) {
-			/* Request a completion on the last successfully
-			 * enqueued packet
-			 */
-			if (last_txbd)
-				last_txbd->flags_type &=
-					~TX_BD_LONG_FLAGS_NO_CMPL;
+		if (unlikely(rc))
 			break;
-		}
 	}
 
-	if (nb_tx_pkts)
+	if (likely(nb_tx_pkts)) {
+		/* Request a completion on the last packet */
+		last_txbd->flags_type &= ~TX_BD_LONG_FLAGS_NO_CMPL;
 		B_TX_DB(txq->tx_ring->tx_doorbell, txq->tx_ring->tx_prod);
+	}
 
 	return nb_tx_pkts;
 }
-- 
2.20.1 (Apple Git-117)

[dpdk-dev] [PATCH 5/6] net/bnxt: optimize Tx batching code

[Ajit Khaparde]

From: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>

This patch adds a few optimizations to Tx batching. Instead of
specifying 'cmpl_next' to bnxt_start_xmit() as a hint to enable
completion for a given packet, request for completion on the last
successfully enqueued TxBD. This takes care of both success and
error cases. It eliminates 'cmpl_next' arg and related per-packet
condition checks around it.

Fixes: 5735eb241947 ("net/bnxt: support Tx batching")
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
Reviewed-by: Ajit Kumar Khaparde <ajit.khaparde@broadcom.com>
Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
---
 drivers/net/bnxt/bnxt_txr.c | 32 +++++++++-----------------------
 1 file changed, 9 insertions(+), 23 deletions(-)

diff --git a/drivers/net/bnxt/bnxt_txr.c b/drivers/net/bnxt/bnxt_txr.c
index 186934136..b15778b39 100644
--- a/drivers/net/bnxt/bnxt_txr.c
+++ b/drivers/net/bnxt/bnxt_txr.c
@@ -121,7 +121,6 @@ static inline uint32_t bnxt_tx_avail(struct bnxt_tx_queue *txq)
 static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 				struct bnxt_tx_queue *txq,
 				uint16_t *coal_pkts,
-				uint16_t *cmpl_next,
 				struct tx_bd_long **last_txbd)
 {
 	struct bnxt_tx_ring_info *txr = txq->tx_ring;
@@ -184,12 +183,7 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 	txbd->opaque = *coal_pkts;
 	txbd->flags_type = nr_bds << TX_BD_LONG_FLAGS_BD_CNT_SFT;
 	txbd->flags_type |= TX_BD_SHORT_FLAGS_COAL_NOW;
-	if (!*cmpl_next) {
-		txbd->flags_type |= TX_BD_LONG_FLAGS_NO_CMPL;
-	} else {
-		*coal_pkts = 0;
-		*cmpl_next = false;
-	}
+	txbd->flags_type |= TX_BD_LONG_FLAGS_NO_CMPL;
 	txbd->len = tx_pkt->data_len;
 	if (tx_pkt->pkt_len >= 2014)
 		txbd->flags_type |= TX_BD_LONG_FLAGS_LHINT_GTE2K;
@@ -415,10 +409,10 @@ static int bnxt_handle_tx_cp(struct bnxt_tx_queue *txq)
 uint16_t bnxt_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 			       uint16_t nb_pkts)
 {
-	struct bnxt_tx_queue *txq = tx_queue;
+	int rc;
 	uint16_t nb_tx_pkts = 0;
 	uint16_t coal_pkts = 0;
-	uint16_t cmpl_next = 0;
+	struct bnxt_tx_queue *txq = tx_queue;
 	struct tx_bd_long *last_txbd = NULL;
 
 	/* Handle TX completions */
@@ -432,27 +426,19 @@ uint16_t bnxt_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 
 	/* Handle TX burst request */
 	for (nb_tx_pkts = 0; nb_tx_pkts < nb_pkts; nb_tx_pkts++) {
-		int rc;
-
-		/* Request a completion on the last packet */
-		cmpl_next |= (nb_pkts == nb_tx_pkts + 1);
 		coal_pkts++;
 		rc = bnxt_start_xmit(tx_pkts[nb_tx_pkts], txq,
-				     &coal_pkts, &cmpl_next, &last_txbd);
+				     &coal_pkts, &last_txbd);
 
-		if (unlikely(rc)) {
-			/* Request a completion on the last successfully
-			 * enqueued packet
-			 */
-			if (last_txbd)
-				last_txbd->flags_type &=
-					~TX_BD_LONG_FLAGS_NO_CMPL;
+		if (unlikely(rc))
 			break;
-		}
 	}
 
-	if (nb_tx_pkts)
+	if (likely(nb_tx_pkts)) {
+		/* Request a completion on the last packet */
+		last_txbd->flags_type &= ~TX_BD_LONG_FLAGS_NO_CMPL;
 		B_TX_DB(txq->tx_ring->tx_doorbell, txq->tx_ring->tx_prod);
+	}
 
 	return nb_tx_pkts;
 }
-- 
2.20.1 (Apple Git-117)

[dpdk-dev] [PATCH 6/6] net/bnxt: support bulk free of Tx mbufs

[Ajit Khaparde]

The driver currently uses rte_pktmbuf_free() to free each mbuf
after transmit completion. This is optimized to free multiple
mbufs using rte_mempool_put_bulk().

Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
---
 drivers/net/bnxt/bnxt_txq.c | 11 ++++++++
 drivers/net/bnxt/bnxt_txq.h |  1 +
 drivers/net/bnxt/bnxt_txr.c | 50 +++++++++++++++++++++++++++++--------
 3 files changed, 52 insertions(+), 10 deletions(-)

diff --git a/drivers/net/bnxt/bnxt_txq.c b/drivers/net/bnxt/bnxt_txq.c
index b9b975e4c..5a7bfaf3e 100644
--- a/drivers/net/bnxt/bnxt_txq.c
+++ b/drivers/net/bnxt/bnxt_txq.c
@@ -69,6 +69,7 @@ void bnxt_tx_queue_release_op(void *tx_queue)
 		rte_memzone_free(txq->mz);
 		txq->mz = NULL;
 
+		rte_free(txq->free);
 		rte_free(txq);
 	}
 }
@@ -110,6 +111,16 @@ int bnxt_tx_queue_setup_op(struct rte_eth_dev *eth_dev,
 		rc = -ENOMEM;
 		goto out;
 	}
+
+	txq->free = rte_zmalloc_socket(NULL,
+				       sizeof(struct rte_mbuf *) * nb_desc,
+				       RTE_CACHE_LINE_SIZE, socket_id);
+	if (!txq->free) {
+		PMD_DRV_LOG(ERR, "allocation of tx mbuf free array failed!");
+		rte_free(txq);
+		rc = -ENOMEM;
+		goto out;
+	}
 	txq->bp = bp;
 	txq->nb_tx_desc = nb_desc;
 	txq->tx_free_thresh = tx_conf->tx_free_thresh;
diff --git a/drivers/net/bnxt/bnxt_txq.h b/drivers/net/bnxt/bnxt_txq.h
index 720ca90cf..a0d4678d9 100644
--- a/drivers/net/bnxt/bnxt_txq.h
+++ b/drivers/net/bnxt/bnxt_txq.h
@@ -33,6 +33,7 @@ struct bnxt_tx_queue {
 	unsigned int		cp_nr_rings;
 	struct bnxt_cp_ring_info	*cp_ring;
 	const struct rte_memzone *mz;
+	struct rte_mbuf **free;
 };
 
 void bnxt_free_txq_stats(struct bnxt_tx_queue *txq);
diff --git a/drivers/net/bnxt/bnxt_txr.c b/drivers/net/bnxt/bnxt_txr.c
index b15778b39..9de12e0d0 100644
--- a/drivers/net/bnxt/bnxt_txr.c
+++ b/drivers/net/bnxt/bnxt_txr.c
@@ -320,6 +320,7 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 		RTE_VERIFY(m_seg->data_len);
 		txr->tx_prod = RING_NEXT(txr->tx_ring_struct, txr->tx_prod);
 		tx_buf = &txr->tx_buf_ring[txr->tx_prod];
+		tx_buf->mbuf = m_seg;
 
 		txbd = &txr->tx_desc_ring[txr->tx_prod];
 		txbd->address = rte_cpu_to_le_64(rte_mbuf_data_iova(m_seg));
@@ -339,24 +340,53 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 static void bnxt_tx_cmp(struct bnxt_tx_queue *txq, int nr_pkts)
 {
 	struct bnxt_tx_ring_info *txr = txq->tx_ring;
+	struct rte_mempool *pool = NULL;
+	struct rte_mbuf **free = txq->free;
 	uint16_t cons = txr->tx_cons;
+	unsigned int blk = 0;
 	int i, j;
 
 	for (i = 0; i < nr_pkts; i++) {
-		struct bnxt_sw_tx_bd *tx_buf;
 		struct rte_mbuf *mbuf;
+		struct bnxt_sw_tx_bd *tx_buf = &txr->tx_buf_ring[cons];
+		unsigned short nr_bds = tx_buf->nr_bds;
 
-		tx_buf = &txr->tx_buf_ring[cons];
-		cons = RING_NEXT(txr->tx_ring_struct, cons);
-		mbuf = tx_buf->mbuf;
-		tx_buf->mbuf = NULL;
-
-		/* EW - no need to unmap DMA memory? */
-
-		for (j = 1; j < tx_buf->nr_bds; j++)
+		for (j = 0; j < nr_bds; j++) {
+			mbuf = tx_buf->mbuf;
+			tx_buf->mbuf = NULL;
 			cons = RING_NEXT(txr->tx_ring_struct, cons);
-		rte_pktmbuf_free(mbuf);
+			tx_buf = &txr->tx_buf_ring[cons];
+			if (!mbuf)	/* long_bd's tx_buf ? */
+				continue;
+
+			mbuf = rte_pktmbuf_prefree_seg(mbuf);
+			if (unlikely(!mbuf))
+				continue;
+
+			/* EW - no need to unmap DMA memory? */
+
+			if (likely(mbuf->pool == pool)) {
+				/* Add mbuf to the bulk free array */
+				free[blk++] = mbuf;
+			} else {
+				/* Found an mbuf from a different pool. Free
+				 * mbufs accumulated so far to the previous
+				 * pool
+				 */
+				if (likely(pool != NULL))
+					rte_mempool_put_bulk(pool,
+							     (void *)free,
+							     blk);
+
+				/* Start accumulating mbufs in a new pool */
+				free[0] = mbuf;
+				pool = mbuf->pool;
+				blk = 1;
+			}
+		}
 	}
+	if (blk)
+		rte_mempool_put_bulk(pool, (void *)free, blk);
 
 	txr->tx_cons = cons;
 }
-- 
2.20.1 (Apple Git-117)

[dpdk-dev] [PATCH 6/6] net/bnxt: support bulk free of Tx mbufs

[Ajit Khaparde]

The driver currently uses rte_pktmbuf_free() to free each mbuf
after transmit completion. This is optimized to free multiple
mbufs using rte_mempool_put_bulk().

Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
Signed-off-by: Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>
---
 drivers/net/bnxt/bnxt_txq.c | 11 ++++++++
 drivers/net/bnxt/bnxt_txq.h |  1 +
 drivers/net/bnxt/bnxt_txr.c | 50 +++++++++++++++++++++++++++++--------
 3 files changed, 52 insertions(+), 10 deletions(-)

diff --git a/drivers/net/bnxt/bnxt_txq.c b/drivers/net/bnxt/bnxt_txq.c
index b9b975e4c..5a7bfaf3e 100644
--- a/drivers/net/bnxt/bnxt_txq.c
+++ b/drivers/net/bnxt/bnxt_txq.c
@@ -69,6 +69,7 @@ void bnxt_tx_queue_release_op(void *tx_queue)
 		rte_memzone_free(txq->mz);
 		txq->mz = NULL;
 
+		rte_free(txq->free);
 		rte_free(txq);
 	}
 }
@@ -110,6 +111,16 @@ int bnxt_tx_queue_setup_op(struct rte_eth_dev *eth_dev,
 		rc = -ENOMEM;
 		goto out;
 	}
+
+	txq->free = rte_zmalloc_socket(NULL,
+				       sizeof(struct rte_mbuf *) * nb_desc,
+				       RTE_CACHE_LINE_SIZE, socket_id);
+	if (!txq->free) {
+		PMD_DRV_LOG(ERR, "allocation of tx mbuf free array failed!");
+		rte_free(txq);
+		rc = -ENOMEM;
+		goto out;
+	}
 	txq->bp = bp;
 	txq->nb_tx_desc = nb_desc;
 	txq->tx_free_thresh = tx_conf->tx_free_thresh;
diff --git a/drivers/net/bnxt/bnxt_txq.h b/drivers/net/bnxt/bnxt_txq.h
index 720ca90cf..a0d4678d9 100644
--- a/drivers/net/bnxt/bnxt_txq.h
+++ b/drivers/net/bnxt/bnxt_txq.h
@@ -33,6 +33,7 @@ struct bnxt_tx_queue {
 	unsigned int		cp_nr_rings;
 	struct bnxt_cp_ring_info	*cp_ring;
 	const struct rte_memzone *mz;
+	struct rte_mbuf **free;
 };
 
 void bnxt_free_txq_stats(struct bnxt_tx_queue *txq);
diff --git a/drivers/net/bnxt/bnxt_txr.c b/drivers/net/bnxt/bnxt_txr.c
index b15778b39..9de12e0d0 100644
--- a/drivers/net/bnxt/bnxt_txr.c
+++ b/drivers/net/bnxt/bnxt_txr.c
@@ -320,6 +320,7 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 		RTE_VERIFY(m_seg->data_len);
 		txr->tx_prod = RING_NEXT(txr->tx_ring_struct, txr->tx_prod);
 		tx_buf = &txr->tx_buf_ring[txr->tx_prod];
+		tx_buf->mbuf = m_seg;
 
 		txbd = &txr->tx_desc_ring[txr->tx_prod];
 		txbd->address = rte_cpu_to_le_64(rte_mbuf_data_iova(m_seg));
@@ -339,24 +340,53 @@ static uint16_t bnxt_start_xmit(struct rte_mbuf *tx_pkt,
 static void bnxt_tx_cmp(struct bnxt_tx_queue *txq, int nr_pkts)
 {
 	struct bnxt_tx_ring_info *txr = txq->tx_ring;
+	struct rte_mempool *pool = NULL;
+	struct rte_mbuf **free = txq->free;
 	uint16_t cons = txr->tx_cons;
+	unsigned int blk = 0;
 	int i, j;
 
 	for (i = 0; i < nr_pkts; i++) {
-		struct bnxt_sw_tx_bd *tx_buf;
 		struct rte_mbuf *mbuf;
+		struct bnxt_sw_tx_bd *tx_buf = &txr->tx_buf_ring[cons];
+		unsigned short nr_bds = tx_buf->nr_bds;
 
-		tx_buf = &txr->tx_buf_ring[cons];
-		cons = RING_NEXT(txr->tx_ring_struct, cons);
-		mbuf = tx_buf->mbuf;
-		tx_buf->mbuf = NULL;
-
-		/* EW - no need to unmap DMA memory? */
-
-		for (j = 1; j < tx_buf->nr_bds; j++)
+		for (j = 0; j < nr_bds; j++) {
+			mbuf = tx_buf->mbuf;
+			tx_buf->mbuf = NULL;
 			cons = RING_NEXT(txr->tx_ring_struct, cons);
-		rte_pktmbuf_free(mbuf);
+			tx_buf = &txr->tx_buf_ring[cons];
+			if (!mbuf)	/* long_bd's tx_buf ? */
+				continue;
+
+			mbuf = rte_pktmbuf_prefree_seg(mbuf);
+			if (unlikely(!mbuf))
+				continue;
+
+			/* EW - no need to unmap DMA memory? */
+
+			if (likely(mbuf->pool == pool)) {
+				/* Add mbuf to the bulk free array */
+				free[blk++] = mbuf;
+			} else {
+				/* Found an mbuf from a different pool. Free
+				 * mbufs accumulated so far to the previous
+				 * pool
+				 */
+				if (likely(pool != NULL))
+					rte_mempool_put_bulk(pool,
+							     (void *)free,
+							     blk);
+
+				/* Start accumulating mbufs in a new pool */
+				free[0] = mbuf;
+				pool = mbuf->pool;
+				blk = 1;
+			}
+		}
 	}
+	if (blk)
+		rte_mempool_put_bulk(pool, (void *)free, blk);
 
 	txr->tx_cons = cons;
 }
-- 
2.20.1 (Apple Git-117)

Re: [dpdk-dev] [PATCH 0/6] bnxt patchset for Tx performance optimization

[Ferruh Yigit]

On 5/15/2019 7:08 PM, Ajit Khaparde wrote:
> This patchset brings in few fixes and some changes tuning
> the Tx datapath performance in bnxt PMD.
> Please apply.
> 
> Ajit Khaparde (4):
>   net/bnxt: fix an issue seen with TSO
>   net/bnxt: add support to update ipid
>   net/bnxt: check for some error conditions in Tx path
>   net/bnxt: support bulk free of Tx mbufs
> 
> Sriharsha Basavapatna (2):
>   net/bnxt: fix a couple of issues with Tx batching
>   net/bnxt: optimize Tx batching code

Series applied to dpdk-next-net/master, thanks.