DPDK patches and discussions
 help / color / mirror / Atom feed
From: Ajit Khaparde <ajit.khaparde@broadcom.com>
To: dev@dpdk.org
Cc: ferruh.yigit@intel.com,
	Sriharsha Basavapatna <sriharsha.basavapatna@broadcom.com>,
	Ajit Kumar Khaparde <ajit.khaparde@broadcom.com>
Subject: [dpdk-dev] [PATCH 4/6] net/bnxt: fix a couple of issues with Tx batching
Date: Wed, 15 May 2019 11:08:15 -0700	[thread overview]
Message-ID: <20190515180817.71523-5-ajit.khaparde@broadcom.com> (raw)
In-Reply-To: <20190515180817.71523-1-ajit.khaparde@broadcom.com>

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)

  parent reply	other threads:[~2019-05-15 18:08 UTC|newest]

Thread overview: 15+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2019-05-15 18:08 [dpdk-dev] [PATCH 0/6] bnxt patchset for Tx performance optimization Ajit Khaparde
2019-05-15 18:08 ` Ajit Khaparde
2019-05-15 18:08 ` [dpdk-dev] [PATCH 1/6] net/bnxt: fix an issue seen with TSO Ajit Khaparde
2019-05-15 18:08   ` Ajit Khaparde
2019-05-15 18:08 ` [dpdk-dev] [PATCH 2/6] net/bnxt: add support to update ipid Ajit Khaparde
2019-05-15 18:08   ` Ajit Khaparde
2019-05-15 18:08 ` [dpdk-dev] [PATCH 3/6] net/bnxt: check for some error conditions in Tx path Ajit Khaparde
2019-05-15 18:08   ` Ajit Khaparde
2019-05-15 18:08 ` Ajit Khaparde [this message]
2019-05-15 18:08   ` [dpdk-dev] [PATCH 4/6] net/bnxt: fix a couple of issues with Tx batching Ajit Khaparde
2019-05-15 18:08 ` [dpdk-dev] [PATCH 5/6] net/bnxt: optimize Tx batching code Ajit Khaparde
2019-05-15 18:08   ` Ajit Khaparde
2019-05-15 18:08 ` [dpdk-dev] [PATCH 6/6] net/bnxt: support bulk free of Tx mbufs Ajit Khaparde
2019-05-15 18:08   ` Ajit Khaparde
2019-05-27 16:16 ` [dpdk-dev] [PATCH 0/6] bnxt patchset for Tx performance optimization Ferruh Yigit

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20190515180817.71523-5-ajit.khaparde@broadcom.com \
    --to=ajit.khaparde@broadcom.com \
    --cc=dev@dpdk.org \
    --cc=ferruh.yigit@intel.com \
    --cc=sriharsha.basavapatna@broadcom.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).