patches for DPDK stable branches
 help / color / mirror / Atom feed
* [PATCH] net/mana: suppress TX CQE generation whenever possible
@ 2023-04-28  3:36 longli
  2023-05-06  1:32 ` [PATCH v2] " longli
  0 siblings, 1 reply; 7+ messages in thread
From: longli @ 2023-04-28  3:36 UTC (permalink / raw)
  To: Ferruh Yigit; +Cc: dev, Ajay Sharma, Long Li, stable

From: Long Li <longli@microsoft.com>

When sending TX packets, we don't need a completion for every packet sent.
If packets are sent in a series, the completion of the last packet can be
used to indicate completion of all prior packets.

Signed-off-by: Long Li <longli@microsoft.com>
Cc: stable@dpdk.org
---
 drivers/net/mana/mana.h |  3 ++-
 drivers/net/mana/tx.c   | 33 ++++++++++++++++++++++++++++++---
 2 files changed, 32 insertions(+), 4 deletions(-)

diff --git a/drivers/net/mana/mana.h b/drivers/net/mana/mana.h
index c79d39daa2..f280d66f6e 100644
--- a/drivers/net/mana/mana.h
+++ b/drivers/net/mana/mana.h
@@ -353,6 +353,7 @@ struct mana_priv {
 struct mana_txq_desc {
 	struct rte_mbuf *pkt;
 	uint32_t wqe_size_in_bu;
+	bool suppress_tx_cqe;
 };
 
 struct mana_rxq_desc {
@@ -401,7 +402,7 @@ struct mana_txq {
 	/* desc_ring_head is where we put pending requests to ring,
 	 * completion pull off desc_ring_tail
 	 */
-	uint32_t desc_ring_head, desc_ring_tail;
+	uint32_t desc_ring_head, desc_ring_tail, desc_ring_len;
 
 	struct mana_mr_btree mr_btree;
 	struct mana_stats stats;
diff --git a/drivers/net/mana/tx.c b/drivers/net/mana/tx.c
index ee0319c71d..c8d3911f85 100644
--- a/drivers/net/mana/tx.c
+++ b/drivers/net/mana/tx.c
@@ -43,9 +43,11 @@ mana_stop_tx_queues(struct rte_eth_dev *dev)
 
 			txq->desc_ring_tail =
 				(txq->desc_ring_tail + 1) % txq->num_desc;
+			txq->desc_ring_len--;
 		}
 		txq->desc_ring_head = 0;
 		txq->desc_ring_tail = 0;
+		txq->desc_ring_len = 0;
 
 		memset(&txq->gdma_sq, 0, sizeof(txq->gdma_sq));
 		memset(&txq->gdma_cq, 0, sizeof(txq->gdma_cq));
@@ -173,13 +175,14 @@ mana_tx_burst(void *dpdk_txq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 	int ret;
 	void *db_page;
 	uint16_t pkt_sent = 0;
-	uint32_t num_comp;
+	uint32_t num_comp, i;
 
 	/* Process send completions from GDMA */
 	num_comp = gdma_poll_completion_queue(&txq->gdma_cq,
 			txq->gdma_comp_buf, txq->num_desc);
 
-	for (uint32_t i = 0; i < num_comp; i++) {
+	i = 0;
+	while (i < num_comp) {
 		struct mana_txq_desc *desc =
 			&txq->desc_ring[txq->desc_ring_tail];
 		struct mana_tx_comp_oob *oob = (struct mana_tx_comp_oob *)
@@ -204,7 +207,16 @@ mana_tx_burst(void *dpdk_txq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 
 		desc->pkt = NULL;
 		txq->desc_ring_tail = (txq->desc_ring_tail + 1) % txq->num_desc;
+		txq->desc_ring_len--;
 		txq->gdma_sq.tail += desc->wqe_size_in_bu;
+
+		/* If TX CQE suppression is used, don't read more CQE but move
+		 * on to the next packet
+		 */
+		if (desc->suppress_tx_cqe)
+			continue;
+
+		i++;
 	}
 
 	/* Post send requests to GDMA */
@@ -215,6 +227,9 @@ mana_tx_burst(void *dpdk_txq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 		struct one_sgl sgl;
 		uint16_t seg_idx;
 
+		if (txq->desc_ring_len >= txq->num_desc)
+			break;
+
 		/* Drop the packet if it exceeds max segments */
 		if (m_pkt->nb_segs > priv->max_send_sge) {
 			DRV_LOG(ERR, "send packet segments %d exceeding max",
@@ -310,7 +325,6 @@ mana_tx_burst(void *dpdk_txq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 			tx_oob.short_oob.tx_compute_UDP_checksum = 0;
 		}
 
-		tx_oob.short_oob.suppress_tx_CQE_generation = 0;
 		tx_oob.short_oob.VCQ_number = txq->gdma_cq.id;
 
 		tx_oob.short_oob.VSQ_frame_num =
@@ -362,6 +376,16 @@ mana_tx_burst(void *dpdk_txq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 		if (seg_idx != m_pkt->nb_segs)
 			continue;
 
+		/* If we can at least queue post two WQEs and there are at
+		 * least two packets to send, use TX CQE suppression for the
+		 * current WQE
+		 */
+		if (txq->desc_ring_len + 1 < txq->num_desc &&
+		    pkt_idx + 1 < nb_pkts)
+			tx_oob.short_oob.suppress_tx_CQE_generation = 1;
+		else
+			tx_oob.short_oob.suppress_tx_CQE_generation = 0;
+
 		struct gdma_work_request work_req;
 		uint32_t wqe_size_in_bu;
 
@@ -384,8 +408,11 @@ mana_tx_burst(void *dpdk_txq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 			/* Update queue for tracking pending requests */
 			desc->pkt = m_pkt;
 			desc->wqe_size_in_bu = wqe_size_in_bu;
+			desc->suppress_tx_cqe =
+				tx_oob.short_oob.suppress_tx_CQE_generation;
 			txq->desc_ring_head =
 				(txq->desc_ring_head + 1) % txq->num_desc;
+			txq->desc_ring_len++;
 
 			pkt_sent++;
 
-- 
2.32.0


^ permalink raw reply	[flat|nested] 7+ messages in thread

end of thread, other threads:[~2023-05-29  1:41 UTC | newest]

Thread overview: 7+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-04-28  3:36 [PATCH] net/mana: suppress TX CQE generation whenever possible longli
2023-05-06  1:32 ` [PATCH v2] " longli
2023-05-19 16:34   ` Ferruh Yigit
2023-05-19 16:40   ` Ferruh Yigit
2023-05-26  0:19   ` EXTERNAL] [PATCH ] Add checks for the port capabilties Ajay Sharma
2023-05-26  3:30     ` Stephen Hemminger
2023-05-29  1:41       ` [EXTERNAL] " Ajay Sharma

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).