[PATCH] net/mana: implement RX CQE coalescing

[PATCH] net/mana: implement RX CQE coalescing

[longli]

From: Long Li <longli@microsoft.com>

With RX coalescing, one CQE entry can be used to indicate up to 4 packets
on the receive queue. This saves processing time and PCI bandwidth over
the CQ.

Cc: stable@dpdk.org
Signed-off-by: Long Li <longli@microsoft.com>
---
 drivers/net/mana/mana.h |  4 +++
 drivers/net/mana/rx.c   | 71 ++++++++++++++++++++++++++++++++++-------
 2 files changed, 64 insertions(+), 11 deletions(-)

diff --git a/drivers/net/mana/mana.h b/drivers/net/mana/mana.h
index be88537f1a..dfeda531eb 100644
--- a/drivers/net/mana/mana.h
+++ b/drivers/net/mana/mana.h
@@ -429,6 +429,10 @@ struct mana_rxq {
 	struct mana_gdma_queue gdma_cq;
 	struct gdma_comp *gdma_comp_buf;
 
+	uint32_t comp_buf_len;
+	uint32_t comp_buf_idx;
+	uint32_t backlog_idx;
+
 	struct mana_stats stats;
 	struct mana_mr_btree mr_btree;
 
diff --git a/drivers/net/mana/rx.c b/drivers/net/mana/rx.c
index 6e1c397be8..cacfd9ae1b 100644
--- a/drivers/net/mana/rx.c
+++ b/drivers/net/mana/rx.c
@@ -361,6 +361,10 @@ mana_start_rx_queues(struct rte_eth_dev *dev)
 		DRV_LOG(INFO, "rxq rq id %u buf %p count %u size %u",
 			rxq->gdma_rq.id, rxq->gdma_rq.buffer,
 			rxq->gdma_rq.count, rxq->gdma_rq.size);
+
+		rxq->comp_buf_len = 0;
+		rxq->comp_buf_idx = 0;
+		rxq->backlog_idx = 0;
 	}
 
 	for (i = 0; i < priv->num_queues; i++) {
@@ -385,16 +389,29 @@ mana_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 	struct mana_priv *priv = rxq->priv;
 	struct rte_mbuf *mbuf;
 	int ret;
-	uint32_t num_pkts;
+	uint32_t pkt_idx = rxq->backlog_idx;
+	uint32_t pkt_len;
+	uint32_t i;
+	int polled = 0;
+
+repoll:
+	/* Polling on new completions if we have no backlog */
+	if (rxq->comp_buf_idx == rxq->comp_buf_len) {
+		RTE_ASSERT(!pkt_idx);
+		rxq->comp_buf_len =
+			gdma_poll_completion_queue(&rxq->gdma_cq,
+						   rxq->gdma_comp_buf, pkts_n);
+		rxq->comp_buf_idx = 0;
+		polled = 1;
+	}
 
-	num_pkts = gdma_poll_completion_queue(&rxq->gdma_cq, rxq->gdma_comp_buf, pkts_n);
-	for (uint32_t i = 0; i < num_pkts; i++) {
+	i = rxq->comp_buf_idx;
+	while (i < rxq->comp_buf_len) {
 		struct mana_rx_comp_oob *oob = (struct mana_rx_comp_oob *)
 			rxq->gdma_comp_buf[i].cqe_data;
 		struct mana_rxq_desc *desc =
 			&rxq->desc_ring[rxq->desc_ring_tail];
 
-		rxq->gdma_rq.tail += desc->wqe_size_in_bu;
 		mbuf = desc->pkt;
 
 		switch (oob->cqe_hdr.cqe_type) {
@@ -409,8 +426,8 @@ mana_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 			goto drop;
 
 		case CQE_RX_COALESCED_4:
-			DP_LOG(ERR, "RX coalescing is not supported");
-			continue;
+			/* Proceed to process mbuf */
+			break;
 
 		default:
 			DP_LOG(ERR, "Unknown RX CQE type %d",
@@ -418,13 +435,22 @@ mana_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 			continue;
 		}
 
-		DP_LOG(DEBUG, "mana_rx_comp_oob CQE_RX_OKAY rxq %p", rxq);
+		DP_LOG(DEBUG, "mana_rx_comp_oob type %d rxq %p",
+		       oob->cqe_hdr.cqe_type, rxq);
+
+		pkt_len = oob->packet_info[pkt_idx].packet_length;
+		if (!pkt_len) {
+			/* Move on to the next completion */
+			pkt_idx = 0;
+			i++;
+			continue;
+		}
 
 		mbuf->data_off = RTE_PKTMBUF_HEADROOM;
 		mbuf->nb_segs = 1;
 		mbuf->next = NULL;
-		mbuf->pkt_len = oob->packet_info[0].packet_length;
-		mbuf->data_len = oob->packet_info[0].packet_length;
+		mbuf->data_len = pkt_len;
+		mbuf->pkt_len = pkt_len;
 		mbuf->port = priv->port_id;
 
 		if (oob->rx_ip_header_checksum_succeeded)
@@ -447,19 +473,28 @@ mana_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 		if (oob->rx_hash_type == MANA_HASH_L3 ||
 		    oob->rx_hash_type == MANA_HASH_L4) {
 			mbuf->ol_flags |= RTE_MBUF_F_RX_RSS_HASH;
-			mbuf->hash.rss = oob->packet_info[0].packet_hash;
+			mbuf->hash.rss = oob->packet_info[pkt_idx].packet_hash;
 		}
 
 		pkts[pkt_received++] = mbuf;
 		rxq->stats.packets++;
 		rxq->stats.bytes += mbuf->data_len;
 
+		pkt_idx++;
+		/* Move on the next completion if all packets are processed */
+		if (pkt_idx >= RX_COM_OOB_NUM_PACKETINFO_SEGMENTS) {
+			pkt_idx = 0;
+			i++;
+		}
+
 drop:
 		rxq->desc_ring_tail++;
 		if (rxq->desc_ring_tail >= rxq->num_desc)
 			rxq->desc_ring_tail = 0;
 
-		/* Post another request */
+		rxq->gdma_rq.tail += desc->wqe_size_in_bu;
+
+		/* Consume this request and post another request */
 		ret = mana_alloc_and_post_rx_wqe(rxq);
 		if (ret) {
 			DP_LOG(ERR, "failed to post rx wqe ret=%d", ret);
@@ -467,6 +502,20 @@ mana_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 		}
 
 		wqe_posted++;
+		if (pkt_received == pkts_n)
+			break;
+	}
+
+	rxq->backlog_idx = pkt_idx;
+	rxq->comp_buf_idx = i;
+
+	/* If all CQEs are processed but there are more packets to read, poll the
+	 * completion queue again because we may have not polled on the completion
+	 * queue due to CQE not fully processed in the previous rx_burst
+	 */
+	if (pkt_received < pkts_n && !polled) {
+		polled = 1;
+		goto repoll;
 	}
 
 	if (wqe_posted)
-- 
2.34.1

Re: [PATCH] net/mana: implement RX CQE coalescing

[Ferruh Yigit]

On 5/6/2023 2:33 AM, longli@linuxonhyperv.com wrote:
> From: Long Li <longli@microsoft.com>
> 
> With RX coalescing, one CQE entry can be used to indicate up to 4 packets
> on the receive queue. This saves processing time and PCI bandwidth over
> the CQ.
> 
> Signed-off-by: Long Li <longli@microsoft.com>
>

Applied to dpdk-next-net/main, thanks.

Dropped stable tag in next-net, can add back again based on discussion
with LTS maintainers.

Re: [PATCH] net/mana: implement RX CQE coalescing

[Ferruh Yigit]

On 5/6/2023 2:33 AM, longli@linuxonhyperv.com wrote:
> From: Long Li <longli@microsoft.com>
> 
> With RX coalescing, one CQE entry can be used to indicate up to 4 packets
> on the receive queue. This saves processing time and PCI bandwidth over
> the CQ.
> 
> Cc: stable@dpdk.org
> Signed-off-by: Long Li <longli@microsoft.com>
>

Backport to stable tree is requested but this is not a fix patch, cc'ed
stable tree maintainers for comment.

stable tag removed while merging to next-net for now.