[dpdk-stable] patch 'net/qede: fix performance bottleneck in Rx path' has been queued to LTS release 18.11.1

[Kevin Traynor <ktraynor@redhat.com>]

Hi,

FYI, your patch has been queued to LTS release 18.11.1

Note it hasn't been pushed to http://dpdk.org/browse/dpdk-stable yet.
It will be pushed if I get no objections before 02/14/19. So please
shout if anyone has objections.

Also note that after the patch there's a diff of the upstream commit vs the
patch applied to the branch. This will indicate if there was any rebasing
needed to apply to the stable branch. If there were code changes for rebasing
(ie: not only metadata diffs), please double check that the rebase was
correctly done.

Thanks.

Kevin Traynor

---
>From 7e22a66545262ffcbd07ff6a266614dfeff643e9 Mon Sep 17 00:00:00 2001
From: Shahed Shaikh <shshaikh@marvell.com>
Date: Fri, 18 Jan 2019 02:29:29 -0800
Subject: [PATCH] net/qede: fix performance bottleneck in Rx path

[ upstream commit 8f2312474529ad7ff0e4b65b82efc8530e7484ce ]

Allocating replacement buffer per received packet is expensive.
Instead, process received packets first and allocate
replacement buffers in bulk later.

This improves performance by ~25% in terms of PPS on AMD
platforms.

Fixes: 2ea6f76aff40 ("qede: add core driver")

Signed-off-by: Shahed Shaikh <shshaikh@marvell.com>
Acked-by: Rasesh Mody <rmody@marvell.com>
---
 drivers/net/qede/qede_rxtx.c | 97 +++++++++++++++++++++++++++---------
 drivers/net/qede/qede_rxtx.h |  2 +
 2 files changed, 75 insertions(+), 24 deletions(-)

diff --git a/drivers/net/qede/qede_rxtx.c b/drivers/net/qede/qede_rxtx.c
index 0e33be1a3..684c4aeef 100644
--- a/drivers/net/qede/qede_rxtx.c
+++ b/drivers/net/qede/qede_rxtx.c
@@ -36,4 +36,50 @@ static inline int qede_alloc_rx_buffer(struct qede_rx_queue *rxq)
 }
 
+#define QEDE_MAX_BULK_ALLOC_COUNT 512
+
+static inline int qede_alloc_rx_bulk_mbufs(struct qede_rx_queue *rxq, int count)
+{
+	void *obj_p[QEDE_MAX_BULK_ALLOC_COUNT] __rte_cache_aligned;
+	struct rte_mbuf *mbuf = NULL;
+	struct eth_rx_bd *rx_bd;
+	dma_addr_t mapping;
+	int i, ret = 0;
+	uint16_t idx;
+
+	idx = rxq->sw_rx_prod & NUM_RX_BDS(rxq);
+
+	if (count > QEDE_MAX_BULK_ALLOC_COUNT)
+		count = QEDE_MAX_BULK_ALLOC_COUNT;
+
+	ret = rte_mempool_get_bulk(rxq->mb_pool, obj_p, count);
+	if (unlikely(ret)) {
+		PMD_RX_LOG(ERR, rxq,
+			   "Failed to allocate %d rx buffers "
+			    "sw_rx_prod %u sw_rx_cons %u mp entries %u free %u",
+			    count, idx, rxq->sw_rx_cons & NUM_RX_BDS(rxq),
+			    rte_mempool_avail_count(rxq->mb_pool),
+			    rte_mempool_in_use_count(rxq->mb_pool));
+		return -ENOMEM;
+	}
+
+	for (i = 0; i < count; i++) {
+		mbuf = obj_p[i];
+		if (likely(i < count - 1))
+			rte_prefetch0(obj_p[i + 1]);
+
+		idx = rxq->sw_rx_prod & NUM_RX_BDS(rxq);
+		rxq->sw_rx_ring[idx].mbuf = mbuf;
+		rxq->sw_rx_ring[idx].page_offset = 0;
+		mapping = rte_mbuf_data_iova_default(mbuf);
+		rx_bd = (struct eth_rx_bd *)
+			ecore_chain_produce(&rxq->rx_bd_ring);
+		rx_bd->addr.hi = rte_cpu_to_le_32(U64_HI(mapping));
+		rx_bd->addr.lo = rte_cpu_to_le_32(U64_LO(mapping));
+		rxq->sw_rx_prod++;
+	}
+
+	return 0;
+}
+
 /* Criterias for calculating Rx buffer size -
  * 1) rx_buf_size should not exceed the size of mbuf
@@ -1132,5 +1178,5 @@ qede_reuse_page(__rte_unused struct qede_dev *qdev,
 {
 	struct eth_rx_bd *rx_bd_prod = ecore_chain_produce(&rxq->rx_bd_ring);
-	uint16_t idx = rxq->sw_rx_cons & NUM_RX_BDS(rxq);
+	uint16_t idx = rxq->sw_rx_prod & NUM_RX_BDS(rxq);
 	struct qede_rx_entry *curr_prod;
 	dma_addr_t new_mapping;
@@ -1365,5 +1411,4 @@ qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 #endif
 	uint8_t tunn_parse_flag;
-	uint8_t j;
 	struct eth_fast_path_rx_tpa_start_cqe *cqe_start_tpa;
 	uint64_t ol_flags;
@@ -1374,4 +1419,5 @@ qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 	struct qede_agg_info *tpa_info = NULL;
 	uint32_t rss_hash;
+	int rx_alloc_count = 0;
 
 	hw_comp_cons = rte_le_to_cpu_16(*rxq->hw_cons_ptr);
@@ -1383,4 +1429,23 @@ qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 		return 0;
 
+	/* Allocate buffers that we used in previous loop */
+	if (rxq->rx_alloc_count) {
+		if (unlikely(qede_alloc_rx_bulk_mbufs(rxq,
+			     rxq->rx_alloc_count))) {
+			struct rte_eth_dev *dev;
+
+			PMD_RX_LOG(ERR, rxq,
+				   "New buffer allocation failed,"
+				   "dropping incoming packetn");
+			dev = &rte_eth_devices[rxq->port_id];
+			dev->data->rx_mbuf_alloc_failed +=
+							rxq->rx_alloc_count;
+			rxq->rx_alloc_errors += rxq->rx_alloc_count;
+			return 0;
+		}
+		qede_update_rx_prod(qdev, rxq);
+		rxq->rx_alloc_count = 0;
+	}
+
 	while (sw_comp_cons != hw_comp_cons) {
 		ol_flags = 0;
@@ -1554,14 +1619,5 @@ qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 		}
 
-		if (unlikely(qede_alloc_rx_buffer(rxq) != 0)) {
-			PMD_RX_LOG(ERR, rxq,
-				   "New buffer allocation failed,"
-				   "dropping incoming packet\n");
-			qede_recycle_rx_bd_ring(rxq, qdev, fp_cqe->bd_num);
-			rte_eth_devices[rxq->port_id].
-			    data->rx_mbuf_alloc_failed++;
-			rxq->rx_alloc_errors++;
-			break;
-		}
+		rx_alloc_count++;
 		qede_rx_bd_ring_consume(rxq);
 
@@ -1575,15 +1631,7 @@ qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 						 pkt_len - len))
 				goto next_cqe;
-			for (j = 0; j < num_segs; j++) {
-				if (qede_alloc_rx_buffer(rxq)) {
-					PMD_RX_LOG(ERR, rxq,
-						"Buffer allocation failed");
-					rte_eth_devices[rxq->port_id].
-						data->rx_mbuf_alloc_failed++;
-					rxq->rx_alloc_errors++;
-					break;
-				}
-				rxq->rx_segs++;
-			}
+
+			rx_alloc_count += num_segs;
+			rxq->rx_segs += num_segs;
 		}
 		rxq->rx_segs++; /* for the first segment */
@@ -1627,5 +1675,6 @@ next_cqe:
 	}
 
-	qede_update_rx_prod(qdev, rxq);
+	/* Request number of bufferes to be allocated in next loop */
+	rxq->rx_alloc_count = rx_alloc_count;
 
 	rxq->rcv_pkts += rx_pkt;
diff --git a/drivers/net/qede/qede_rxtx.h b/drivers/net/qede/qede_rxtx.h
index 454daa07b..5b249cbb2 100644
--- a/drivers/net/qede/qede_rxtx.h
+++ b/drivers/net/qede/qede_rxtx.h
@@ -193,4 +193,6 @@ struct qede_rx_queue {
 	uint16_t port_id;
 	uint16_t rx_buf_size;
+	uint16_t rx_alloc_count;
+	uint16_t unused;
 	uint64_t rcv_pkts;
 	uint64_t rx_segs;
-- 
2.19.0

---
  Diff of the applied patch vs upstream commit (please double-check if non-empty:
---
--- -	2019-02-07 13:19:56.429751219 +0000
+++ 0038-net-qede-fix-performance-bottleneck-in-Rx-path.patch	2019-02-07 13:19:55.000000000 +0000
@@ -1,8 +1,10 @@
-From 8f2312474529ad7ff0e4b65b82efc8530e7484ce Mon Sep 17 00:00:00 2001
+From 7e22a66545262ffcbd07ff6a266614dfeff643e9 Mon Sep 17 00:00:00 2001
 From: Shahed Shaikh <shshaikh@marvell.com>
 Date: Fri, 18 Jan 2019 02:29:29 -0800
 Subject: [PATCH] net/qede: fix performance bottleneck in Rx path
 
+[ upstream commit 8f2312474529ad7ff0e4b65b82efc8530e7484ce ]
+
 Allocating replacement buffer per received packet is expensive.
 Instead, process received packets first and allocate
 replacement buffers in bulk later.
@@ -11,7 +13,6 @@
 platforms.
 
 Fixes: 2ea6f76aff40 ("qede: add core driver")
-Cc: stable@dpdk.org
 
 Signed-off-by: Shahed Shaikh <shshaikh@marvell.com>
 Acked-by: Rasesh Mody <rmody@marvell.com>