patch 'net/qede: fix Tx completion' has been queued to stable release 21.11.1

[Kevin Traynor]

Hi,

FYI, your patch has been queued to stable release 21.11.1

Note it hasn't been pushed to http://dpdk.org/browse/dpdk-stable yet.
It will be pushed if I get no objections before 03/21/22. 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.

Queued patches are on a temporary branch at:
https://github.com/kevintraynor/dpdk-stable

This queued commit can be viewed at:
https://github.com/kevintraynor/dpdk-stable/commit/d300f3f060d2fb096758bdceb125de820f71756b

Thanks.

Kevin

---
>From d300f3f060d2fb096758bdceb125de820f71756b Mon Sep 17 00:00:00 2001
From: Devendra Singh Rawat <dsinghrawat at marvell.com>
Date: Fri, 4 Mar 2022 17:38:31 +0530
Subject: [PATCH] net/qede: fix Tx completion

[ upstream commit bc3045626517c26bd1cca4342d8df8f852407726 ]

Tx completion routine was first incrementing the number of free
slots in Tx ring and then freeing corresponding mbufs in bulk.
In some situations, the number of mbufs freed were less than
number of Tx ring slots freed. This caused Tx ring to get into an
inconsistent state and ultimately application fails to transmit
further traffic.

The fix first updates the Tx ring SW consumer index, then
increments Tx ring free slot number and finally frees the mbuf,
this is done in a single iteration of loop.

Fixes: 2c41740bf19e ("net/qede: get consumer index once")
Fixes: 4996b959cde6 ("net/qede: free packets in bulk")

Signed-off-by: Devendra Singh Rawat <dsinghrawat at marvell.com>
Signed-off-by: Rasesh Mody <rmody at marvell.com>
---
 drivers/net/qede/qede_rxtx.c | 79 +++++++++++++++---------------------
 1 file changed, 33 insertions(+), 46 deletions(-)

diff --git a/drivers/net/qede/qede_rxtx.c b/drivers/net/qede/qede_rxtx.c
index 7088c57b50..0b3b4c7ef6 100644
--- a/drivers/net/qede/qede_rxtx.c
+++ b/drivers/net/qede/qede_rxtx.c
@@ -888,66 +888,53 @@ qede_tx_queue_start(struct rte_eth_dev *eth_dev, uint16_t tx_queue_id)
 
 static inline void
-qede_process_tx_compl(__rte_unused struct ecore_dev *edev,
-		      struct qede_tx_queue *txq)
+qede_free_tx_pkt(struct qede_tx_queue *txq)
 {
-	uint16_t hw_bd_cons;
-	uint16_t sw_tx_cons;
-	uint16_t remaining;
-	uint16_t mask;
 	struct rte_mbuf *mbuf;
 	uint16_t nb_segs;
 	uint16_t idx;
-	uint16_t first_idx;
 
-	rte_compiler_barrier();
-	rte_prefetch0(txq->hw_cons_ptr);
-	sw_tx_cons = ecore_chain_get_cons_idx(&txq->tx_pbl);
-	hw_bd_cons = rte_le_to_cpu_16(*txq->hw_cons_ptr);
-#ifdef RTE_LIBRTE_QEDE_DEBUG_TX
-	PMD_TX_LOG(DEBUG, txq, "Tx Completions = %u\n",
-		   abs(hw_bd_cons - sw_tx_cons));
-#endif
-
-	mask = NUM_TX_BDS(txq);
-	idx = txq->sw_tx_cons & mask;
-
-	remaining = hw_bd_cons - sw_tx_cons;
-	txq->nb_tx_avail += remaining;
-	first_idx = idx;
-
-	while (remaining) {
-		mbuf = txq->sw_tx_ring[idx];
-		RTE_ASSERT(mbuf);
+	idx = TX_CONS(txq);
+	mbuf = txq->sw_tx_ring[idx];
+	if (mbuf) {
 		nb_segs = mbuf->nb_segs;
-		remaining -= nb_segs;
-
-		/* Prefetch the next mbuf. Note that at least the last 4 mbufs
-		 * that are prefetched will not be used in the current call.
-		 */
-		rte_mbuf_prefetch_part1(txq->sw_tx_ring[(idx + 4) & mask]);
-		rte_mbuf_prefetch_part2(txq->sw_tx_ring[(idx + 4) & mask]);
-
 		PMD_TX_LOG(DEBUG, txq, "nb_segs to free %u\n", nb_segs);
-
 		while (nb_segs) {
+			/* It's like consuming rxbuf in recv() */
 			ecore_chain_consume(&txq->tx_pbl);
+			txq->nb_tx_avail++;
 			nb_segs--;
 		}
-
-		idx = (idx + 1) & mask;
+		rte_pktmbuf_free(mbuf);
+		txq->sw_tx_ring[idx] = NULL;
+		txq->sw_tx_cons++;
 		PMD_TX_LOG(DEBUG, txq, "Freed tx packet\n");
-	}
-	txq->sw_tx_cons = idx;
-
-	if (first_idx > idx) {
-		rte_pktmbuf_free_bulk(&txq->sw_tx_ring[first_idx],
-							  mask - first_idx + 1);
-		rte_pktmbuf_free_bulk(&txq->sw_tx_ring[0], idx);
 	} else {
-		rte_pktmbuf_free_bulk(&txq->sw_tx_ring[first_idx],
-							  idx - first_idx);
+		ecore_chain_consume(&txq->tx_pbl);
+		txq->nb_tx_avail++;
 	}
 }
 
+static inline void
+qede_process_tx_compl(__rte_unused struct ecore_dev *edev,
+		      struct qede_tx_queue *txq)
+{
+	uint16_t hw_bd_cons;
+#ifdef RTE_LIBRTE_QEDE_DEBUG_TX
+	uint16_t sw_tx_cons;
+#endif
+
+	hw_bd_cons = rte_le_to_cpu_16(*txq->hw_cons_ptr);
+	/* read barrier prevents speculative execution on stale data */
+	rte_rmb();
+
+#ifdef RTE_LIBRTE_QEDE_DEBUG_TX
+	sw_tx_cons = ecore_chain_get_cons_idx(&txq->tx_pbl);
+	PMD_TX_LOG(DEBUG, txq, "Tx Completions = %u\n",
+		   abs(hw_bd_cons - sw_tx_cons));
+#endif
+	while (hw_bd_cons !=  ecore_chain_get_cons_idx(&txq->tx_pbl))
+		qede_free_tx_pkt(txq);
+}
+
 static int qede_drain_txq(struct qede_dev *qdev,
 			  struct qede_tx_queue *txq, bool allow_drain)
-- 
2.34.1

---
  Diff of the applied patch vs upstream commit (please double-check if non-empty:
---
--- -	2022-03-16 15:14:12.309354115 +0000
+++ 0006-net-qede-fix-Tx-completion.patch	2022-03-16 15:14:12.102847584 +0000
@@ -1 +1 @@
-From bc3045626517c26bd1cca4342d8df8f852407726 Mon Sep 17 00:00:00 2001
+From d300f3f060d2fb096758bdceb125de820f71756b Mon Sep 17 00:00:00 2001
@@ -5,0 +6,2 @@
+[ upstream commit bc3045626517c26bd1cca4342d8df8f852407726 ]
+
@@ -19 +20,0 @@
-Cc: stable at dpdk.org
@@ -28 +29 @@
-index 911bb1a260..0c52568180 100644
+index 7088c57b50..0b3b4c7ef6 100644
@@ -31 +32 @@
-@@ -886,66 +886,53 @@ qede_tx_queue_start(struct rte_eth_dev *eth_dev, uint16_t tx_queue_id)
+@@ -888,66 +888,53 @@ qede_tx_queue_start(struct rte_eth_dev *eth_dev, uint16_t tx_queue_id)