[PATCH 1/3] net/qede: fix Tx callback completion routine

[Devendra Singh Rawat]

Tx completion routine was first incrementing no. of free slots in Tx
ring and then freeing corresponding mbufs in bulk. In some situations
no. of mbufs freed were less than no. 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 Tx ring SW consumer index, then increments Tx ring
free slot no. 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")
Cc: stable at dpdk.org

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 911bb1a260..0c52568180 100644
--- a/drivers/net/qede/qede_rxtx.c
+++ b/drivers/net/qede/qede_rxtx.c
@@ -885,68 +885,55 @@ 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)
 {
-- 
2.18.2