net: thunderx: Fix transmit queue timeout issue

Transmit queue timeout issue is seen in two cases
- Due to a race condition btw setting stop_queue at xmit()
  and checking for stopped_queue in NAPI poll routine, at times
  transmission from a SQ comes to a halt. This is fixed
  by using barriers and also added a check for SQ free descriptors,
  incase SQ is stopped and there are only CQE_RX i.e no CQE_TX.
- Contrary to an assumption, a HW errata where HW doesn't stop transmission
  even though there are not enough CQEs available for a CQE_TX is
  not fixed in T88 pass 2.x. This results in a Qset error with
  'CQ_WR_FULL' stalling transmission. This is fixed by adjusting
  RXQ's  RED levels for CQ level such that there is always enough
  space left for CQE_TXs.
Signed-off-by: Sunil Goutham <sgoutham@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/cavium/thunder/nicvf_main.c

@@ -644,6 +644,7 @@ static int nicvf_cq_intr_handler(struct net_device *netdev, u8 cq_idx,
 	struct cmp_queue *cq = &qs->cq[cq_idx];
 	struct cqe_rx_t *cq_desc;
 	struct netdev_queue *txq;
+	struct snd_queue *sq;
 	unsigned int tx_pkts = 0, tx_bytes = 0;
 
 	spin_lock_bh(&cq->lock);
@@ -709,16 +710,20 @@ static int nicvf_cq_intr_handler(struct net_device *netdev, u8 cq_idx,
 
 done:
 	/* Wakeup TXQ if its stopped earlier due to SQ full */
-	if (tx_done) {
+	sq = &nic->qs->sq[cq_idx];
+	if (tx_done ||
+	    (atomic_read(&sq->free_cnt) >= MIN_SQ_DESC_PER_PKT_XMIT)) {
 		netdev = nic->pnicvf->netdev;
 		txq = netdev_get_tx_queue(netdev,
 					  nicvf_netdev_qidx(nic, cq_idx));
 		if (tx_pkts)
 			netdev_tx_completed_queue(txq, tx_pkts, tx_bytes);
 
-		nic = nic->pnicvf;
+		/* To read updated queue and carrier status */
+		smp_mb();
 		if (netif_tx_queue_stopped(txq) && netif_carrier_ok(netdev)) {
-			netif_tx_start_queue(txq);
+			netif_tx_wake_queue(txq);
+			nic = nic->pnicvf;
 			this_cpu_inc(nic->drv_stats->txq_wake);
 			if (netif_msg_tx_err(nic))
 				netdev_warn(netdev,
@@ -1054,6 +1059,9 @@ static netdev_tx_t nicvf_xmit(struct sk_buff *skb, struct net_device *netdev)
 	struct nicvf *nic = netdev_priv(netdev);
 	int qid = skb_get_queue_mapping(skb);
 	struct netdev_queue *txq = netdev_get_tx_queue(netdev, qid);
+	struct nicvf *snic;
+	struct snd_queue *sq;
+	int tmp;
 
 	/* Check for minimum packet length */
 	if (skb->len <= ETH_HLEN) {
@@ -1061,13 +1069,39 @@ static netdev_tx_t nicvf_xmit(struct sk_buff *skb, struct net_device *netdev)
 		return NETDEV_TX_OK;
 	}
 
-	if (!netif_tx_queue_stopped(txq) && !nicvf_sq_append_skb(nic, skb)) {
+	snic = nic;
+	/* Get secondary Qset's SQ structure */
+	if (qid >= MAX_SND_QUEUES_PER_QS) {
+		tmp = qid / MAX_SND_QUEUES_PER_QS;
+		snic = (struct nicvf *)nic->snicvf[tmp - 1];
+		if (!snic) {
+			netdev_warn(nic->netdev,
+				    "Secondary Qset#%d's ptr not initialized\n",
+				    tmp - 1);
+			dev_kfree_skb(skb);
+			return NETDEV_TX_OK;
+		}
+		qid = qid % MAX_SND_QUEUES_PER_QS;
+	}
+
+	sq = &snic->qs->sq[qid];
+	if (!netif_tx_queue_stopped(txq) &&
+	    !nicvf_sq_append_skb(snic, sq, skb, qid)) {
 		netif_tx_stop_queue(txq);
+
+		/* Barrier, so that stop_queue visible to other cpus */
+		smp_mb();
+
+		/* Check again, incase another cpu freed descriptors */
+		if (atomic_read(&sq->free_cnt) > MIN_SQ_DESC_PER_PKT_XMIT) {
+			netif_tx_wake_queue(txq);
+		} else {
 			this_cpu_inc(nic->drv_stats->txq_stop);
 			if (netif_msg_tx_err(nic))
 				netdev_warn(netdev,
 					    "%s: Transmit ring full, stopping SQ%d\n",
 					    netdev->name, qid);
+		}
 		return NETDEV_TX_BUSY;
 	}
 

drivers/net/ethernet/cavium/thunder/nicvf_queues.c

@@ -1190,30 +1190,12 @@ static int nicvf_sq_append_tso(struct nicvf *nic, struct snd_queue *sq,
 }
 
 /* Append an skb to a SQ for packet transfer. */
-int nicvf_sq_append_skb(struct nicvf *nic, struct sk_buff *skb)
+int nicvf_sq_append_skb(struct nicvf *nic, struct snd_queue *sq,
+			struct sk_buff *skb, u8 sq_num)
 {
 	int i, size;
 	int subdesc_cnt, tso_sqe = 0;
-	int sq_num, qentry;
-	struct queue_set *qs;
-	struct snd_queue *sq;
-
-	sq_num = skb_get_queue_mapping(skb);
-	if (sq_num >= MAX_SND_QUEUES_PER_QS) {
-		/* Get secondary Qset's SQ structure */
-		i = sq_num / MAX_SND_QUEUES_PER_QS;
-		if (!nic->snicvf[i - 1]) {
-			netdev_warn(nic->netdev,
-				    "Secondary Qset#%d's ptr not initialized\n",
-				    i - 1);
-			return 1;
-		}
-		nic = (struct nicvf *)nic->snicvf[i - 1];
-		sq_num = sq_num % MAX_SND_QUEUES_PER_QS;
-	}
-
-	qs = nic->qs;
-	sq = &qs->sq[sq_num];
+	int qentry;
 
 	subdesc_cnt = nicvf_sq_subdesc_required(nic, skb);
 	if (subdesc_cnt > atomic_read(&sq->free_cnt))

drivers/net/ethernet/cavium/thunder/nicvf_queues.h

@@ -88,13 +88,13 @@
 
 /* RED and Backpressure levels of CQ for pkt reception
  * For CQ, level is a measure of emptiness i.e 0x0 means full
- * eg: For CQ of size 4K, and for pass/drop levels of 128/96
- * HW accepts pkt if unused CQE >= 2048
- * RED accepts pkt if unused CQE < 2048 & >= 1536
- * DROPs pkts if unused CQE < 1536
+ * eg: For CQ of size 4K, and for pass/drop levels of 160/144
+ * HW accepts pkt if unused CQE >= 2560
+ * RED accepts pkt if unused CQE < 2304 & >= 2560
+ * DROPs pkts if unused CQE < 2304
  */
-#define RQ_PASS_CQ_LVL		128ULL
-#define RQ_DROP_CQ_LVL		96ULL
+#define RQ_PASS_CQ_LVL		160ULL
+#define RQ_DROP_CQ_LVL		144ULL
 
 /* RED and Backpressure levels of RBDR for pkt reception
  * For RBDR, level is a measure of fullness i.e 0x0 means empty
@@ -306,7 +306,8 @@ void nicvf_sq_disable(struct nicvf *nic, int qidx);
 void nicvf_put_sq_desc(struct snd_queue *sq, int desc_cnt);
 void nicvf_sq_free_used_descs(struct net_device *netdev,
 			      struct snd_queue *sq, int qidx);
-int nicvf_sq_append_skb(struct nicvf *nic, struct sk_buff *skb);
+int nicvf_sq_append_skb(struct nicvf *nic, struct snd_queue *sq,
+			struct sk_buff *skb, u8 sq_num);
 
 struct sk_buff *nicvf_get_rcv_skb(struct nicvf *nic, struct cqe_rx_t *cqe_rx);
 void nicvf_rbdr_task(unsigned long data);