On 6/29/19 11:57 AM, Jakub Kicinski wrote:

On Fri, 28 Jun 2019 14:39:29 -0700, Shannon Nelson wrote:

quoted

+static int ionic_tx(struct queue *q, struct sk_buff *skb)
+{
+	struct tx_stats *stats = q_to_tx_stats(q);
+	int err;
+
+	if (skb->ip_summed == CHECKSUM_PARTIAL)
+		err = ionic_tx_calc_csum(q, skb);
+	else
+		err = ionic_tx_calc_no_csum(q, skb);
+	if (err)
+		return err;
+
+	err = ionic_tx_skb_frags(q, skb);
+	if (err)
+		return err;
+
+	skb_tx_timestamp(skb);
+	stats->pkts++;
+	stats->bytes += skb->len;

Presumably this is 64bit so you should use
u64_stats_update_begin()
u64_stats_update_end()
around it (and all other stats).

Since the device won't work in a 32-bit arch and I have the the Kconfig 
set to depend on 64BIT, I wasn't sure I needed to bother with the extra 
syntactic sugar.

quoted

+
+	ionic_txq_post(q, !netdev_xmit_more(), ionic_tx_clean, skb);
+
+	return 0;
+}
+
+static int ionic_tx_descs_needed(struct queue *q, struct sk_buff *skb)
+{
+	struct tx_stats *stats = q_to_tx_stats(q);
+	int err;
+
+	/* If TSO, need roundup(skb->len/mss) descs */
+	if (skb_is_gso(skb))
+		return (skb->len / skb_shinfo(skb)->gso_size) + 1;
+
+	/* If non-TSO, just need 1 desc and nr_frags sg elems */
+	if (skb_shinfo(skb)->nr_frags <= IONIC_TX_MAX_SG_ELEMS)
+		return 1;
+
+	/* Too many frags, so linearize */
+	err = skb_linearize(skb);
+	if (err)
+		return err;
+
+	stats->linearize++;
+
+	/* Need 1 desc and zero sg elems */
+	return 1;
+}
+
+netdev_tx_t ionic_start_xmit(struct sk_buff *skb, struct net_device *netdev)
+{
+	u16 queue_index = skb_get_queue_mapping(skb);
+	struct lif *lif = netdev_priv(netdev);
+	struct queue *q;
+	int ndescs;
+	int err;
+
+	if (unlikely(!test_bit(LIF_UP, lif->state))) {
+		dev_kfree_skb(skb);
+		return NETDEV_TX_OK;
+	}
+
+	if (likely(lif_to_txqcq(lif, queue_index)))
+		q = lif_to_txq(lif, queue_index);
+	else
+		q = lif_to_txq(lif, 0);
+
+	ndescs = ionic_tx_descs_needed(q, skb);
+	if (ndescs < 0)
+		goto err_out_drop;
+
+	if (!ionic_q_has_space(q, ndescs)) {
+		netif_stop_subqueue(netdev, queue_index);
+		q->stop++;
+
+		/* Might race with ionic_tx_clean, check again */
+		smp_rmb();
+		if (ionic_q_has_space(q, ndescs)) {
+			netif_wake_subqueue(netdev, queue_index);
+			q->wake++;
+		} else {
+			return NETDEV_TX_BUSY;

This should never really happen..

Couldn't we have an Rx interrupt and a call to ionic_tx_clean() in the 
middle of this?

quoted

+		}
+	}
+
+	if (skb_is_gso(skb))
+		err = ionic_tx_tso(q, skb);
+	else
+		err = ionic_tx(q, skb);
+
+	if (err)
+		goto err_out_drop;

.. at this point if you can't guarantee fitting biggest possible frame
in, you have to stop the ring.

Yep, that would work.

Thanks,
sln