[netdrvr] sfc: Add TSO support

The SFC4000 controller does not have hardware support for TSO, and the
core GSO code incurs a high cost in allocating and freeing skbs.  This
TSO implementation uses lightweight packet header structures and is
substantially faster.
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>

drivers/net/sfc/efx.c

@@ -1873,6 +1873,7 @@ static int efx_init_struct(struct efx_nic *efx, struct efx_nic_type *type,
 		tx_queue->queue = i;
 		tx_queue->buffer = NULL;
 		tx_queue->channel = &efx->channel[0]; /* for safety */
+		tx_queue->tso_headers_free = NULL;
 	}
 	for (i = 0; i < EFX_MAX_RX_QUEUES; i++) {
 		rx_queue = &efx->rx_queue[i];
@@ -2071,7 +2072,8 @@ static int __devinit efx_pci_probe(struct pci_dev *pci_dev,
 	net_dev = alloc_etherdev(sizeof(*efx));
 	if (!net_dev)
 		return -ENOMEM;
-	net_dev->features |= NETIF_F_IP_CSUM | NETIF_F_SG | NETIF_F_HIGHDMA;
+	net_dev->features |= (NETIF_F_IP_CSUM | NETIF_F_SG |
+			      NETIF_F_HIGHDMA | NETIF_F_TSO);
 	if (lro)
 		net_dev->features |= NETIF_F_LRO;
 	efx = net_dev->priv;

drivers/net/sfc/ethtool.c

@@ -272,6 +272,22 @@ static void efx_ethtool_get_stats(struct net_device *net_dev,
 	}
 }
 
+static int efx_ethtool_set_tso(struct net_device *net_dev, u32 enable)
+{
+	int rc;
+
+	/* Our TSO requires TX checksumming, so force TX checksumming
+	 * on when TSO is enabled.
+	 */
+	if (enable) {
+		rc = efx_ethtool_set_tx_csum(net_dev, 1);
+		if (rc)
+			return rc;
+	}
+
+	return ethtool_op_set_tso(net_dev, enable);
+}
+
 static int efx_ethtool_set_tx_csum(struct net_device *net_dev, u32 enable)
 {
 	struct efx_nic *efx = net_dev->priv;
@@ -283,6 +299,15 @@ static int efx_ethtool_set_tx_csum(struct net_device *net_dev, u32 enable)
 
 	efx_flush_queues(efx);
 
+	/* Our TSO requires TX checksumming, so disable TSO when
+	 * checksumming is disabled
+	 */
+	if (!enable) {
+		rc = efx_ethtool_set_tso(net_dev, 0);
+		if (rc)
+			return rc;
+	}
+
 	return 0;
 }
 
@@ -451,6 +476,8 @@ struct ethtool_ops efx_ethtool_ops = {
 	.set_tx_csum		= efx_ethtool_set_tx_csum,
 	.get_sg			= ethtool_op_get_sg,
 	.set_sg			= ethtool_op_set_sg,
+	.get_tso		= ethtool_op_get_tso,
+	.set_tso		= efx_ethtool_set_tso,
 	.get_flags		= ethtool_op_get_flags,
 	.set_flags		= ethtool_op_set_flags,
 	.get_strings		= efx_ethtool_get_strings,

drivers/net/sfc/net_driver.h

@@ -134,6 +134,8 @@ struct efx_special_buffer {
  *	Set only on the final fragment of a packet; %NULL for all other
  *	fragments.  When this fragment completes, then we can free this
  *	skb.
+ * @tsoh: The associated TSO header structure, or %NULL if this
+ *	buffer is not a TSO header.
  * @dma_addr: DMA address of the fragment.
  * @len: Length of this fragment.
  *	This field is zero when the queue slot is empty.
@@ -144,6 +146,7 @@ struct efx_special_buffer {
  */
 struct efx_tx_buffer {
 	const struct sk_buff *skb;
+	struct efx_tso_header *tsoh;
 	dma_addr_t dma_addr;
 	unsigned short len;
 	unsigned char continuation;
@@ -187,6 +190,13 @@ struct efx_tx_buffer {
  *	variable indicates that the queue is full.  This is to
  *	avoid cache-line ping-pong between the xmit path and the
  *	completion path.
+ * @tso_headers_free: A list of TSO headers allocated for this TX queue
+ *	that are not in use, and so available for new TSO sends. The list
+ *	is protected by the TX queue lock.
+ * @tso_bursts: Number of times TSO xmit invoked by kernel
+ * @tso_long_headers: Number of packets with headers too long for standard
+ *	blocks
+ * @tso_packets: Number of packets via the TSO xmit path
  */
 struct efx_tx_queue {
 	/* Members which don't change on the fast path */
@@ -206,6 +216,10 @@ struct efx_tx_queue {
 	unsigned int insert_count ____cacheline_aligned_in_smp;
 	unsigned int write_count;
 	unsigned int old_read_count;
+	struct efx_tso_header *tso_headers_free;
+	unsigned int tso_bursts;
+	unsigned int tso_long_headers;
+	unsigned int tso_packets;
 };
 
 /**