IPoIB: add LRO support

Add "ipoib_use_lro" module parameter to enable LRO and an
"ipoib_lro_max_aggr" module parameter to set the max number of packets
to be aggregated.  Make LRO controllable and LRO statistics accessible
through ethtool.
Signed-off-by: Vladimir Sokolovsky <vlad@mellanox.co.il>
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>

drivers/infiniband/ulp/ipoib/Kconfig

 config INFINIBAND_IPOIB
 	tristate "IP-over-InfiniBand"
 	depends on NETDEVICES && INET && (IPV6 || IPV6=n)
+	select INET_LRO
 	---help---
 	  Support for the IP-over-InfiniBand protocol (IPoIB). This
 	  transports IP packets over InfiniBand so you can use your IB

drivers/infiniband/ulp/ipoib/ipoib.h

@@ -50,6 +50,7 @@
 #include <rdma/ib_verbs.h>
 #include <rdma/ib_pack.h>
 #include <rdma/ib_sa.h>
+#include <linux/inet_lro.h>
 
 /* constants */
 
@@ -94,6 +95,9 @@ enum {
 	IPOIB_MCAST_FLAG_BUSY	  = 2,	/* joining or already joined */
 	IPOIB_MCAST_FLAG_ATTACHED = 3,
 
+	IPOIB_MAX_LRO_DESCRIPTORS = 8,
+	IPOIB_LRO_MAX_AGGR 	  = 64,
+
 	MAX_SEND_CQE		  = 16,
 	IPOIB_CM_COPYBREAK	  = 256,
 };
@@ -248,6 +252,11 @@ struct ipoib_ethtool_st {
 	u16     max_coalesced_frames;
 };
 
+struct ipoib_lro {
+	struct net_lro_mgr lro_mgr;
+	struct net_lro_desc lro_desc[IPOIB_MAX_LRO_DESCRIPTORS];
+};
+
 /*
  * Device private locking: tx_lock protects members used in TX fast
  * path (and we use LLTX so upper layers don't do extra locking).
@@ -334,6 +343,8 @@ struct ipoib_dev_priv {
 	int	hca_caps;
 	struct ipoib_ethtool_st ethtool;
 	struct timer_list poll_timer;
+
+	struct ipoib_lro lro;
 };
 
 struct ipoib_ah {

drivers/infiniband/ulp/ipoib/ipoib_ethtool.c

@@ -86,11 +86,57 @@ static int ipoib_set_coalesce(struct net_device *dev,
 	return 0;
 }
 
+static const char ipoib_stats_keys[][ETH_GSTRING_LEN] = {
+	"LRO aggregated", "LRO flushed",
+	"LRO avg aggr", "LRO no desc"
+};
+
+static void ipoib_get_strings(struct net_device *netdev, u32 stringset, u8 *data)
+{
+	switch (stringset) {
+	case ETH_SS_STATS:
+		memcpy(data, *ipoib_stats_keys,	sizeof(ipoib_stats_keys));
+		break;
+	}
+}
+
+static int ipoib_get_sset_count(struct net_device *dev, int sset)
+{
+	switch (sset) {
+	case ETH_SS_STATS:
+		return ARRAY_SIZE(ipoib_stats_keys);
+	default:
+		return -EOPNOTSUPP;
+	}
+}
+
+static void ipoib_get_ethtool_stats(struct net_device *dev,
+				struct ethtool_stats *stats, uint64_t *data)
+{
+	struct ipoib_dev_priv *priv = netdev_priv(dev);
+	int index = 0;
+
+	/* Get LRO statistics */
+	data[index++] = priv->lro.lro_mgr.stats.aggregated;
+	data[index++] = priv->lro.lro_mgr.stats.flushed;
+	if (priv->lro.lro_mgr.stats.flushed)
+		data[index++] = priv->lro.lro_mgr.stats.aggregated /
+				priv->lro.lro_mgr.stats.flushed;
+	else
+		data[index++] = 0;
+	data[index++] = priv->lro.lro_mgr.stats.no_desc;
+}
+
 static const struct ethtool_ops ipoib_ethtool_ops = {
 	.get_drvinfo		= ipoib_get_drvinfo,
 	.get_tso		= ethtool_op_get_tso,
 	.get_coalesce		= ipoib_get_coalesce,
 	.set_coalesce		= ipoib_set_coalesce,
+	.get_flags		= ethtool_op_get_flags,
+	.set_flags		= ethtool_op_set_flags,
+	.get_strings		= ipoib_get_strings,
+	.get_sset_count		= ipoib_get_sset_count,
+	.get_ethtool_stats	= ipoib_get_ethtool_stats,
 };
 
 void ipoib_set_ethtool_ops(struct net_device *dev)

drivers/infiniband/ulp/ipoib/ipoib_ib.c

@@ -288,6 +288,9 @@ static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 	if (test_bit(IPOIB_FLAG_CSUM, &priv->flags) && likely(wc->csum_ok))
 		skb->ip_summed = CHECKSUM_UNNECESSARY;
 
+	if (dev->features & NETIF_F_LRO)
+		lro_receive_skb(&priv->lro.lro_mgr, skb, NULL);
+	else
 		netif_receive_skb(skb);
 
 repost:
@@ -440,6 +443,9 @@ poll_more:
 	}
 
 	if (done < budget) {
+		if (dev->features & NETIF_F_LRO)
+			lro_flush_all(&priv->lro.lro_mgr);
+
 		netif_rx_complete(dev, napi);
 		if (unlikely(ib_req_notify_cq(priv->recv_cq,
 					      IB_CQ_NEXT_COMP |

drivers/infiniband/ulp/ipoib/ipoib_main.c

@@ -60,6 +60,15 @@ MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
 
+static int lro;
+module_param(lro, bool, 0444);
+MODULE_PARM_DESC(lro,  "Enable LRO (Large Receive Offload)");
+
+static int lro_max_aggr = IPOIB_LRO_MAX_AGGR;
+module_param(lro_max_aggr, int, 0644);
+MODULE_PARM_DESC(lro_max_aggr, "LRO: Max packets to be aggregated "
+		"(default = 64)");
+
 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
 int ipoib_debug_level;
 
@@ -936,6 +945,54 @@ static const struct header_ops ipoib_header_ops = {
 	.create	= ipoib_hard_header,
 };
 
+static int get_skb_hdr(struct sk_buff *skb, void **iphdr,
+		       void **tcph, u64 *hdr_flags, void *priv)
+{
+	unsigned int ip_len;
+	struct iphdr *iph;
+
+	if (unlikely(skb->protocol != htons(ETH_P_IP)))
+		return -1;
+
+	/*
+	 * In the future we may add an else clause that verifies the
+	 * checksum and allows devices which do not calculate checksum
+	 * to use LRO.
+	 */
+	if (unlikely(skb->ip_summed != CHECKSUM_UNNECESSARY))
+		return -1;
+
+	/* Check for non-TCP packet */
+	skb_reset_network_header(skb);
+	iph = ip_hdr(skb);
+	if (iph->protocol != IPPROTO_TCP)
+		return -1;
+
+	ip_len = ip_hdrlen(skb);
+	skb_set_transport_header(skb, ip_len);
+	*tcph = tcp_hdr(skb);
+
+	/* check if IP header and TCP header are complete */
+	if (ntohs(iph->tot_len) < ip_len + tcp_hdrlen(skb))
+		return -1;
+
+	*hdr_flags = LRO_IPV4 | LRO_TCP;
+	*iphdr = iph;
+
+	return 0;
+}
+
+static void ipoib_lro_setup(struct ipoib_dev_priv *priv)
+{
+	priv->lro.lro_mgr.max_aggr	 = lro_max_aggr;
+	priv->lro.lro_mgr.max_desc	 = IPOIB_MAX_LRO_DESCRIPTORS;
+	priv->lro.lro_mgr.lro_arr	 = priv->lro.lro_desc;
+	priv->lro.lro_mgr.get_skb_header = get_skb_hdr;
+	priv->lro.lro_mgr.features	 = LRO_F_NAPI;
+	priv->lro.lro_mgr.dev		 = priv->dev;
+	priv->lro.lro_mgr.ip_summed_aggr = CHECKSUM_UNNECESSARY;
+}
+
 static void ipoib_setup(struct net_device *dev)
 {
 	struct ipoib_dev_priv *priv = netdev_priv(dev);
@@ -975,6 +1032,8 @@ static void ipoib_setup(struct net_device *dev)
 
 	priv->dev = dev;
 
+	ipoib_lro_setup(priv);
+
 	spin_lock_init(&priv->lock);
 	spin_lock_init(&priv->tx_lock);
 
@@ -1152,6 +1211,9 @@ static struct net_device *ipoib_add_port(const char *format,
 		priv->dev->features |= NETIF_F_SG | NETIF_F_IP_CSUM;
 	}
 
+	if (lro)
+		priv->dev->features |= NETIF_F_LRO;
+
 	/*
 	 * Set the full membership bit, so that we join the right
 	 * broadcast group, etc.