net/mlx5e: Add UDP GSO support

This patch enables UDP GSO support. We enable this by using two WQEs
the first is a UDP LSO WQE for all segments with equal length, and the
second is for the last segment in case it has different length.
Due to HW limitation, before sending, we must adjust the packet length fields.

We measure performance between two Intel(R) Xeon(R) CPU E5-2643 v2 @3.50GHz
machines connected back-to-back with Connectx4-Lx (40Gbps) NICs.
We compare single stream UDP, UDP GSO and UDP GSO with offload.
Performance:
		| MSS (bytes)	| Throughput (Gbps)	| CPU utilization (%)
UDP GSO offload	| 1472		| 35.6			| 8%
UDP GSO 	| 1472		| 25.5			| 17%
UDP 		| 1472		| 10.2			| 17%
UDP GSO offload	| 1024		| 35.6			| 8%
UDP GSO		| 1024		| 19.2			| 17%
UDP 		| 1024		| 5.7			| 17%
UDP GSO offload	| 512		| 33.8			| 16%
UDP GSO		| 512		| 10.4			| 17%
UDP 		| 512		| 3.5			| 17%
Signed-off-by: Boris Pismenny <borisp@mellanox.com>
Signed-off-by: Yossi Kuperman <yossiku@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>

drivers/net/ethernet/mellanox/mlx5/core/Makefile

@@ -14,8 +14,8 @@ mlx5_core-$(CONFIG_MLX5_FPGA) += fpga/cmd.o fpga/core.o fpga/conn.o fpga/sdk.o \
 		fpga/ipsec.o fpga/tls.o
 
 mlx5_core-$(CONFIG_MLX5_CORE_EN) += en_main.o en_common.o en_fs.o en_ethtool.o \
-		en_tx.o en_rx.o en_dim.o en_txrx.o en_stats.o vxlan.o \
-		en_arfs.o en_fs_ethtool.o en_selftest.o en/port.o
+		en_tx.o en_rx.o en_dim.o en_txrx.o en_accel/rxtx.o en_stats.o  \
+		vxlan.o en_arfs.o en_fs_ethtool.o en_selftest.o en/port.o
 
 mlx5_core-$(CONFIG_MLX5_MPFS) += lib/mpfs.o
 

drivers/net/ethernet/mellanox/mlx5/core/en_accel/en_accel.h

@@ -34,12 +34,11 @@
 #ifndef __MLX5E_EN_ACCEL_H__
 #define __MLX5E_EN_ACCEL_H__
 
-#ifdef CONFIG_MLX5_ACCEL
-
 #include <linux/skbuff.h>
 #include <linux/netdevice.h>
 #include "en_accel/ipsec_rxtx.h"
 #include "en_accel/tls_rxtx.h"
+#include "en_accel/rxtx.h"
 #include "en.h"
 
 static inline struct sk_buff *mlx5e_accel_handle_tx(struct sk_buff *skb,
@@ -64,9 +63,13 @@ static inline struct sk_buff *mlx5e_accel_handle_tx(struct sk_buff *skb,
 	}
 #endif
 
+	if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) {
+		skb = mlx5e_udp_gso_handle_tx_skb(dev, sq, skb, wqe, pi);
+		if (unlikely(!skb))
+			return NULL;
+	}
+
 	return skb;
 }
 
-#endif /* CONFIG_MLX5_ACCEL */
-
 #endif /* __MLX5E_EN_ACCEL_H__ */

drivers/net/ethernet/mellanox/mlx5/core/en_accel/rxtx.c

+#include "en_accel/rxtx.h"
+
+static void mlx5e_udp_gso_prepare_last_skb(struct sk_buff *skb,
+					   struct sk_buff *nskb,
+					   int remaining)
+{
+	int bytes_needed = remaining, remaining_headlen, remaining_page_offset;
+	int headlen = skb_transport_offset(skb) + sizeof(struct udphdr);
+	int payload_len = remaining + sizeof(struct udphdr);
+	int k = 0, i, j;
+
+	skb_copy_bits(skb, 0, nskb->data, headlen);
+	nskb->dev = skb->dev;
+	skb_reset_mac_header(nskb);
+	skb_set_network_header(nskb, skb_network_offset(skb));
+	skb_set_transport_header(nskb, skb_transport_offset(skb));
+	skb_set_tail_pointer(nskb, headlen);
+
+	/* How many frags do we need? */
+	for (i = skb_shinfo(skb)->nr_frags - 1; i >= 0; i--) {
+		bytes_needed -= skb_frag_size(&skb_shinfo(skb)->frags[i]);
+		k++;
+		if (bytes_needed <= 0)
+			break;
+	}
+
+	/* Fill the first frag and split it if necessary */
+	j = skb_shinfo(skb)->nr_frags - k;
+	remaining_page_offset = -bytes_needed;
+	skb_fill_page_desc(nskb, 0,
+			   skb_shinfo(skb)->frags[j].page.p,
+			   skb_shinfo(skb)->frags[j].page_offset + remaining_page_offset,
+			   skb_shinfo(skb)->frags[j].size - remaining_page_offset);
+
+	skb_frag_ref(skb, j);
+
+	/* Fill the rest of the frags */
+	for (i = 1; i < k; i++) {
+		j = skb_shinfo(skb)->nr_frags - k + i;
+
+		skb_fill_page_desc(nskb, i,
+				   skb_shinfo(skb)->frags[j].page.p,
+				   skb_shinfo(skb)->frags[j].page_offset,
+				   skb_shinfo(skb)->frags[j].size);
+		skb_frag_ref(skb, j);
+	}
+	skb_shinfo(nskb)->nr_frags = k;
+
+	remaining_headlen = remaining - skb->data_len;
+
+	/* headlen contains remaining data? */
+	if (remaining_headlen > 0)
+		skb_copy_bits(skb, skb->len - remaining, nskb->data + headlen,
+			      remaining_headlen);
+	nskb->len = remaining + headlen;
+	nskb->data_len =  payload_len - sizeof(struct udphdr) +
+		max_t(int, 0, remaining_headlen);
+	nskb->protocol = skb->protocol;
+	if (nskb->protocol == htons(ETH_P_IP)) {
+		ip_hdr(nskb)->id = htons(ntohs(ip_hdr(nskb)->id) +
+					 skb_shinfo(skb)->gso_segs);
+		ip_hdr(nskb)->tot_len =
+			htons(payload_len + sizeof(struct iphdr));
+	} else {
+		ipv6_hdr(nskb)->payload_len = htons(payload_len);
+	}
+	udp_hdr(nskb)->len = htons(payload_len);
+	skb_shinfo(nskb)->gso_size = 0;
+	nskb->ip_summed = skb->ip_summed;
+	nskb->csum_start = skb->csum_start;
+	nskb->csum_offset = skb->csum_offset;
+	nskb->queue_mapping = skb->queue_mapping;
+}
+
+/* might send skbs and update wqe and pi */
+struct sk_buff *mlx5e_udp_gso_handle_tx_skb(struct net_device *netdev,
+					    struct mlx5e_txqsq *sq,
+					    struct sk_buff *skb,
+					    struct mlx5e_tx_wqe **wqe,
+					    u16 *pi)
+{
+	int payload_len = skb_shinfo(skb)->gso_size + sizeof(struct udphdr);
+	int headlen = skb_transport_offset(skb) + sizeof(struct udphdr);
+	int remaining = (skb->len - headlen) % skb_shinfo(skb)->gso_size;
+	struct sk_buff *nskb;
+
+	if (skb->protocol == htons(ETH_P_IP))
+		ip_hdr(skb)->tot_len = htons(payload_len + sizeof(struct iphdr));
+	else
+		ipv6_hdr(skb)->payload_len = htons(payload_len);
+	udp_hdr(skb)->len = htons(payload_len);
+	if (!remaining)
+		return skb;
+
+	nskb = alloc_skb(max_t(int, headlen, headlen + remaining - skb->data_len), GFP_ATOMIC);
+	if (unlikely(!nskb)) {
+		sq->stats->dropped++;
+		return NULL;
+	}
+
+	mlx5e_udp_gso_prepare_last_skb(skb, nskb, remaining);
+
+	skb_shinfo(skb)->gso_segs--;
+	pskb_trim(skb, skb->len - remaining);
+	mlx5e_sq_xmit(sq, skb, *wqe, *pi);
+	mlx5e_sq_fetch_wqe(sq, wqe, pi);
+	return nskb;
+}

drivers/net/ethernet/mellanox/mlx5/core/en_accel/rxtx.h

+
+#ifndef __MLX5E_EN_ACCEL_RX_TX_H__
+#define __MLX5E_EN_ACCEL_RX_TX_H__
+
+#include <linux/skbuff.h>
+#include "en.h"
+
+struct sk_buff *mlx5e_udp_gso_handle_tx_skb(struct net_device *netdev,
+					    struct mlx5e_txqsq *sq,
+					    struct sk_buff *skb,
+					    struct mlx5e_tx_wqe **wqe,
+					    u16 *pi);
+
+#endif

drivers/net/ethernet/mellanox/mlx5/core/en_main.c

@@ -4592,6 +4592,9 @@ static void mlx5e_build_nic_netdev(struct net_device *netdev)
 	netdev->features         |= NETIF_F_HIGHDMA;
 	netdev->features         |= NETIF_F_HW_VLAN_STAG_FILTER;
 
+	netdev->features         |= NETIF_F_GSO_UDP_L4;
+	netdev->hw_features      |= NETIF_F_GSO_UDP_L4;
+
 	netdev->priv_flags       |= IFF_UNICAST_FLT;
 
 	mlx5e_set_netdev_dev_addr(netdev);

drivers/net/ethernet/mellanox/mlx5/core/en_tx.c

@@ -228,6 +228,9 @@ mlx5e_tx_get_gso_ihs(struct mlx5e_txqsq *sq, struct sk_buff *skb)
 		stats->tso_inner_packets++;
 		stats->tso_inner_bytes += skb->len - ihs;
 	} else {
+		if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4)
+			ihs = skb_transport_offset(skb) + sizeof(struct udphdr);
+		else
 			ihs = skb_transport_offset(skb) + tcp_hdrlen(skb);
 		stats->tso_packets++;
 		stats->tso_bytes += skb->len - ihs;
@@ -443,12 +446,11 @@ netdev_tx_t mlx5e_xmit(struct sk_buff *skb, struct net_device *dev)
 	sq = priv->txq2sq[skb_get_queue_mapping(skb)];
 	mlx5e_sq_fetch_wqe(sq, &wqe, &pi);
 
-#ifdef CONFIG_MLX5_ACCEL
 	/* might send skbs and update wqe and pi */
 	skb = mlx5e_accel_handle_tx(skb, sq, dev, &wqe, &pi);
 	if (unlikely(!skb))
 		return NETDEV_TX_OK;
-#endif
+
 	return mlx5e_sq_xmit(sq, skb, wqe, pi);
 }