[dpdk-dev] [PATCH v2 20/23] mlx4: add L2 tunnel (VXLAN) checksum offload support

[Adrien Mazarguil <adrien.mazarguil@6wind.com>]

Depending on adapters features and VXLAN support in the kernel, VXLAN frames
can be automatically recognized, in which case checksum validation and
generation occurs on inner and outer L3 and L4.

Signed-off-by: Olga Shern <olgas@mellanox.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
---
 drivers/net/mlx4/mlx4.c | 49 ++++++++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 48 insertions(+), 1 deletion(-)

diff --git a/drivers/net/mlx4/mlx4.c b/drivers/net/mlx4/mlx4.c
index fa9216f..3c72235 100644
--- a/drivers/net/mlx4/mlx4.c
+++ b/drivers/net/mlx4/mlx4.c
@@ -203,6 +203,7 @@ struct rxq {
 	} elts;
 	unsigned int sp:1; /* Use scattered RX elements. */
 	unsigned int csum:1; /* Enable checksum offloading. */
+	unsigned int csum_l2tun:1; /* Same for L2 tunnels. */
 	uint32_t mb_len; /* Length of a mp-issued mbuf. */
 	struct mlx4_rxq_stats stats; /* RX queue counters. */
 	unsigned int socket; /* CPU socket ID for allocations. */
@@ -276,6 +277,7 @@ struct priv {
 	unsigned int hw_tss:1; /* TSS is supported. */
 	unsigned int hw_rss:1; /* RSS is supported. */
 	unsigned int hw_csum:1; /* Checksum offload is supported. */
+	unsigned int hw_csum_l2tun:1; /* Same for L2 tunnels. */
 	unsigned int rss:1; /* RSS is enabled. */
 	unsigned int vf:1; /* This is a VF device. */
 #ifdef INLINE_RECV
@@ -1243,8 +1245,21 @@ mlx4_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 		}
 		/* Should we enable HW CKSUM offload */
 		if (buf->ol_flags &
-		    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM))
+		    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
 			send_flags |= IBV_EXP_QP_BURST_IP_CSUM;
+			/* HW does not support checksum offloads at arbitrary
+			 * offsets but automatically recognizes the packet
+			 * type. For inner L3/L4 checksums, only VXLAN (UDP)
+			 * tunnels are currently supported.
+			 *
+			 * FIXME: since PKT_TX_UDP_TUNNEL_PKT has been removed,
+			 * the outer packet type is unknown. All we know is
+			 * that the L2 header is of unusual length (not
+			 * ETHER_HDR_LEN with or without 802.1Q header). */
+			if ((buf->l2_len != ETHER_HDR_LEN) &&
+			    (buf->l2_len != (ETHER_HDR_LEN + 4)))
+				send_flags |= IBV_EXP_QP_BURST_TUNNEL;
+		}
 		if (likely(segs == 1)) {
 			uintptr_t addr;
 			uint32_t length;
@@ -2443,6 +2458,25 @@ rxq_cq_to_ol_flags(const struct rxq *rxq, uint32_t flags)
 			TRANSPOSE(~flags,
 				  IBV_EXP_CQ_RX_TCP_UDP_CSUM_OK,
 				  PKT_RX_L4_CKSUM_BAD);
+	/*
+	 * PKT_RX_IP_CKSUM_BAD and PKT_RX_L4_CKSUM_BAD are used in place
+	 * of PKT_RX_EIP_CKSUM_BAD because the latter is not functional
+	 * (its value is 0).
+	 */
+	if ((flags & IBV_EXP_CQ_RX_TUNNEL_PACKET) && (rxq->csum_l2tun))
+		ol_flags |=
+			TRANSPOSE(flags,
+				  IBV_EXP_CQ_RX_OUTER_IPV4_PACKET,
+				  PKT_RX_TUNNEL_IPV4_HDR) |
+			TRANSPOSE(flags,
+				  IBV_EXP_CQ_RX_OUTER_IPV6_PACKET,
+				  PKT_RX_TUNNEL_IPV6_HDR) |
+			TRANSPOSE(~flags,
+				  IBV_EXP_CQ_RX_OUTER_IP_CSUM_OK,
+				  PKT_RX_IP_CKSUM_BAD) |
+			TRANSPOSE(~flags,
+				  IBV_EXP_CQ_RX_OUTER_TCP_UDP_CSUM_OK,
+				  PKT_RX_L4_CKSUM_BAD);
 	return ol_flags;
 }
 
@@ -2976,6 +3010,10 @@ rxq_rehash(struct rte_eth_dev *dev, struct rxq *rxq)
 		tmpl.csum = !!dev->data->dev_conf.rxmode.hw_ip_checksum;
 		rxq->csum = tmpl.csum;
 	}
+	if (priv->hw_csum_l2tun) {
+		tmpl.csum_l2tun = !!dev->data->dev_conf.rxmode.hw_ip_checksum;
+		rxq->csum_l2tun = tmpl.csum_l2tun;
+	}
 	/* Enable scattered packets support for this queue if necessary. */
 	if ((dev->data->dev_conf.rxmode.jumbo_frame) &&
 	    (dev->data->dev_conf.rxmode.max_rx_pkt_len >
@@ -3200,6 +3238,8 @@ rxq_setup(struct rte_eth_dev *dev, struct rxq *rxq, uint16_t desc,
 	/* Toggle RX checksum offload if hardware supports it. */
 	if (priv->hw_csum)
 		tmpl.csum = !!dev->data->dev_conf.rxmode.hw_ip_checksum;
+	if (priv->hw_csum_l2tun)
+		tmpl.csum_l2tun = !!dev->data->dev_conf.rxmode.hw_ip_checksum;
 	/* Enable scattered packets support for this queue if necessary. */
 	if ((dev->data->dev_conf.rxmode.jumbo_frame) &&
 	    (dev->data->dev_conf.rxmode.max_rx_pkt_len >
@@ -4427,6 +4467,8 @@ static const struct eth_dev_ops mlx4_dev_ops = {
 	.mac_addr_remove = mlx4_mac_addr_remove,
 	.mac_addr_add = mlx4_mac_addr_add,
 	.mtu_set = mlx4_dev_set_mtu,
+	.udp_tunnel_add = NULL,
+	.udp_tunnel_del = NULL,
 	.fdir_add_signature_filter = NULL,
 	.fdir_update_signature_filter = NULL,
 	.fdir_remove_signature_filter = NULL,
@@ -4757,6 +4799,11 @@ mlx4_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
 		DEBUG("checksum offloading is %ssupported",
 		      (priv->hw_csum ? "" : "not "));
 
+		priv->hw_csum_l2tun = !!(exp_device_attr.exp_device_cap_flags &
+					 IBV_EXP_DEVICE_VXLAN_SUPPORT);
+		DEBUG("L2 tunnel checksum offloads are %ssupported",
+		      (priv->hw_csum_l2tun ? "" : "not "));
+
 #ifdef INLINE_RECV
 		priv->inl_recv_size = mlx4_getenv_int("MLX4_INLINE_RECV_SIZE");
 
-- 
2.1.0