#define E1000_TX_FLAGS_VLAN_SHIFT      16
 
 static int e1000_tso(struct e1000_adapter *adapter,
-                    struct e1000_tx_ring *tx_ring, struct sk_buff *skb)
+                    struct e1000_tx_ring *tx_ring, struct sk_buff *skb,
+                    __be16 protocol)
 {
        struct e1000_context_desc *context_desc;
        struct e1000_buffer *buffer_info;
 
                hdr_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
                mss = skb_shinfo(skb)->gso_size;
-               if (skb->protocol == htons(ETH_P_IP)) {
+               if (protocol == htons(ETH_P_IP)) {
                        struct iphdr *iph = ip_hdr(skb);
                        iph->tot_len = 0;
                        iph->check = 0;
                                                                 0);
                        cmd_length = E1000_TXD_CMD_IP;
                        ipcse = skb_transport_offset(skb) - 1;
-               } else if (skb->protocol == htons(ETH_P_IPV6)) {
+               } else if (skb_is_gso_v6(skb)) {
                        ipv6_hdr(skb)->payload_len = 0;
                        tcp_hdr(skb)->check =
                                ~csum_ipv6_magic(&ipv6_hdr(skb)->saddr,
 }
 
 static bool e1000_tx_csum(struct e1000_adapter *adapter,
-                         struct e1000_tx_ring *tx_ring, struct sk_buff *skb)
+                         struct e1000_tx_ring *tx_ring, struct sk_buff *skb,
+                         __be16 protocol)
 {
        struct e1000_context_desc *context_desc;
        struct e1000_buffer *buffer_info;
        if (skb->ip_summed != CHECKSUM_PARTIAL)
                return false;
 
-       switch (skb->protocol) {
+       switch (protocol) {
        case cpu_to_be16(ETH_P_IP):
                if (ip_hdr(skb)->protocol == IPPROTO_TCP)
                        cmd_len |= E1000_TXD_CMD_TCP;
        int count = 0;
        int tso;
        unsigned int f;
+       __be16 protocol = vlan_get_protocol(skb);
 
        /* This goes back to the question of how to logically map a Tx queue
         * to a flow.  Right now, performance is impacted slightly negatively
 
        first = tx_ring->next_to_use;
 
-       tso = e1000_tso(adapter, tx_ring, skb);
+       tso = e1000_tso(adapter, tx_ring, skb, protocol);
        if (tso < 0) {
                dev_kfree_skb_any(skb);
                return NETDEV_TX_OK;
                if (likely(hw->mac_type != e1000_82544))
                        tx_ring->last_tx_tso = true;
                tx_flags |= E1000_TX_FLAGS_TSO;
-       } else if (likely(e1000_tx_csum(adapter, tx_ring, skb)))
+       } else if (likely(e1000_tx_csum(adapter, tx_ring, skb, protocol)))
                tx_flags |= E1000_TX_FLAGS_CSUM;
 
-       if (likely(skb->protocol == htons(ETH_P_IP)))
+       if (protocol == htons(ETH_P_IP))
                tx_flags |= E1000_TX_FLAGS_IPV4;
 
        if (unlikely(skb->no_fcs))