diff --git a/src/vnet/gso/node.c b/src/vnet/gso/node.c
index 663f9cc..b57830f 100644
--- a/src/vnet/gso/node.c
+++ b/src/vnet/gso/node.c
@@ -56,6 +56,50 @@
   return s;
 }
 
+static_always_inline u16
+tso_segment_ipip_tunnel_fixup (vlib_main_t * vm,
+			       vnet_interface_per_thread_data_t * ptd,
+			       vlib_buffer_t * sb0,
+			       generic_header_offset_t * gho)
+{
+  u16 n_tx_bufs = vec_len (ptd->split_buffers);
+  u16 i = 0, n_tx_bytes = 0;
+
+  while (i < n_tx_bufs)
+    {
+      vlib_buffer_t *b0 = vlib_get_buffer (vm, ptd->split_buffers[i]);
+      vnet_get_outer_header (b0, gho);
+      clib_memcpy_fast (vlib_buffer_get_current (b0),
+			vlib_buffer_get_current (sb0), gho->outer_hdr_sz);
+
+      ip4_header_t *ip4 =
+	(ip4_header_t *) (vlib_buffer_get_current (b0) +
+			  gho->outer_l3_hdr_offset);
+      ip6_header_t *ip6 =
+	(ip6_header_t *) (vlib_buffer_get_current (b0) +
+			  gho->outer_l3_hdr_offset);
+
+      if (gho->gho_flags & GHO_F_OUTER_IP4)
+	{
+	  ip4->length =
+	    clib_host_to_net_u16 (b0->current_length -
+				  gho->outer_l3_hdr_offset);
+	  ip4->checksum = ip4_header_checksum (ip4);
+	}
+      else if (gho->gho_flags & GHO_F_OUTER_IP6)
+	{
+	  ip6->payload_length =
+	    clib_host_to_net_u16 (b0->current_length -
+				  gho->outer_l4_hdr_offset);
+	}
+
+      n_tx_bytes += gho->outer_hdr_sz;
+      i++;
+    }
+  return n_tx_bytes;
+
+}
+
 static_always_inline void
 tso_segment_vxlan_tunnel_headers_fixup (vlib_main_t * vm, vlib_buffer_t * b,
 					generic_header_offset_t * gho)
@@ -177,14 +221,22 @@
 tso_init_buf_from_template_base (vlib_buffer_t * nb0, vlib_buffer_t * b0,
 				 u32 flags, u16 length)
 {
+  /* copying objects from cacheline 0 */
   nb0->current_data = b0->current_data;
-  nb0->total_length_not_including_first_buffer = 0;
+  nb0->current_length = length;
   nb0->flags = VLIB_BUFFER_TOTAL_LENGTH_VALID | flags;
-  nb0->trace_handle = b0->trace_handle;
+  nb0->flow_id = b0->flow_id;
+  nb0->error = b0->error;
+  nb0->current_config_index = b0->current_config_index;
   clib_memcpy_fast (&nb0->opaque, &b0->opaque, sizeof (nb0->opaque));
+
+  /* copying objects from cacheline 1 */
+  nb0->trace_handle = b0->trace_handle;
+  nb0->total_length_not_including_first_buffer = 0;
+
+  /* copying data */
   clib_memcpy_fast (vlib_buffer_get_current (nb0),
 		    vlib_buffer_get_current (b0), length);
-  nb0->current_length = length;
 }
 
 static_always_inline void
@@ -209,8 +261,7 @@
 
 static_always_inline void
 tso_fixup_segmented_buf (vlib_main_t * vm, vlib_buffer_t * b0, u8 tcp_flags,
-			 int is_ip6, generic_header_offset_t * gho,
-			 u32 do_csums)
+			 int is_ip6, generic_header_offset_t * gho)
 {
   ip4_header_t *ip4 =
     (ip4_header_t *) (vlib_buffer_get_current (b0) + gho->l3_hdr_offset);
@@ -224,9 +275,8 @@
   if (is_ip6)
     {
       ip6->payload_length =
-	clib_host_to_net_u16 (b0->current_length -
-			      (gho->l4_hdr_offset - gho->l2_hdr_offset));
-      if (do_csums && (b0->flags & VNET_BUFFER_F_OFFLOAD_TCP_CKSUM))
+	clib_host_to_net_u16 (b0->current_length - gho->l4_hdr_offset);
+      if (gho->gho_flags & GHO_F_TCP)
 	{
 	  int bogus = 0;
 	  tcp->checksum = 0;
@@ -238,20 +288,29 @@
   else
     {
       ip4->length =
-	clib_host_to_net_u16 (b0->current_length -
-			      (gho->l3_hdr_offset - gho->l2_hdr_offset));
-      if (do_csums)
+	clib_host_to_net_u16 (b0->current_length - gho->l3_hdr_offset);
+      if (gho->gho_flags & GHO_F_IP4)
+	ip4->checksum = ip4_header_checksum (ip4);
+      if (gho->gho_flags & GHO_F_TCP)
 	{
-	  if (b0->flags & VNET_BUFFER_F_OFFLOAD_IP_CKSUM)
-	    ip4->checksum = ip4_header_checksum (ip4);
-	  if (b0->flags & VNET_BUFFER_F_OFFLOAD_TCP_CKSUM)
-	    {
-	      tcp->checksum = 0;
-	      tcp->checksum = ip4_tcp_udp_compute_checksum (vm, b0, ip4);
-	    }
-	  b0->flags &= ~VNET_BUFFER_F_OFFLOAD_TCP_CKSUM;
-	  b0->flags &= ~VNET_BUFFER_F_OFFLOAD_IP_CKSUM;
+	  tcp->checksum = 0;
+	  tcp->checksum = ip4_tcp_udp_compute_checksum (vm, b0, ip4);
 	}
+      b0->flags &= ~VNET_BUFFER_F_OFFLOAD_TCP_CKSUM;
+      b0->flags &= ~VNET_BUFFER_F_OFFLOAD_IP_CKSUM;
+    }
+
+  if ((gho->gho_flags & GHO_F_TUNNEL) == 0)
+    {
+      u32 adj_index0 = vnet_buffer (b0)->ip.adj_index[VLIB_TX];
+
+      ip_adjacency_t *adj0 = adj_get (adj_index0);
+
+      if (adj0->lookup_next_index == IP_LOOKUP_NEXT_MIDCHAIN &&
+	  adj0->sub_type.midchain.fixup_func)
+	/* calls e.g. ipip44_fixup */
+	adj0->sub_type.midchain.fixup_func
+	  (vm, adj0, b0, adj0->sub_type.midchain.fixup_data);
     }
 }
 
@@ -267,8 +326,7 @@
 static_always_inline u32
 tso_segment_buffer (vlib_main_t * vm, vnet_interface_per_thread_data_t * ptd,
 		    u32 sbi0, vlib_buffer_t * sb0,
-		    generic_header_offset_t * gho, u32 n_bytes_b0, int is_ip6,
-		    u32 do_csums)
+		    generic_header_offset_t * gho, u32 n_bytes_b0, int is_ip6)
 {
   u32 n_tx_bytes = 0;
   u16 gso_size = vnet_buffer2 (sb0)->gso_size;
@@ -316,8 +374,7 @@
       src_ptr = vlib_buffer_get_current (sb0) + l234_sz + first_data_size;
       src_left = sb0->current_length - l234_sz - first_data_size;
 
-      tso_fixup_segmented_buf (vm, b0, tcp_flags_no_fin_psh, is_ip6, gho,
-			       do_csums);
+      tso_fixup_segmented_buf (vm, b0, tcp_flags_no_fin_psh, is_ip6, gho);
 
       /* grab a second buffer and prepare the loop */
       ASSERT (dbi < vec_len (ptd->split_buffers));
@@ -367,7 +424,7 @@
 	    {
 	      n_tx_bytes += cdb0->current_length;
 	      tso_fixup_segmented_buf (vm, cdb0, tcp_flags_no_fin_psh, is_ip6,
-				       gho, do_csums);
+				       gho);
 	      ASSERT (dbi < vec_len (ptd->split_buffers));
 	      cdb0 = vlib_get_buffer (vm, ptd->split_buffers[dbi++]);
 	      tso_init_buf_from_template (vm, cdb0, b0, l234_sz,
@@ -376,8 +433,7 @@
 	    }
 	}
 
-      tso_fixup_segmented_buf (vm, cdb0, save_tcp_flags, is_ip6, gho,
-			       do_csums);
+      tso_fixup_segmented_buf (vm, cdb0, save_tcp_flags, is_ip6, gho);
 
       n_tx_bytes += cdb0->current_length;
     }
@@ -411,7 +467,7 @@
 		      vlib_frame_t * frame,
 		      vnet_main_t * vnm,
 		      vnet_hw_interface_t * hi,
-		      int is_ip6, int do_segmentation)
+		      int is_l2, int is_ip4, int is_ip6, int do_segmentation)
 {
   u32 *to_next;
   u32 next_index = node->cached_next_index;
@@ -496,7 +552,8 @@
 		t0->flags = b[0]->flags & VNET_BUFFER_F_GSO;
 		t0->gso_size = vnet_buffer2 (b[0])->gso_size;
 		t0->gso_l4_hdr_sz = vnet_buffer2 (b[0])->gso_l4_hdr_sz;
-		vnet_generic_header_offset_parser (b[0], &t0->gho);
+		vnet_generic_header_offset_parser (b[0], &t0->gho, is_l2,
+						   is_ip4, is_ip6);
 	      }
 	    if (b[1]->flags & VLIB_BUFFER_IS_TRACED)
 	      {
@@ -504,7 +561,8 @@
 		t1->flags = b[1]->flags & VNET_BUFFER_F_GSO;
 		t1->gso_size = vnet_buffer2 (b[1])->gso_size;
 		t1->gso_l4_hdr_sz = vnet_buffer2 (b[1])->gso_l4_hdr_sz;
-		vnet_generic_header_offset_parser (b[1], &t1->gho);
+		vnet_generic_header_offset_parser (b[1], &t1->gho, is_l2,
+						   is_ip4, is_ip6);
 	      }
 	    if (b[2]->flags & VLIB_BUFFER_IS_TRACED)
 	      {
@@ -512,7 +570,8 @@
 		t2->flags = b[2]->flags & VNET_BUFFER_F_GSO;
 		t2->gso_size = vnet_buffer2 (b[2])->gso_size;
 		t2->gso_l4_hdr_sz = vnet_buffer2 (b[2])->gso_l4_hdr_sz;
-		vnet_generic_header_offset_parser (b[2], &t2->gho);
+		vnet_generic_header_offset_parser (b[2], &t2->gho, is_l2,
+						   is_ip4, is_ip6);
 	      }
 	    if (b[3]->flags & VLIB_BUFFER_IS_TRACED)
 	      {
@@ -520,7 +579,8 @@
 		t3->flags = b[3]->flags & VNET_BUFFER_F_GSO;
 		t3->gso_size = vnet_buffer2 (b[3])->gso_size;
 		t3->gso_l4_hdr_sz = vnet_buffer2 (b[3])->gso_l4_hdr_sz;
-		vnet_generic_header_offset_parser (b[3], &t3->gho);
+		vnet_generic_header_offset_parser (b[3], &t3->gho, is_l2,
+						   is_ip4, is_ip6);
 	      }
 
 	    from += 4;
@@ -547,7 +607,6 @@
 	  vnet_hw_interface_t *hi0;
 	  u32 next0 = 0;
 	  u32 do_segmentation0 = 0;
-	  u32 do_csums = 0;
 
 	  swif0 = vnet_buffer (b[0])->sw_if_index[VLIB_TX];
 	  if (PREDICT_FALSE (hi->sw_if_index != swif0))
@@ -573,7 +632,8 @@
 	      t0->flags = b[0]->flags & VNET_BUFFER_F_GSO;
 	      t0->gso_size = vnet_buffer2 (b[0])->gso_size;
 	      t0->gso_l4_hdr_sz = vnet_buffer2 (b[0])->gso_l4_hdr_sz;
-	      vnet_generic_header_offset_parser (b[0], &t0->gho);
+	      vnet_generic_header_offset_parser (b[0], &t0->gho, is_l2,
+						 is_ip4, is_ip6);
 	    }
 
 	  if (do_segmentation0)
@@ -592,12 +652,14 @@
 		  u32 n_bytes_b0 = vlib_buffer_length_in_chain (vm, b[0]);
 		  u32 n_tx_bytes = 0;
 
-		  vnet_generic_header_offset_parser (b[0], &gho);
+		  vnet_generic_header_offset_parser (b[0], &gho, is_l2,
+						     is_ip4, is_ip6);
 
 		  if (PREDICT_FALSE (gho.gho_flags & GHO_F_TUNNEL))
 		    {
 		      if (PREDICT_FALSE
-			  ((gho.gho_flags & GHO_F_VXLAN_TUNNEL) == 0))
+			  (gho.gho_flags & (GHO_F_GRE_TUNNEL |
+					    GHO_F_GENEVE_TUNNEL)))
 			{
 			  /* not supported yet */
 			  drop_one_buffer_and_count (vm, vnm, node, from - 1,
@@ -610,17 +672,12 @@
 		      vnet_get_inner_header (b[0], &gho);
 
 		      n_bytes_b0 -= gho.outer_hdr_sz;
-		      /*
-		       * In case of tunnel encapsulated packet, we will
-		       * calculate the checksums for segmented inner packets.
-		       */
-		      do_csums = 1;
 		      is_ip6 = (gho.gho_flags & GHO_F_IP6) != 0;
 		    }
 
 		  n_tx_bytes =
 		    tso_segment_buffer (vm, ptd, bi0, b[0], &gho, n_bytes_b0,
-					is_ip6, do_csums);
+					is_ip6);
 
 		  if (PREDICT_FALSE (n_tx_bytes == 0))
 		    {
@@ -631,12 +688,22 @@
 		      continue;
 		    }
 
+
 		  if (PREDICT_FALSE (gho.gho_flags & GHO_F_VXLAN_TUNNEL))
 		    {
 		      vnet_get_outer_header (b[0], &gho);
 		      n_tx_bytes +=
 			tso_segment_vxlan_tunnel_fixup (vm, ptd, b[0], &gho);
 		    }
+		  else
+		    if (PREDICT_FALSE
+			(gho.gho_flags & (GHO_F_IPIP_TUNNEL |
+					  GHO_F_IPIP6_TUNNEL)))
+		    {
+		      vnet_get_outer_header (b[0], &gho);
+		      n_tx_bytes +=
+			tso_segment_ipip_tunnel_fixup (vm, ptd, b[0], &gho);
+		    }
 
 		  u16 n_tx_bufs = vec_len (ptd->split_buffers);
 		  u32 *from_seg = ptd->split_buffers;
@@ -645,30 +712,7 @@
 		    {
 		      u32 sbi0;
 		      vlib_buffer_t *sb0;
-		      if (n_tx_bufs >= n_left_to_next)
-			{
-			  while (n_left_to_next > 0)
-			    {
-			      sbi0 = to_next[0] = from_seg[0];
-			      sb0 = vlib_get_buffer (vm, sbi0);
-			      ASSERT (sb0->current_length > 0);
-			      to_next += 1;
-			      from_seg += 1;
-			      n_left_to_next -= 1;
-			      n_tx_bufs -= 1;
-			      vnet_feature_next (&next0, sb0);
-			      vlib_validate_buffer_enqueue_x1 (vm, node,
-							       next_index,
-							       to_next,
-							       n_left_to_next,
-							       sbi0, next0);
-			    }
-			  vlib_put_next_frame (vm, node, next_index,
-					       n_left_to_next);
-			  vlib_get_new_next_frame (vm, node, next_index,
-						   to_next, n_left_to_next);
-			}
-		      while (n_tx_bufs > 0)
+		      while (n_tx_bufs > 0 && n_left_to_next > 0)
 			{
 			  sbi0 = to_next[0] = from_seg[0];
 			  sb0 = vlib_get_buffer (vm, sbi0);
@@ -677,6 +721,7 @@
 			  from_seg += 1;
 			  n_left_to_next -= 1;
 			  n_tx_bufs -= 1;
+			  next0 = 0;
 			  vnet_feature_next (&next0, sb0);
 			  vlib_validate_buffer_enqueue_x1 (vm, node,
 							   next_index,
@@ -684,6 +729,11 @@
 							   n_left_to_next,
 							   sbi0, next0);
 			}
+		      vlib_put_next_frame (vm, node, next_index,
+					   n_left_to_next);
+		      if (n_tx_bufs > 0)
+			vlib_get_next_frame (vm, node, next_index,
+					     to_next, n_left_to_next);
 		    }
 		  /* The buffers were enqueued. Reset the length */
 		  _vec_len (ptd->split_buffers) = 0;
@@ -707,7 +757,8 @@
 
 static_always_inline uword
 vnet_gso_inline (vlib_main_t * vm,
-		 vlib_node_runtime_t * node, vlib_frame_t * frame, int is_ip6)
+		 vlib_node_runtime_t * node, vlib_frame_t * frame, int is_l2,
+		 int is_ip4, int is_ip6)
 {
   vnet_main_t *vnm = vnet_get_main ();
   vnet_hw_interface_t *hi;
@@ -721,10 +772,12 @@
 
       if (hi->flags & VNET_HW_INTERFACE_FLAG_SUPPORTS_GSO)
 	return vnet_gso_node_inline (vm, node, frame, vnm, hi,
-				     is_ip6, /* do_segmentation */ 0);
+				     is_l2, is_ip4, is_ip6,
+				     /* do_segmentation */ 0);
       else
 	return vnet_gso_node_inline (vm, node, frame, vnm, hi,
-				     is_ip6, /* do_segmentation */ 1);
+				     is_l2, is_ip4, is_ip6,
+				     /* do_segmentation */ 1);
     }
   return 0;
 }
@@ -732,25 +785,29 @@
 VLIB_NODE_FN (gso_l2_ip4_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
 				vlib_frame_t * frame)
 {
-  return vnet_gso_inline (vm, node, frame, 0 /* ip6 */ );
+  return vnet_gso_inline (vm, node, frame, 1 /* l2 */ , 1 /* ip4 */ ,
+			  0 /* ip6 */ );
 }
 
 VLIB_NODE_FN (gso_l2_ip6_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
 				vlib_frame_t * frame)
 {
-  return vnet_gso_inline (vm, node, frame, 1 /* ip6 */ );
+  return vnet_gso_inline (vm, node, frame, 1 /* l2 */ , 0 /* ip4 */ ,
+			  1 /* ip6 */ );
 }
 
 VLIB_NODE_FN (gso_ip4_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
 			     vlib_frame_t * frame)
 {
-  return vnet_gso_inline (vm, node, frame, 0 /* ip6 */ );
+  return vnet_gso_inline (vm, node, frame, 0 /* l2 */ , 1 /* ip4 */ ,
+			  0 /* ip6 */ );
 }
 
 VLIB_NODE_FN (gso_ip6_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
 			     vlib_frame_t * frame)
 {
-  return vnet_gso_inline (vm, node, frame, 1 /* ip6 */ );
+  return vnet_gso_inline (vm, node, frame, 0 /* l2 */ , 0 /* ip4 */ ,
+			  1 /* ip6 */ );
 }
 
 /* *INDENT-OFF* */
@@ -806,15 +863,13 @@
 VNET_FEATURE_INIT (gso_ip4_node, static) = {
   .arc_name = "ip4-output",
   .node_name = "gso-ip4",
-  .runs_after = VNET_FEATURES ("ipsec4-output-feature"),
-  .runs_before = VNET_FEATURES ("interface-output"),
+  .runs_before = VNET_FEATURES ("esp4-encrypt-tun","ipsec4-output-feature"),
 };
 
 VNET_FEATURE_INIT (gso_ip6_node, static) = {
   .arc_name = "ip6-output",
   .node_name = "gso-ip6",
-  .runs_after = VNET_FEATURES ("ipsec6-output-feature"),
-  .runs_before = VNET_FEATURES ("interface-output"),
+  .runs_before = VNET_FEATURES ("esp6-encrypt-tun","ipsec6-output-feature"),
 };
 
 /*