gso: add support for IP-IP

Type: feature

Change-Id: I37752af8496e0042a1da91124f3d94216b39ff11
Signed-off-by: Mohsin Kazmi <sykazmi@cisco.com>

7 files changed, 266 insertions, 138 deletions

diff --git a/src/vnet/devices/virtio/device.c b/src/vnet/devices/virtio/device.c
index d110946dede..9a35bb8162f 100644
--- a/src/vnet/devices/virtio/device.c
+++ b/src/vnet/devices/virtio/device.c
@@ -168,13 +168,15 @@ virtio_free_used_device_desc (vlib_main_t * vm, virtio_vring_t * vring,
 }
 
 static_always_inline void
-set_checksum_offsets (vlib_buffer_t * b, struct virtio_net_hdr_v1 *hdr)
+set_checksum_offsets (vlib_buffer_t * b, struct virtio_net_hdr_v1 *hdr,
+		      int is_l2)
 {
   if (b->flags & VNET_BUFFER_F_IS_IP4)
     {
       ip4_header_t *ip4;
       generic_header_offset_t gho = { 0 };
-      vnet_generic_header_offset_parser (b, &gho);
+      vnet_generic_header_offset_parser (b, &gho, is_l2, 1 /* ip4 */ ,
+					 0 /* ip6 */ );
       hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
       hdr->csum_start = gho.l4_hdr_offset;	// 0x22;
       if (b->flags & VNET_BUFFER_F_OFFLOAD_TCP_CKSUM)
@@ -198,7 +200,8 @@ set_checksum_offsets (vlib_buffer_t * b, struct virtio_net_hdr_v1 *hdr)
   else if (b->flags & VNET_BUFFER_F_IS_IP6)
     {
       generic_header_offset_t gho = { 0 };
-      vnet_generic_header_offset_parser (b, &gho);
+      vnet_generic_header_offset_parser (b, &gho, is_l2, 0 /* ip4 */ ,
+					 1 /* ip6 */ );
       hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
       hdr->csum_start = gho.l4_hdr_offset;	// 0x36;
       if (b->flags & VNET_BUFFER_F_OFFLOAD_TCP_CKSUM)
@@ -213,13 +216,14 @@ set_checksum_offsets (vlib_buffer_t * b, struct virtio_net_hdr_v1 *hdr)
 }
 
 static_always_inline void
-set_gso_offsets (vlib_buffer_t * b, struct virtio_net_hdr_v1 *hdr)
+set_gso_offsets (vlib_buffer_t * b, struct virtio_net_hdr_v1 *hdr, int is_l2)
 {
   if (b->flags & VNET_BUFFER_F_IS_IP4)
     {
       ip4_header_t *ip4;
       generic_header_offset_t gho = { 0 };
-      vnet_generic_header_offset_parser (b, &gho);
+      vnet_generic_header_offset_parser (b, &gho, is_l2, 1 /* ip4 */ ,
+					 0 /* ip6 */ );
       hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
       hdr->gso_size = vnet_buffer2 (b)->gso_size;
       hdr->hdr_len = gho.hdr_sz;
@@ -238,7 +242,8 @@ set_gso_offsets (vlib_buffer_t * b, struct virtio_net_hdr_v1 *hdr)
   else if (b->flags & VNET_BUFFER_F_IS_IP6)
     {
       generic_header_offset_t gho = { 0 };
-      vnet_generic_header_offset_parser (b, &gho);
+      vnet_generic_header_offset_parser (b, &gho, is_l2, 0 /* ip4 */ ,
+					 1 /* ip6 */ );
       hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
       hdr->gso_size = vnet_buffer2 (b)->gso_size;
       hdr->hdr_len = gho.hdr_sz;
@@ -260,13 +265,14 @@ add_buffer_to_slot (vlib_main_t * vm, virtio_if_t * vif,
   d = &vring->desc[next];
   vlib_buffer_t *b = vlib_get_buffer (vm, bi);
   struct virtio_net_hdr_v1 *hdr = vlib_buffer_get_current (b) - hdr_sz;
+  int is_l2 = (vif->type & (VIRTIO_IF_TYPE_TAP | VIRTIO_IF_TYPE_PCI));
 
   clib_memset (hdr, 0, hdr_sz);
 
   if (b->flags & VNET_BUFFER_F_GSO)
     {
       if (do_gso)
-	set_gso_offsets (b, hdr);
+	set_gso_offsets (b, hdr, is_l2);
       else
 	{
 	  virtio_interface_drop_inline (vm, node_index, &bi, 1,
@@ -278,7 +284,7 @@ add_buffer_to_slot (vlib_main_t * vm, virtio_if_t * vif,
 		       VNET_BUFFER_F_OFFLOAD_UDP_CKSUM))
     {
       if (csum_offload)
-	set_checksum_offsets (b, hdr);
+	set_checksum_offsets (b, hdr, is_l2);
       else
 	{
 	  virtio_interface_drop_inline (vm, node_index, &bi, 1,
diff --git a/src/vnet/devices/virtio/vhost_user_output.c b/src/vnet/devices/virtio/vhost_user_output.c
index 4f5eb3c1d76..d48e43738a9 100644
--- a/src/vnet/devices/virtio/vhost_user_output.c
+++ b/src/vnet/devices/virtio/vhost_user_output.c
@@ -237,7 +237,11 @@ vhost_user_handle_tx_offload (vhost_user_intf_t * vui, vlib_buffer_t * b,
 			      virtio_net_hdr_t * hdr)
 {
   generic_header_offset_t gho = { 0 };
-  vnet_generic_header_offset_parser (b, &gho);
+  int is_ip4 = b->flags & VNET_BUFFER_F_IS_IP4;
+  int is_ip6 = b->flags & VNET_BUFFER_F_IS_IP6;
+
+  ASSERT (!(is_ip4 && is_ip6));
+  vnet_generic_header_offset_parser (b, &gho, 1 /* l2 */ , is_ip4, is_ip6);
   if (b->flags & VNET_BUFFER_F_OFFLOAD_IP_CKSUM)
     {
       ip4_header_t *ip4;
@@ -272,13 +276,13 @@ vhost_user_handle_tx_offload (vhost_user_intf_t * vui, vlib_buffer_t * b,
     {
       if (b->flags & VNET_BUFFER_F_OFFLOAD_TCP_CKSUM)
 	{
-	  if ((b->flags & VNET_BUFFER_F_IS_IP4) &&
+	  if (is_ip4 &&
 	      (vui->features & (1ULL << FEAT_VIRTIO_NET_F_GUEST_TSO4)))
 	    {
 	      hdr->gso_size = vnet_buffer2 (b)->gso_size;
 	      hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
 	    }
-	  else if ((b->flags & VNET_BUFFER_F_IS_IP6) &&
+	  else if (is_ip6 &&
 		   (vui->features & (1ULL << FEAT_VIRTIO_NET_F_GUEST_TSO6)))
 	    {
 	      hdr->gso_size = vnet_buffer2 (b)->gso_size;
diff --git a/src/vnet/gso/FEATURE.yaml b/src/vnet/gso/FEATURE.yaml
index b3296d6efa6..79b506df69c 100644
--- a/src/vnet/gso/FEATURE.yaml
+++ b/src/vnet/gso/FEATURE.yaml
@@ -5,9 +5,11 @@ features:
   - Basic GSO support
   - GSO for VLAN tagged packets
   - GSO for VXLAN tunnel
+  - GSO for IP-IP tunnel
+  - GSO for IPSec tunnel  
   - Provide inline function to get header offsets
 description: "Generic Segmentation Offload"
 missing:
-  - Thorough Testing, IPIP, GRE, Geneve, IPSec
+  - Thorough Testing, GRE, Geneve
 state: experimental
 properties: [API, CLI]
diff --git a/src/vnet/gso/gso.c b/src/vnet/gso/gso.c
index c741b17bdd4..854bffb39e3 100644
--- a/src/vnet/gso/gso.c
+++ b/src/vnet/gso/gso.c
@@ -26,30 +26,6 @@ gso_main_t gso_main;
 int
 vnet_sw_interface_gso_enable_disable (u32 sw_if_index, u8 enable)
 {
-  ethernet_interface_t *eif;
-  vnet_sw_interface_t *si;
-  ethernet_main_t *em;
-  vnet_main_t *vnm;
-
-  vnm = vnet_get_main ();
-  em = &ethernet_main;
-  si = vnet_get_sw_interface (vnm, sw_if_index);
-
-  /*
-   * only ethernet HW interfaces are supported at this time
-   */
-  if (si->type != VNET_SW_INTERFACE_TYPE_HARDWARE)
-    {
-      return (VNET_API_ERROR_INVALID_VALUE);
-    }
-
-  eif = ethernet_get_interface (em, si->hw_if_index);
-
-  if (!eif)
-    {
-      return (VNET_API_ERROR_FEATURE_DISABLED);
-    }
-
   vnet_feature_enable_disable ("ip4-output", "gso-ip4", sw_if_index, enable,
 			       0, 0);
   vnet_feature_enable_disable ("ip6-output", "gso-ip6", sw_if_index, enable,
diff --git a/src/vnet/gso/hdr_offset_parser.h b/src/vnet/gso/hdr_offset_parser.h
index 8f9e0cda984..b194333e780 100644
--- a/src/vnet/gso/hdr_offset_parser.h
+++ b/src/vnet/gso/hdr_offset_parser.h
@@ -36,7 +36,8 @@
   _( 8, VXLAN_TUNNEL)           \
   _( 9, GRE_TUNNEL)             \
   _( 10, IPIP_TUNNEL)           \
-  _( 11, GENEVE_TUNNEL)
+  _( 11, IPIP6_TUNNEL)          \
+  _( 12, GENEVE_TUNNEL)
 
 typedef enum gho_flag_t_
 {
@@ -45,9 +46,10 @@ typedef enum gho_flag_t_
 #undef _
 } gho_flag_t;
 
-#define GHO_F_TUNNEL (GHO_F_VXLAN_TUNNEL |  \
-                      GHO_F_GENEVE_TUNNEL | \
-                      GHO_F_IPIP_TUNNEL |   \
+#define GHO_F_TUNNEL (GHO_F_VXLAN_TUNNEL  |  \
+                      GHO_F_GENEVE_TUNNEL |  \
+                      GHO_F_IPIP_TUNNEL   |  \
+                      GHO_F_IPIP6_TUNNEL  |  \
                       GHO_F_GRE_TUNNEL)
 
 #define GHO_F_OUTER_HDR (GHO_F_OUTER_IP4 | \
@@ -173,11 +175,77 @@ static_always_inline void
 vnet_ipip_inner_header_parser_inline (vlib_buffer_t * b0,
 				      generic_header_offset_t * gho)
 {
-  /* not supported yet */
-  if ((gho->gho_flags & GHO_F_IPIP_TUNNEL) == 0)
+  if ((gho->gho_flags & (GHO_F_IPIP_TUNNEL | GHO_F_IPIP6_TUNNEL)) == 0)
     return;
 
-  ASSERT (0);
+  u8 l4_proto = 0;
+  u8 l4_hdr_sz = 0;
+
+  gho->outer_l2_hdr_offset = gho->l2_hdr_offset;
+  gho->outer_l3_hdr_offset = gho->l3_hdr_offset;
+  gho->outer_l4_hdr_offset = gho->l4_hdr_offset;
+  gho->outer_l4_hdr_sz = gho->l4_hdr_sz;
+  gho->outer_hdr_sz = gho->hdr_sz;
+
+  gho->l2_hdr_offset = 0;
+  gho->l3_hdr_offset = 0;
+  gho->l4_hdr_offset = 0;
+  gho->l4_hdr_sz = 0;
+  gho->hdr_sz = 0;
+
+  if (gho->gho_flags & GHO_F_IP4)
+    {
+      gho->gho_flags |= GHO_F_OUTER_IP4;
+    }
+  else if (gho->gho_flags & GHO_F_IP6)
+    {
+      gho->gho_flags |= GHO_F_OUTER_IP6;
+    }
+
+  gho->gho_flags &= ~GHO_F_INNER_HDR;
+
+  vnet_get_inner_header (b0, gho);
+
+  gho->l2_hdr_offset = b0->current_data;
+  gho->l3_hdr_offset = 0;
+
+  if (PREDICT_TRUE (gho->gho_flags & GHO_F_IPIP_TUNNEL))
+    {
+      ip4_header_t *ip4 = (ip4_header_t *) vlib_buffer_get_current (b0);
+      gho->l4_hdr_offset = ip4_header_bytes (ip4);
+      l4_proto = ip4->protocol;
+      gho->gho_flags |= GHO_F_IP4;
+    }
+  else if (PREDICT_TRUE (gho->gho_flags & GHO_F_IPIP6_TUNNEL))
+    {
+      ip6_header_t *ip6 = (ip6_header_t *) vlib_buffer_get_current (b0);
+      /* FIXME IPv6 EH traversal */
+      gho->l4_hdr_offset = sizeof (ip6_header_t);
+      l4_proto = ip6->protocol;
+      gho->gho_flags |= GHO_F_IP6;
+    }
+  if (l4_proto == IP_PROTOCOL_TCP)
+    {
+      tcp_header_t *tcp = (tcp_header_t *) (vlib_buffer_get_current (b0) +
+					    gho->l4_hdr_offset);
+      l4_hdr_sz = tcp_header_bytes (tcp);
+
+      gho->gho_flags |= GHO_F_TCP;
+
+    }
+  else if (l4_proto == IP_PROTOCOL_UDP)
+    {
+      udp_header_t *udp = (udp_header_t *) (vlib_buffer_get_current (b0) +
+					    gho->l4_hdr_offset);
+      l4_hdr_sz = sizeof (*udp);
+
+      gho->gho_flags |= GHO_F_UDP;
+    }
+
+  gho->l4_hdr_sz = l4_hdr_sz;
+  gho->hdr_sz += gho->l4_hdr_offset + l4_hdr_sz;
+
+  vnet_get_outer_header (b0, gho);
 }
 
 static_always_inline void
@@ -290,7 +358,7 @@ vnet_generic_inner_header_parser_inline (vlib_buffer_t * b0,
 
   if (gho->gho_flags & GHO_F_VXLAN_TUNNEL)
     vnet_vxlan_inner_header_parser_inline (b0, gho);
-  else if (gho->gho_flags & GHO_F_IPIP_TUNNEL)
+  else if (gho->gho_flags & (GHO_F_IPIP_TUNNEL | GHO_F_IPIP6_TUNNEL))
     vnet_ipip_inner_header_parser_inline (b0, gho);
   else if (gho->gho_flags & GHO_F_GRE_TUNNEL)
     vnet_gre_inner_header_parser_inline (b0, gho);
@@ -300,33 +368,44 @@ vnet_generic_inner_header_parser_inline (vlib_buffer_t * b0,
 
 static_always_inline void
 vnet_generic_outer_header_parser_inline (vlib_buffer_t * b0,
-					 generic_header_offset_t * gho)
+					 generic_header_offset_t * gho,
+					 int is_l2, int is_ip4, int is_ip6)
 {
   u8 l4_proto = 0;
   u8 l4_hdr_sz = 0;
+  u16 ethertype = 0;
+  u16 l2hdr_sz = 0;
 
-  ethernet_header_t *eh = (ethernet_header_t *) vlib_buffer_get_current (b0);
-  u16 ethertype = clib_net_to_host_u16 (eh->type);
-  u16 l2hdr_sz = sizeof (ethernet_header_t);
+  ASSERT (is_ip4 ^ is_ip6);
 
-  if (ethernet_frame_is_tagged (ethertype))
+  if (is_l2)
     {
-      ethernet_vlan_header_t *vlan = (ethernet_vlan_header_t *) (eh + 1);
+      ethernet_header_t *eh =
+	(ethernet_header_t *) vlib_buffer_get_current (b0);
+      ethertype = clib_net_to_host_u16 (eh->type);
+      l2hdr_sz = sizeof (ethernet_header_t);
 
-      ethertype = clib_net_to_host_u16 (vlan->type);
-      l2hdr_sz += sizeof (*vlan);
-      if (ethertype == ETHERNET_TYPE_VLAN)
+      if (ethernet_frame_is_tagged (ethertype))
 	{
-	  vlan++;
+	  ethernet_vlan_header_t *vlan = (ethernet_vlan_header_t *) (eh + 1);
+
 	  ethertype = clib_net_to_host_u16 (vlan->type);
 	  l2hdr_sz += sizeof (*vlan);
+	  if (ethertype == ETHERNET_TYPE_VLAN)
+	    {
+	      vlan++;
+	      ethertype = clib_net_to_host_u16 (vlan->type);
+	      l2hdr_sz += sizeof (*vlan);
+	    }
 	}
     }
+  else
+    l2hdr_sz = vnet_buffer (b0)->ip.save_rewrite_length;
 
   gho->l2_hdr_offset = b0->current_data;
   gho->l3_hdr_offset = l2hdr_sz;
 
-  if (PREDICT_TRUE (ethertype == ETHERNET_TYPE_IP4))
+  if (PREDICT_TRUE (is_ip4))
     {
       ip4_header_t *ip4 =
 	(ip4_header_t *) (vlib_buffer_get_current (b0) + l2hdr_sz);
@@ -334,7 +413,7 @@ vnet_generic_outer_header_parser_inline (vlib_buffer_t * b0,
       l4_proto = ip4->protocol;
       gho->gho_flags |= GHO_F_IP4;
     }
-  else if (PREDICT_TRUE (ethertype == ETHERNET_TYPE_IP6))
+  else if (PREDICT_TRUE (is_ip6))
     {
       ip6_header_t *ip6 =
 	(ip6_header_t *) (vlib_buffer_get_current (b0) + l2hdr_sz);
@@ -369,12 +448,16 @@ vnet_generic_outer_header_parser_inline (vlib_buffer_t * b0,
 	  gho->gho_flags |= GHO_F_GENEVE_TUNNEL;
 	}
     }
-  else if ((l4_proto == IP_PROTOCOL_IP_IN_IP)
-	   || (l4_proto == IP_PROTOCOL_IPV6))
+  else if (l4_proto == IP_PROTOCOL_IP_IN_IP)
     {
       l4_hdr_sz = 0;
       gho->gho_flags |= GHO_F_IPIP_TUNNEL;
     }
+  else if (l4_proto == IP_PROTOCOL_IPV6)
+    {
+      l4_hdr_sz = 0;
+      gho->gho_flags |= GHO_F_IPIP6_TUNNEL;
+    }
   else if (l4_proto == IP_PROTOCOL_GRE)
     {
       l4_hdr_sz = 0;
@@ -387,9 +470,10 @@ vnet_generic_outer_header_parser_inline (vlib_buffer_t * b0,
 
 static_always_inline void
 vnet_generic_header_offset_parser (vlib_buffer_t * b0,
-				   generic_header_offset_t * gho)
+				   generic_header_offset_t * gho, int is_l2,
+				   int is_ip4, int is_ip6)
 {
-  vnet_generic_outer_header_parser_inline (b0, gho);
+  vnet_generic_outer_header_parser_inline (b0, gho, is_l2, is_ip4, is_ip6);
 
   if (gho->gho_flags & GHO_F_TUNNEL)
     {
diff --git a/src/vnet/gso/node.c b/src/vnet/gso/node.c
index 663f9cc602d..b57830f1ac5 100644
--- a/src/vnet/gso/node.c
+++ b/src/vnet/gso/node.c
@@ -56,6 +56,50 @@ format_gso_trace (u8 * s, va_list * args)
   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 @@ static_always_inline void
 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 @@ tso_init_buf_from_template (vlib_main_t * vm, vlib_buffer_t * nb0,
 
 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 @@ tso_fixup_segmented_buf (vlib_main_t * vm, vlib_buffer_t * b0, u8 tcp_flags,
   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 @@ tso_fixup_segmented_buf (vlib_main_t * vm, vlib_buffer_t * b0, u8 tcp_flags,
   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 @@ tso_fixup_segmented_buf (vlib_main_t * vm, vlib_buffer_t * b0, u8 tcp_flags,
 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 @@ tso_segment_buffer (vlib_main_t * vm, vnet_interface_per_thread_data_t * ptd,
       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 @@ tso_segment_buffer (vlib_main_t * vm, vnet_interface_per_thread_data_t * ptd,
 	    {
 	      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_segment_buffer (vlib_main_t * vm, vnet_interface_per_thread_data_t * ptd,
 	    }
 	}
 
-      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 @@ vnet_gso_node_inline (vlib_main_t * vm,
 		      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 @@ vnet_gso_node_inline (vlib_main_t * vm,
 		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 @@ vnet_gso_node_inline (vlib_main_t * vm,
 		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 @@ vnet_gso_node_inline (vlib_main_t * vm,
 		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 @@ vnet_gso_node_inline (vlib_main_t * vm,
 		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_gso_node_inline (vlib_main_t * vm,
 	  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 @@ vnet_gso_node_inline (vlib_main_t * vm,
 	      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 @@ vnet_gso_node_inline (vlib_main_t * vm,
 		  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_gso_node_inline (vlib_main_t * vm,
 		      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 @@ vnet_gso_node_inline (vlib_main_t * vm,
 		      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 @@ vnet_gso_node_inline (vlib_main_t * vm,
 		    {
 		      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 @@ vnet_gso_node_inline (vlib_main_t * vm,
 			  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 @@ vnet_gso_node_inline (vlib_main_t * vm,
 							   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 @@ vnet_gso_node_inline (vlib_main_t * vm,
 
 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 @@ vnet_gso_inline (vlib_main_t * vm,
 
       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 @@ vnet_gso_inline (vlib_main_t * vm,
 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_l2_ip6_node, static) = {
 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"),
 };
 
 /*
diff --git a/src/vnet/interface_output.h b/src/vnet/interface_output.h
index ef7aaa8999e..c408fca9ca9 100644
--- a/src/vnet/interface_output.h
+++ b/src/vnet/interface_output.h
@@ -92,7 +92,8 @@ vnet_calc_checksums_inline (vlib_main_t * vm, vlib_buffer_t * b,
   if (with_gso)
     {
       generic_header_offset_t gho = { 0 };
-      vnet_generic_header_offset_parser (b, &gho);
+      vnet_generic_header_offset_parser (b, &gho, 1 /* l2 */ , is_ip4,
+					 is_ip6);
 
       ASSERT (gho.gho_flags ^ (GHO_F_IP4 | GHO_F_IP6));