devices: add support for pseudo header checksum

Type: improvement

Linux uses pseudo header checksum when checksum of l4 is offloaded.
This patch adds similar support in virtual interfaces.

Change-Id: I6a94d1104e59356f95057e7c122e3be9cd8659a3
Signed-off-by: Aloys Augustin <aloaugus@cisco.com>
Signed-off-by: Mohsin Kazmi <sykazmi@cisco.com>

5 files changed, 104 insertions, 15 deletions

diff --git a/src/vnet/CMakeLists.txt b/src/vnet/CMakeLists.txt
index 7b01d941d38..5e1cbc24fd1 100644
--- a/src/vnet/CMakeLists.txt
+++ b/src/vnet/CMakeLists.txt
@@ -458,6 +458,7 @@ list(APPEND VNET_HEADERS
   ip/ip_table.h
   ip/ip_interface.h
   ip/ip_packet.h
+  ip/ip_psh_cksum.h
   ip/ip_source_and_port_range_check.h
   ip/ip_types.h
   ip/lookup.h
diff --git a/src/vnet/devices/af_packet/node.c b/src/vnet/devices/af_packet/node.c
index caddcfa416b..e2f87b10b35 100644
--- a/src/vnet/devices/af_packet/node.c
+++ b/src/vnet/devices/af_packet/node.c
@@ -132,7 +132,6 @@ mark_tcp_udp_cksum_calc (vlib_buffer_t *b, u8 *l4_hdr_sz)
 	  tcp_header_t *tcp = (tcp_header_t *) (vlib_buffer_get_current (b) +
 						sizeof (ethernet_header_t) +
 						ip4_header_bytes (ip4));
-	  tcp->checksum = 0;
 	  *l4_hdr_sz = tcp_header_bytes (tcp);
 	}
       else if (ip4->protocol == IP_PROTOCOL_UDP)
@@ -141,7 +140,6 @@ mark_tcp_udp_cksum_calc (vlib_buffer_t *b, u8 *l4_hdr_sz)
 	  udp_header_t *udp = (udp_header_t *) (vlib_buffer_get_current (b) +
 						sizeof (ethernet_header_t) +
 						ip4_header_bytes (ip4));
-	  udp->checksum = 0;
 	  *l4_hdr_sz = sizeof (*udp);
 	}
       vnet_buffer (b)->l3_hdr_offset = sizeof (ethernet_header_t);
@@ -172,7 +170,6 @@ mark_tcp_udp_cksum_calc (vlib_buffer_t *b, u8 *l4_hdr_sz)
 	  tcp_header_t *tcp =
 	    (tcp_header_t *) (vlib_buffer_get_current (b) +
 			      sizeof (ethernet_header_t) + ip6_hdr_len);
-	  tcp->checksum = 0;
 	  *l4_hdr_sz = tcp_header_bytes (tcp);
 	}
       else if (ip6->protocol == IP_PROTOCOL_UDP)
@@ -181,7 +178,6 @@ mark_tcp_udp_cksum_calc (vlib_buffer_t *b, u8 *l4_hdr_sz)
 	  udp_header_t *udp =
 	    (udp_header_t *) (vlib_buffer_get_current (b) +
 			      sizeof (ethernet_header_t) + ip6_hdr_len);
-	  udp->checksum = 0;
 	  *l4_hdr_sz = sizeof (*udp);
 	}
       vnet_buffer (b)->l3_hdr_offset = sizeof (ethernet_header_t);
diff --git a/src/vnet/devices/virtio/device.c b/src/vnet/devices/virtio/device.c
index 6c2fe34abf0..ac9be6b02ae 100644
--- a/src/vnet/devices/virtio/device.c
+++ b/src/vnet/devices/virtio/device.c
@@ -27,6 +27,7 @@
 #include <vnet/gso/hdr_offset_parser.h>
 #include <vnet/ip/ip4_packet.h>
 #include <vnet/ip/ip6_packet.h>
+#include <vnet/ip/ip_psh_cksum.h>
 #include <vnet/tcp/tcp_packet.h>
 #include <vnet/udp/udp_packet.h>
 #include <vnet/devices/virtio/virtio.h>
@@ -296,37 +297,60 @@ set_checksum_offsets (vlib_buffer_t *b, virtio_net_hdr_v1_t *hdr,
 					 0 /* ip6 */ );
       hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
       hdr->csum_start = gho.l4_hdr_offset;	// 0x22;
+
+      /*
+       * virtio devices do not support IP4 checksum offload. So driver takes
+       * care of it while doing tx.
+       */
+      ip4 = (ip4_header_t *) (vlib_buffer_get_current (b) + gho.l3_hdr_offset);
+      if (oflags & VNET_BUFFER_OFFLOAD_F_IP_CKSUM)
+	ip4->checksum = ip4_header_checksum (ip4);
+
+      /*
+       * virtio devices assume the l4 header is set to the checksum of the
+       * l3 pseudo-header, so we compute it before tx-ing
+       */
       if (oflags & VNET_BUFFER_OFFLOAD_F_TCP_CKSUM)
 	{
+	  tcp_header_t *tcp =
+	    (tcp_header_t *) (vlib_buffer_get_current (b) + gho.l4_hdr_offset);
+	  tcp->checksum = ip4_pseudo_header_cksum (ip4);
 	  hdr->csum_offset = STRUCT_OFFSET_OF (tcp_header_t, checksum);
 	}
       else if (oflags & VNET_BUFFER_OFFLOAD_F_UDP_CKSUM)
 	{
+	  udp_header_t *udp =
+	    (udp_header_t *) (vlib_buffer_get_current (b) + gho.l4_hdr_offset);
+	  udp->checksum = ip4_pseudo_header_cksum (ip4);
 	  hdr->csum_offset = STRUCT_OFFSET_OF (udp_header_t, checksum);
 	}
-
-      /*
-       * virtio devices do not support IP4 checksum offload. So driver takes care
-       * of it while doing tx.
-       */
-      ip4 =
-	(ip4_header_t *) (vlib_buffer_get_current (b) + gho.l3_hdr_offset);
-      if (oflags & VNET_BUFFER_OFFLOAD_F_IP_CKSUM)
-	ip4->checksum = ip4_header_checksum (ip4);
     }
   else if (b->flags & VNET_BUFFER_F_IS_IP6)
     {
+      ip6_header_t *ip6;
       generic_header_offset_t gho = { 0 };
       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;
+      ip6 = (ip6_header_t *) (vlib_buffer_get_current (b) + gho.l3_hdr_offset);
+
+      /*
+       * virtio devices assume the l4 header is set to the checksum of the
+       * l3 pseudo-header, so we compute it before tx-ing
+       */
       if (oflags & VNET_BUFFER_OFFLOAD_F_TCP_CKSUM)
 	{
+	  tcp_header_t *tcp =
+	    (tcp_header_t *) (vlib_buffer_get_current (b) + gho.l4_hdr_offset);
+	  tcp->checksum = ip6_pseudo_header_cksum (ip6);
 	  hdr->csum_offset = STRUCT_OFFSET_OF (tcp_header_t, checksum);
 	}
       else if (oflags & VNET_BUFFER_OFFLOAD_F_UDP_CKSUM)
 	{
+	  udp_header_t *udp =
+	    (udp_header_t *) (vlib_buffer_get_current (b) + gho.l4_hdr_offset);
+	  udp->checksum = ip6_pseudo_header_cksum (ip6);
 	  hdr->csum_offset = STRUCT_OFFSET_OF (udp_header_t, checksum);
 	}
     }
diff --git a/src/vnet/devices/virtio/vhost_user_output.c b/src/vnet/devices/virtio/vhost_user_output.c
index 4efafa85333..15e39a11692 100644
--- a/src/vnet/devices/virtio/vhost_user_output.c
+++ b/src/vnet/devices/virtio/vhost_user_output.c
@@ -37,6 +37,7 @@
 #include <vnet/ethernet/ethernet.h>
 #include <vnet/devices/devices.h>
 #include <vnet/feature/feature.h>
+#include <vnet/ip/ip_psh_cksum.h>
 
 #include <vnet/devices/virtio/vhost_user.h>
 #include <vnet/devices/virtio/vhost_user_inline.h>
@@ -208,27 +209,40 @@ vhost_user_handle_tx_offload (vhost_user_intf_t * vui, vlib_buffer_t * b,
   int is_ip4 = b->flags & VNET_BUFFER_F_IS_IP4;
   int is_ip6 = b->flags & VNET_BUFFER_F_IS_IP6;
   vnet_buffer_oflags_t oflags = vnet_buffer (b)->oflags;
+  u16 psh_cksum = 0;
+  ip4_header_t *ip4 = 0;
+  ip6_header_t *ip6 = 0;
 
   ASSERT (!(is_ip4 && is_ip6));
   vnet_generic_header_offset_parser (b, &gho, 1 /* l2 */ , is_ip4, is_ip6);
   if (oflags & VNET_BUFFER_OFFLOAD_F_IP_CKSUM)
     {
-      ip4_header_t *ip4;
-
       ip4 =
 	(ip4_header_t *) (vlib_buffer_get_current (b) + gho.l3_hdr_offset);
       ip4->checksum = ip4_header_checksum (ip4);
+      psh_cksum = ip4_pseudo_header_cksum (ip4);
+    }
+  else
+    {
+      ip6 = (ip6_header_t *) (vlib_buffer_get_current (b) + gho.l3_hdr_offset);
+      psh_cksum = ip6_pseudo_header_cksum (ip6);
     }
 
   /* checksum offload */
   if (oflags & VNET_BUFFER_OFFLOAD_F_UDP_CKSUM)
     {
+      udp_header_t *udp =
+	(udp_header_t *) (vlib_buffer_get_current (b) + gho.l4_hdr_offset);
+      udp->checksum = psh_cksum;
       hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
       hdr->csum_start = gho.l4_hdr_offset;
       hdr->csum_offset = offsetof (udp_header_t, checksum);
     }
   else if (oflags & VNET_BUFFER_OFFLOAD_F_TCP_CKSUM)
     {
+      tcp_header_t *tcp =
+	(tcp_header_t *) (vlib_buffer_get_current (b) + gho.l4_hdr_offset);
+      tcp->checksum = psh_cksum;
       hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
       hdr->csum_start = gho.l4_hdr_offset;
       hdr->csum_offset = offsetof (tcp_header_t, checksum);
diff --git a/src/vnet/ip/ip_psh_cksum.h b/src/vnet/ip/ip_psh_cksum.h
new file mode 100644
index 00000000000..eaac401f223
--- /dev/null
+++ b/src/vnet/ip/ip_psh_cksum.h
@@ -0,0 +1,54 @@
+/*
+ * SPDX-License-Identifier: Apache-2.0
+ * Copyright(c) 2021 Cisco Systems, Inc.
+ */
+
+#ifndef included_ip_psh_cksum_h
+#define included_ip_psh_cksum_h
+
+#include <vnet/ip/ip.h>
+
+typedef struct _ip4_psh
+{
+  ip4_address_t src;
+  ip4_address_t dst;
+  u8 zero;
+  u8 proto;
+  u16 l4len;
+} ip4_psh_t;
+
+typedef struct _ip6_psh
+{
+  ip6_address_t src;
+  ip6_address_t dst;
+  u32 l4len;
+  u32 proto;
+} ip6_psh_t;
+
+STATIC_ASSERT (sizeof (ip4_psh_t) == 12, "ipv4 pseudo header is 12B");
+STATIC_ASSERT (sizeof (ip6_psh_t) == 40, "ipv6 pseudo header is 40B");
+
+static_always_inline u16
+ip4_pseudo_header_cksum (ip4_header_t *ip4)
+{
+  ip4_psh_t psh = { 0 };
+  psh.src = ip4->src_address;
+  psh.dst = ip4->dst_address;
+  psh.proto = ip4->protocol;
+  psh.l4len = clib_host_to_net_u16 (clib_net_to_host_u16 (ip4->length) -
+				    sizeof (ip4_header_t));
+  return ~clib_net_to_host_u16 (ip_csum (&psh, sizeof (ip4_psh_t)));
+}
+
+static_always_inline u16
+ip6_pseudo_header_cksum (ip6_header_t *ip6)
+{
+  ip6_psh_t psh = { 0 };
+  psh.src = ip6->src_address;
+  psh.dst = ip6->dst_address;
+  psh.l4len = ip6->payload_length;
+  psh.proto = clib_host_to_net_u32 ((u32) ip6->protocol);
+  return ~clib_net_to_host_u16 (ip_csum (&psh, sizeof (ip6_psh_t)));
+}
+
+#endif /* included_ip_psh_cksum_h */