Add support for transparent ethernet bridging to GRE

Change-Id: Iaa3cefe223eb48b128893029a17e092b72a5157c
Signed-off-by: David Hotham <david.hotham@metaswitch.com>

6 files changed, 90 insertions, 27 deletions

diff --git a/vnet/vnet/ethernet/node.c b/vnet/vnet/ethernet/node.c
index 74d4f476d7a..a8a4fcd074d 100644
--- a/vnet/vnet/ethernet/node.c
+++ b/vnet/vnet/ethernet/node.c
@@ -238,7 +238,8 @@ determine_next_node (ethernet_main_t * em,
     {
       *next0 = em->l2_next;
       // record the L2 len and reset the buffer so the L2 header is preserved
-      vnet_buffer (b0)->l2.l2_len = b0->current_data;
+      u32 eth_start = vnet_buffer (b0)->ethernet.start_of_ethernet_header;
+      vnet_buffer (b0)->l2.l2_len = b0->current_data - eth_start;
       vlib_buffer_advance (b0, -ethernet_buffer_header_size (b0));
 
       // check for common IP/MPLS ethertypes
diff --git a/vnet/vnet/gre/gre.c b/vnet/vnet/gre/gre.c
index 6d375159818..f00977c8cd6 100644
--- a/vnet/vnet/gre/gre.c
+++ b/vnet/vnet/gre/gre.c
@@ -283,17 +283,25 @@ gre_interface_tx (vlib_main_t * vm,
           vnet_buffer (b0)->sw_if_index[VLIB_TX] = t->outer_fib_index;
           vnet_buffer (b1)->sw_if_index[VLIB_TX] = t->outer_fib_index;
 
-          ip0 = vlib_buffer_get_current (b0);
-          gre_protocol0 = clib_net_to_host_u16 (0x800);
-          gre_protocol0 = 
-              ((ip0->ip_version_and_header_length & 0xF0) == 0x60) ? 
-              0x86DD : gre_protocol0;
-
-          ip1 = vlib_buffer_get_current (b1);
-          gre_protocol1 = clib_net_to_host_u16 (0x800);
-          gre_protocol1 = 
-              ((ip1->ip_version_and_header_length & 0xF0) == 0x60) ? 
-              0x86DD : gre_protocol1;
+          if (PREDICT_FALSE(t->teb))
+          {
+            gre_protocol0 = clib_net_to_host_u16(GRE_PROTOCOL_teb);
+            gre_protocol1 = clib_net_to_host_u16(GRE_PROTOCOL_teb);
+          }
+          else
+          {
+            ip0 = vlib_buffer_get_current (b0);
+            gre_protocol0 = clib_net_to_host_u16 (0x800);
+            gre_protocol0 =
+                ((ip0->ip_version_and_header_length & 0xF0) == 0x60) ?
+                0x86DD : gre_protocol0;
+
+            ip1 = vlib_buffer_get_current (b1);
+            gre_protocol1 = clib_net_to_host_u16 (0x800);
+            gre_protocol1 =
+                ((ip1->ip_version_and_header_length & 0xF0) == 0x60) ?
+                0x86DD : gre_protocol1;
+          }
 
           vlib_buffer_advance (b0, -sizeof(*h0));
           vlib_buffer_advance (b1, -sizeof(*h1));
@@ -366,10 +374,17 @@ gre_interface_tx (vlib_main_t * vm,
 
           vnet_buffer (b0)->sw_if_index[VLIB_TX] = t->outer_fib_index;
           ip0 = vlib_buffer_get_current (b0);
-          gre_protocol0 = clib_net_to_host_u16 (0x800);
-          gre_protocol0 = 
-              ((ip0->ip_version_and_header_length & 0xF0) == 0x60) ? 
-              0x86DD : gre_protocol0;
+          if (PREDICT_FALSE(t->teb))
+          {
+            gre_protocol0 = clib_net_to_host_u16(GRE_PROTOCOL_teb);
+          }
+          else
+          {
+            gre_protocol0 = clib_net_to_host_u16 (0x800);
+            gre_protocol0 =
+                ((ip0->ip_version_and_header_length & 0xF0) == 0x60) ?
+                0x86DD : gre_protocol0;
+          }
 
           vlib_buffer_advance (b0, -sizeof(*h0));
 
diff --git a/vnet/vnet/gre/gre.h b/vnet/vnet/gre/gre.h
index d875be990c6..ad599d2f09e 100644
--- a/vnet/vnet/gre/gre.h
+++ b/vnet/vnet/gre/gre.h
@@ -55,6 +55,7 @@ typedef struct {
   u32 outer_fib_index;
   u32 hw_if_index;
   u32 sw_if_index;
+  u8 teb;
 } gre_tunnel_t;
 
 typedef struct {
@@ -127,6 +128,7 @@ typedef struct {
 
   ip4_address_t src, dst;
   u32 outer_fib_id;
+  u8 teb;
 } vnet_gre_add_del_tunnel_args_t;
 
 int vnet_gre_add_del_tunnel
diff --git a/vnet/vnet/gre/interface.c b/vnet/vnet/gre/interface.c
index 03591e1ff41..864c384b992 100644
--- a/vnet/vnet/gre/interface.c
+++ b/vnet/vnet/gre/interface.c
@@ -26,10 +26,11 @@ u8 * format_gre_tunnel (u8 * s, va_list * args)
   gre_main_t * gm = &gre_main;
 
   s = format (s,
-              "[%d] %U (src) %U (dst) outer_fib_index %d",
+              "[%d] %U (src) %U (dst) payload %s outer_fib_index %d",
               t - gm->tunnels,
               format_ip4_address, &t->tunnel_src,
               format_ip4_address, &t->tunnel_dst,
+              (t->teb ? "teb" : "ip"),
               t->outer_fib_index);
   return s;
 }
@@ -47,6 +48,8 @@ int vnet_gre_add_del_tunnel
   u32 outer_fib_index;
   uword * p;
   u64 key;
+  u8 address[6];
+  clib_error_t *error;
 
   key = (u64)a->src.as_u32 << 32 | (u64)a->dst.as_u32;
   p = hash_get (gm->tunnel_by_key, key);
@@ -87,12 +90,34 @@ int vnet_gre_add_del_tunnel
           (&im->sw_if_counters[VNET_INTERFACE_COUNTER_DROP], sw_if_index);
         vnet_interface_counter_unlock(im);
     } else {
+      if (a->teb)
+      {
+        /* Default MAC address (d00b:eed0:0000 + sw_if_index) */
+        memset (address, 0, sizeof (address));
+        address[0] = 0xd0;
+        address[1] = 0x0b;
+        address[2] = 0xee;
+        address[3] = 0xd0;
+        address[4] = t - gm->tunnels;
+
+        error = ethernet_register_interface
+          (vnm,
+           gre_device_class.index, t - gm->tunnels, address, &hw_if_index,
+           0);
+
+        if (error)
+        {
+          clib_error_report (error);
+          return VNET_API_ERROR_INVALID_REGISTRATION;
+        }
+      } else {
         hw_if_index = vnet_register_interface
           (vnm, gre_device_class.index, t - gm->tunnels,
            gre_hw_interface_class.index,
            t - gm->tunnels);
-        hi = vnet_get_hw_interface (vnm, hw_if_index);
-        sw_if_index = hi->sw_if_index;
+      }
+      hi = vnet_get_hw_interface (vnm, hw_if_index);
+      sw_if_index = hi->sw_if_index;
     }
 
     t->hw_if_index = hw_if_index;
@@ -112,6 +137,7 @@ int vnet_gre_add_del_tunnel
     /* Standard default gre MTU. */
     hi->max_l3_packet_bytes[VLIB_RX] = hi->max_l3_packet_bytes[VLIB_TX] = 9000;
 
+    t->teb = a->teb;
     clib_memcpy (&t->tunnel_src, &a->src, sizeof (t->tunnel_src));
     clib_memcpy (&t->tunnel_dst, &a->dst, sizeof (t->tunnel_dst));
 
@@ -156,6 +182,7 @@ create_gre_tunnel_command_fn (vlib_main_t * vm,
   vnet_gre_add_del_tunnel_args_t _a, * a = &_a;
   ip4_address_t src, dst;
   u32 outer_fib_id = 0;
+  u8 teb = 0;
   int rv;
   u32 num_m_args = 0;
   u8 is_add = 1;
@@ -174,6 +201,8 @@ create_gre_tunnel_command_fn (vlib_main_t * vm,
       num_m_args++;
     else if (unformat (line_input, "outer-fib-id %d", &outer_fib_id))
       ;
+    else if (unformat (line_input, "teb"))
+      teb = 1;
     else
       return clib_error_return (0, "unknown input `%U'",
                                 format_unformat_error, input);
@@ -189,6 +218,7 @@ create_gre_tunnel_command_fn (vlib_main_t * vm,
   memset (a, 0, sizeof (*a));
   a->is_add = is_add;
   a->outer_fib_id = outer_fib_id;
+  a->teb = teb;
   clib_memcpy(&a->src, &src, sizeof(src));
   clib_memcpy(&a->dst, &dst, sizeof(dst));
 
@@ -214,7 +244,7 @@ create_gre_tunnel_command_fn (vlib_main_t * vm,
 VLIB_CLI_COMMAND (create_gre_tunnel_command, static) = {
   .path = "create gre tunnel",
   .short_help = "create gre tunnel src <addr> dst <addr> "
-                "[outer-fib-id <fib>] [del]",
+                "[outer-fib-id <fib>] [teb] [del]",
   .function = create_gre_tunnel_command_fn,
 };
 
diff --git a/vnet/vnet/gre/node.c b/vnet/vnet/gre/node.c
index 97f9dac2e2b..d5ea4b65ddb 100644
--- a/vnet/vnet/gre/node.c
+++ b/vnet/vnet/gre/node.c
@@ -23,6 +23,7 @@
 #define foreach_gre_input_next			\
 _(PUNT, "error-punt")                           \
 _(DROP, "error-drop")                           \
+_(ETHERNET_INPUT, "ethernet-input")             \
 _(IP4_INPUT, "ip4-input")                       \
 _(IP6_INPUT, "ip6-input")			
 
@@ -160,9 +161,11 @@ gre_input (vlib_main_t * vm,
               : b1->error;
           next1 = verr1 ? GRE_INPUT_NEXT_DROP : next1;
 
+
           /* RPF check for ip4/ip6 input */
-          if (PREDICT_FALSE(next0 == GRE_INPUT_NEXT_IP4_INPUT 
-                            || next0 == GRE_INPUT_NEXT_IP6_INPUT))
+          if (PREDICT_FALSE(next0 == GRE_INPUT_NEXT_IP4_INPUT
+                            || next0 == GRE_INPUT_NEXT_IP6_INPUT
+                            || next0 == GRE_INPUT_NEXT_ETHERNET_INPUT))
             {
               u64 key = ((u64)(vnet_buffer(b0)->gre.dst) << 32) |
                          (u64)(vnet_buffer(b0)->gre.src);
@@ -207,11 +210,13 @@ gre_input (vlib_main_t * vm,
                                                len /* bytes */);
 
               vnet_buffer(b0)->sw_if_index[VLIB_TX] = tunnel_fib_index;
+              vnet_buffer(b0)->sw_if_index[VLIB_RX] = tunnel_sw_if_index;
             }
 
 drop0:
-          if (PREDICT_FALSE(next1 == GRE_INPUT_NEXT_IP4_INPUT 
-                            || next1 == GRE_INPUT_NEXT_IP6_INPUT))
+          if (PREDICT_FALSE(next1 == GRE_INPUT_NEXT_IP4_INPUT
+                            || next1 == GRE_INPUT_NEXT_IP6_INPUT
+                            || next1 == GRE_INPUT_NEXT_ETHERNET_INPUT))
             {
               u64 key = ((u64)(vnet_buffer(b1)->gre.dst) << 32) |
                          (u64)(vnet_buffer(b1)->gre.src);
@@ -256,6 +261,7 @@ drop0:
                                                len /* bytes */);
 
               vnet_buffer(b1)->sw_if_index[VLIB_TX] = tunnel_fib_index;
+              vnet_buffer(b1)->sw_if_index[VLIB_RX] = tunnel_sw_if_index;
             }
 drop1:
           if (PREDICT_FALSE(b0->flags & VLIB_BUFFER_IS_TRACED)) 
@@ -326,11 +332,13 @@ drop1:
               : b0->error;
           next0 = verr0 ? GRE_INPUT_NEXT_DROP : next0;
 
+
           /* For IP payload we need to find source interface
              so we can increase counters and help forward node to
              pick right FIB */
-          if (PREDICT_FALSE(next0 == GRE_INPUT_NEXT_IP4_INPUT 
-                            || next0 == GRE_INPUT_NEXT_IP6_INPUT))
+          if (PREDICT_FALSE(next0 == GRE_INPUT_NEXT_IP4_INPUT
+                            || next0 == GRE_INPUT_NEXT_IP6_INPUT
+                            || next0 == GRE_INPUT_NEXT_ETHERNET_INPUT))
             {
               u64 key = ((u64)(vnet_buffer(b0)->gre.dst) << 32) |
                          (u64)(vnet_buffer(b0)->gre.src);
@@ -375,6 +383,7 @@ drop1:
                                                len /* bytes */);
 
               vnet_buffer(b0)->sw_if_index[VLIB_TX] = tunnel_fib_index;
+              vnet_buffer(b0)->sw_if_index[VLIB_RX] = tunnel_sw_if_index;
             }
 
 drop:
@@ -476,7 +485,7 @@ gre_setup_node (vlib_main_t * vm, u32 node_index)
 static clib_error_t * gre_input_init (vlib_main_t * vm)
 {
   gre_input_runtime_t * rt;
-  vlib_node_t *ip4_input, *ip6_input, *mpls_unicast_input;
+  vlib_node_t *ethernet_input, *ip4_input, *ip6_input, *mpls_unicast_input;
 
   {
     clib_error_t * error; 
@@ -494,6 +503,8 @@ static clib_error_t * gre_input_init (vlib_main_t * vm)
      /* bits in index */ BITS (((gre_header_t *) 0)->protocol));
 
   /* These could be moved to the supported protocol input node defn's */
+  ethernet_input = vlib_get_node_by_name (vm, (u8 *)"ethernet-input");
+  ASSERT(ethernet_input);
   ip4_input = vlib_get_node_by_name (vm, (u8 *)"ip4-input");
   ASSERT(ip4_input);
   ip6_input = vlib_get_node_by_name (vm, (u8 *)"ip6-input");
@@ -501,6 +512,9 @@ static clib_error_t * gre_input_init (vlib_main_t * vm)
   mpls_unicast_input = vlib_get_node_by_name (vm, (u8 *)"mpls-gre-input");
   ASSERT(mpls_unicast_input);
 
+  gre_register_input_protocol (vm, GRE_PROTOCOL_teb,
+                               ethernet_input->index);
+
   gre_register_input_protocol (vm, GRE_PROTOCOL_ip4, 
                                ip4_input->index);
 
diff --git a/vnet/vnet/gre/packet.h b/vnet/vnet/gre/packet.h
index 573f2624bf8..cc2ccda9eff 100644
--- a/vnet/vnet/gre/packet.h
+++ b/vnet/vnet/gre/packet.h
@@ -21,6 +21,7 @@
 #define foreach_gre_protocol			\
 _ (0x0800, ip4)                                 \
 _ (0x86DD, ip6)                                 \
+_ (0x6558, teb)                                 \
 _ (0x0806, arp)					\
 _ (0x8847, mpls_unicast)			\
 _ (0x894F, nsh)