MTU: IP fragmentation added to ip4-rewrite and ip6-rewrite

Change-Id: Ibae260273f25a319153be37470aed49ff73e957a
Signed-off-by: Ole Troan <ot@cisco.com>

4 files changed, 59 insertions, 19 deletions

diff --git a/src/vnet/ip/ip4_forward.c b/src/vnet/ip/ip4_forward.c
index 6d5f07bda2c..3444eb88b7c 100644
--- a/src/vnet/ip/ip4_forward.c
+++ b/src/vnet/ip/ip4_forward.c
@@ -39,6 +39,7 @@
 
 #include <vnet/vnet.h>
 #include <vnet/ip/ip.h>
+#include <vnet/ip/ip_frag.h>
 #include <vnet/ethernet/ethernet.h>	/* for ethernet_header_t */
 #include <vnet/ethernet/arp_packet.h>	/* for ethernet_arp_header_t */
 #include <vnet/ppp/ppp.h>
@@ -2013,6 +2014,8 @@ typedef enum
 {
   IP4_REWRITE_NEXT_DROP,
   IP4_REWRITE_NEXT_ICMP_ERROR,
+  IP4_REWRITE_NEXT_FRAGMENT,
+  IP4_REWRITE_N_NEXT		/* Last */
 } ip4_rewrite_next_t;
 
 /**
@@ -2042,8 +2045,10 @@ ip4_mtu_check (vlib_buffer_t * b, u16 packet_len,
 	}
       else
 	{
-	  /* Add support for fragmentation here */
-	  *next = IP4_REWRITE_NEXT_DROP;
+	  /* IP fragmentation */
+	  ip_frag_set_vnet_buffer (b, 0, adj_packet_bytes,
+				   IP4_FRAG_NEXT_IP4_LOOKUP, 0);
+	  *next = IP4_REWRITE_NEXT_FRAGMENT;
 	}
     }
 }
@@ -2539,10 +2544,11 @@ VLIB_REGISTER_NODE (ip4_rewrite_node) = {
 
   .format_trace = format_ip4_rewrite_trace,
 
-  .n_next_nodes = 2,
+  .n_next_nodes = IP4_REWRITE_N_NEXT,
   .next_nodes = {
     [IP4_REWRITE_NEXT_DROP] = "ip4-drop",
     [IP4_REWRITE_NEXT_ICMP_ERROR] = "ip4-icmp-error",
+    [IP4_REWRITE_NEXT_FRAGMENT] = "ip4-frag",
   },
 };
 VLIB_NODE_FUNCTION_MULTIARCH (ip4_rewrite_node, ip4_rewrite)
diff --git a/src/vnet/ip/ip6_forward.c b/src/vnet/ip/ip6_forward.c
index a66070302c0..f7946b0b3af 100644
--- a/src/vnet/ip/ip6_forward.c
+++ b/src/vnet/ip/ip6_forward.c
@@ -39,6 +39,7 @@
 
 #include <vnet/vnet.h>
 #include <vnet/ip/ip.h>
+#include <vnet/ip/ip_frag.h>
 #include <vnet/ip/ip6_neighbor.h>
 #include <vnet/ethernet/ethernet.h>	/* for ethernet_header_t */
 #include <vnet/srp/srp.h>	/* for srp_hw_interface_class */
@@ -1542,6 +1543,8 @@ typedef enum
 {
   IP6_REWRITE_NEXT_DROP,
   IP6_REWRITE_NEXT_ICMP_ERROR,
+  IP6_REWRITE_NEXT_FRAGMENT,
+  IP6_REWRITE_N_NEXT		/* Last */
 } ip6_rewrite_next_t;
 
 /**
@@ -1552,14 +1555,25 @@ typedef enum
 
 always_inline void
 ip6_mtu_check (vlib_buffer_t * b, u16 packet_bytes,
-	       u16 adj_packet_bytes, u32 * next, u32 * error)
+	       u16 adj_packet_bytes, bool is_locally_generated,
+	       u32 * next, u32 * error)
 {
   if (adj_packet_bytes >= 1280 && packet_bytes > adj_packet_bytes)
     {
-      *error = IP6_ERROR_MTU_EXCEEDED;
-      icmp6_error_set_vnet_buffer (b, ICMP6_packet_too_big, 0,
-				   adj_packet_bytes);
-      *next = IP6_REWRITE_NEXT_ICMP_ERROR;
+      if (is_locally_generated)
+	{
+	  /* IP fragmentation */
+	  ip_frag_set_vnet_buffer (b, 0, adj_packet_bytes,
+				   IP6_FRAG_NEXT_IP6_LOOKUP, 0);
+	  *next = IP6_REWRITE_NEXT_FRAGMENT;
+	}
+      else
+	{
+	  *error = IP6_ERROR_MTU_EXCEEDED;
+	  icmp6_error_set_vnet_buffer (b, ICMP6_packet_too_big, 0,
+				       adj_packet_bytes);
+	  *next = IP6_REWRITE_NEXT_ICMP_ERROR;
+	}
     }
 }
 
@@ -1591,6 +1605,7 @@ ip6_rewrite_inline (vlib_main_t * vm,
 	  u32 pi0, rw_len0, next0, error0, adj_index0;
 	  u32 pi1, rw_len1, next1, error1, adj_index1;
 	  u32 tx_sw_if_index0, tx_sw_if_index1;
+	  bool is_locally_originated0, is_locally_originated1;
 
 	  /* Prefetch next iteration. */
 	  {
@@ -1629,7 +1644,9 @@ ip6_rewrite_inline (vlib_main_t * vm,
 	  error0 = error1 = IP6_ERROR_NONE;
 	  next0 = next1 = IP6_REWRITE_NEXT_DROP;
 
-	  if (PREDICT_TRUE (!(p0->flags & VNET_BUFFER_F_LOCALLY_ORIGINATED)))
+	  is_locally_originated0 =
+	    p0->flags & VNET_BUFFER_F_LOCALLY_ORIGINATED;
+	  if (PREDICT_TRUE (!is_locally_originated0))
 	    {
 	      i32 hop_limit0 = ip0->hop_limit;
 
@@ -1658,7 +1675,9 @@ ip6_rewrite_inline (vlib_main_t * vm,
 	    {
 	      p0->flags &= ~VNET_BUFFER_F_LOCALLY_ORIGINATED;
 	    }
-	  if (PREDICT_TRUE (!(p1->flags & VNET_BUFFER_F_LOCALLY_ORIGINATED)))
+	  is_locally_originated1 =
+	    p1->flags & VNET_BUFFER_F_LOCALLY_ORIGINATED;
+	  if (PREDICT_TRUE (!is_locally_originated1))
 	    {
 	      i32 hop_limit1 = ip1->hop_limit;
 
@@ -1711,11 +1730,11 @@ ip6_rewrite_inline (vlib_main_t * vm,
 	  ip6_mtu_check (p0, clib_net_to_host_u16 (ip0->payload_length) +
 			 sizeof (ip6_header_t),
 			 adj0[0].rewrite_header.max_l3_packet_bytes,
-			 &next0, &error0);
+			 is_locally_originated0, &next0, &error0);
 	  ip6_mtu_check (p1, clib_net_to_host_u16 (ip1->payload_length) +
 			 sizeof (ip6_header_t),
 			 adj1[0].rewrite_header.max_l3_packet_bytes,
-			 &next1, &error1);
+			 is_locally_originated1, &next1, &error1);
 
 	  /* Don't adjust the buffer for hop count issue; icmp-error node
 	   * wants to see the IP headerr */
@@ -1789,6 +1808,7 @@ ip6_rewrite_inline (vlib_main_t * vm,
 	  u32 pi0, rw_len0;
 	  u32 adj_index0, next0, error0;
 	  u32 tx_sw_if_index0;
+	  bool is_locally_originated0;
 
 	  pi0 = to_next[0] = from[0];
 
@@ -1804,7 +1824,9 @@ ip6_rewrite_inline (vlib_main_t * vm,
 	  next0 = IP6_REWRITE_NEXT_DROP;
 
 	  /* Check hop limit */
-	  if (PREDICT_TRUE (!(p0->flags & VNET_BUFFER_F_LOCALLY_ORIGINATED)))
+	  is_locally_originated0 =
+	    p0->flags & VNET_BUFFER_F_LOCALLY_ORIGINATED;
+	  if (PREDICT_TRUE (!is_locally_originated0))
 	    {
 	      i32 hop_limit0 = ip0->hop_limit;
 
@@ -1852,7 +1874,7 @@ ip6_rewrite_inline (vlib_main_t * vm,
 	  ip6_mtu_check (p0, clib_net_to_host_u16 (ip0->payload_length) +
 			 sizeof (ip6_header_t),
 			 adj0[0].rewrite_header.max_l3_packet_bytes,
-			 &next0, &error0);
+			 is_locally_originated0, &next0, &error0);
 
 	  /* Don't adjust the buffer for hop count issue; icmp-error node
 	   * wants to see the IP header */
@@ -1968,11 +1990,12 @@ VLIB_REGISTER_NODE (ip6_rewrite_node) =
   .name = "ip6-rewrite",
   .vector_size = sizeof (u32),
   .format_trace = format_ip6_rewrite_trace,
-  .n_next_nodes = 2,
+  .n_next_nodes = IP6_REWRITE_N_NEXT,
   .next_nodes =
   {
     [IP6_REWRITE_NEXT_DROP] = "ip6-drop",
     [IP6_REWRITE_NEXT_ICMP_ERROR] = "ip6-icmp-error",
+    [IP6_REWRITE_NEXT_FRAGMENT] = "ip6-frag",
   },
 };
 /* *INDENT-ON* */
diff --git a/src/vnet/ip/ip_frag.c b/src/vnet/ip/ip_frag.c
index fba25fffff7..1207ec54b49 100644
--- a/src/vnet/ip/ip_frag.c
+++ b/src/vnet/ip/ip_frag.c
@@ -85,6 +85,12 @@ ip4_frag_do_fragment (vlib_main_t * vm, u32 pi, u32 ** buffer,
       return;
     }
 
+  if (p->flags & VLIB_BUFFER_NEXT_PRESENT)
+    {
+      *error = IP_FRAG_ERROR_MALFORMED;
+      return;
+    }
+
   if (ip4_is_fragment (ip4))
     {
       ip_frag_id = ip4->fragment_id;
@@ -342,6 +348,12 @@ ip6_frag_do_fragment (vlib_main_t * vm, u32 pi, u32 ** buffer,
       return;
     }
 
+  if (p->flags & VLIB_BUFFER_NEXT_PRESENT)
+    {
+      *error = IP_FRAG_ERROR_MALFORMED;
+      return;
+    }
+
   u8 has_more;
   u16 initial_offset;
   if (*next_header == IP_PROTOCOL_IPV6_FRAGMENTATION)
diff --git a/test/test_mtu.py b/test/test_mtu.py
index 1327c4bfd8e..d6be2e2afb5 100644
--- a/test/test_mtu.py
+++ b/test/test_mtu.py
@@ -119,7 +119,6 @@ class TestMTU(VppTestCase):
             # n.show2()
             self.validate_bytes(str(p[1]), icmp4_reply)
 
-        '''
         # Now with DF off. Expect fragments.
         # First go with 1500 byte packets.
         p_payload = UDP(sport=1234, dport=1234) / self.payload(
@@ -127,19 +126,18 @@ class TestMTU(VppTestCase):
         p4 = p_ether / p_ip4 / p_payload
         p4.flags = 0
         p4_reply = p_ip4 / p_payload
-        p4_reply.ttl = 62 # check this
+        p4_reply.ttl = 62  # check this
         p4_reply.flags = 0
         p4_reply.id = 256
         self.pg_enable_capture()
         self.pg0.add_stream(p4*1)
         self.pg_start()
         rx = self.pg1.get_capture(3)
-        print('RX', len(rx))
         reass_pkt = reassemble(rx)
         self.validate(reass_pkt, p4_reply)
+
         '''
         # Now what happens with a 9K frame
-        '''
         p_payload = UDP(sport=1234, dport=1234) / self.payload(
             current_mtu - 20 - 8)
         p4 = p_ether / p_ip4 / p_payload
@@ -158,6 +156,7 @@ class TestMTU(VppTestCase):
         p4_reply.show2()
         self.validate(reass_pkt, p4_reply)
         '''
+
         # Reset MTU
         self.vapi.sw_interface_set_mtu(self.pg1.sw_if_index,
                                        [current_mtu, 0, 0, 0])