reassembly: prevent long chain attack

limit max # of fragments to 3 per packet by default
add API option to configure the limit at runtime

Change-Id: Ie4b9507bf5c6095b9a5925972b37fe0032f4f9e8
Signed-off-by: Klement Sekera <ksekera@cisco.com>

8 files changed, 86 insertions, 13 deletions

diff --git a/src/vnet/ip/ip.api b/src/vnet/ip/ip.api
index 39d394f709d..afb0960c78a 100644
--- a/src/vnet/ip/ip.api
+++ b/src/vnet/ip/ip.api
@@ -20,7 +20,7 @@
     called through a shared memory interface. 
 */
 
-option version = "2.0.0";
+option version = "2.0.1";
 import "vnet/ip/ip_types.api";
 import "vnet/fib/fib_types.api";
 import "vnet/ethernet/ethernet_types.api";
@@ -1085,6 +1085,7 @@ autoreply define ip_reassembly_set
   u32 context;
   u32 timeout_ms;
   u32 max_reassemblies;
+  u32 max_reassembly_length;
   u32 expire_walk_interval_ms;
   u8 is_ip6;
 };
@@ -1102,6 +1103,7 @@ define ip_reassembly_get_reply
   i32 retval;
   u32 timeout_ms;
   u32 max_reassemblies;
+  u32 max_reassembly_length;
   u32 expire_walk_interval_ms;
   u8 is_ip6;
 };
diff --git a/src/vnet/ip/ip4_error.h b/src/vnet/ip/ip4_error.h
index badcc6609e9..d3bf6d79714 100644
--- a/src/vnet/ip/ip4_error.h
+++ b/src/vnet/ip/ip4_error.h
@@ -86,6 +86,8 @@
   /* Errors signalled by ip4-reassembly */                              \
   _ (REASS_DUPLICATE_FRAGMENT, "duplicate/overlapping fragments")       \
   _ (REASS_LIMIT_REACHED, "drops due to concurrent reassemblies limit") \
+  _ (REASS_FRAGMENT_CHAIN_TOO_LONG, "fragment chain too long (drop)")   \
+  _ (REASS_NO_BUF, "out of buffers (drop)")                             \
   _ (REASS_MALFORMED_PACKET, "malformed packets")                       \
   _ (REASS_INTERNAL_ERROR, "drops due to internal reassembly error")
 
diff --git a/src/vnet/ip/ip4_reassembly.c b/src/vnet/ip/ip4_reassembly.c
index b38ade09efd..f27351038fe 100644
--- a/src/vnet/ip/ip4_reassembly.c
+++ b/src/vnet/ip/ip4_reassembly.c
@@ -30,6 +30,7 @@
 #define IP4_REASS_TIMEOUT_DEFAULT_MS 100
 #define IP4_REASS_EXPIRE_WALK_INTERVAL_DEFAULT_MS 10000	// 10 seconds default
 #define IP4_REASS_MAX_REASSEMBLIES_DEFAULT 1024
+#define IP4_REASS_MAX_REASSEMBLY_LENGTH_DEFAULT 3
 #define IP4_REASS_HT_LOAD_FACTOR (0.75)
 
 #define IP4_REASS_DEBUG_BUFFERS 0
@@ -57,6 +58,7 @@
 typedef enum
 {
   IP4_REASS_RC_OK,
+  IP4_REASS_RC_TOO_MANY_FRAGMENTS,
   IP4_REASS_RC_INTERNAL_ERROR,
   IP4_REASS_RC_NO_BUF,
 } ip4_reass_rc_t;
@@ -133,7 +135,8 @@ typedef struct
   u8 next_index;
   // minimum fragment length for this reassembly - used to estimate MTU
   u16 min_fragment_length;
-
+  // number of fragments in this reassembly
+  u32 fragments_n;
 } ip4_reass_t;
 
 typedef struct
@@ -150,6 +153,9 @@ typedef struct
   u32 timeout_ms;
   f64 timeout;
   u32 expire_walk_interval_ms;
+  // maximum number of fragments in one reassembly
+  u32 max_reass_len;
+  // maximum number of reassemblies
   u32 max_reass_n;
 
   // IPv4 runtime
@@ -750,6 +756,7 @@ ip4_reass_update (vlib_main_t * vm, vlib_node_runtime_t * node,
 	}
       *bi0 = ~0;
       reass->min_fragment_length = clib_net_to_host_u16 (fip->length);
+      reass->fragments_n = 1;
       return IP4_REASS_RC_OK;
     }
   reass->min_fragment_length = clib_min (clib_net_to_host_u16 (fip->length),
@@ -907,6 +914,7 @@ ip4_reass_update (vlib_main_t * vm, vlib_node_runtime_t * node,
 	}
       break;
     }
+  ++reass->fragments_n;
   if (consumed)
     {
       if (PREDICT_FALSE (fb->flags & VLIB_BUFFER_IS_TRACED))
@@ -925,6 +933,10 @@ ip4_reass_update (vlib_main_t * vm, vlib_node_runtime_t * node,
       if (consumed)
 	{
 	  *bi0 = ~0;
+	  if (reass->fragments_n > rm->max_reass_len)
+	    {
+	      rc = IP4_REASS_RC_TOO_MANY_FRAGMENTS;
+	    }
 	}
       else
 	{
@@ -1022,10 +1034,26 @@ ip4_reassembly_inline (vlib_main_t * vm,
 			case IP4_REASS_RC_OK:
 			  /* nothing to do here */
 			  break;
+			case IP4_REASS_RC_TOO_MANY_FRAGMENTS:
+			  vlib_node_increment_counter (vm, node->node_index,
+						       IP4_ERROR_REASS_FRAGMENT_CHAIN_TOO_LONG,
+						       1);
+			  ip4_reass_on_timeout (vm, rm, reass);
+			  ip4_reass_free (rm, rt, reass);
+			  goto next_packet;
+			  break;
 			case IP4_REASS_RC_NO_BUF:
-			  /* fallthrough */
+			  vlib_node_increment_counter (vm, node->node_index,
+						       IP4_ERROR_REASS_NO_BUF,
+						       1);
+			  ip4_reass_on_timeout (vm, rm, reass);
+			  ip4_reass_free (rm, rt, reass);
+			  goto next_packet;
+			  break;
 			case IP4_REASS_RC_INTERNAL_ERROR:
-			  /* drop everything and start with a clean slate */
+			  vlib_node_increment_counter (vm, node->node_index,
+						       IP4_ERROR_REASS_INTERNAL_ERROR,
+						       1);
 			  ip4_reass_on_timeout (vm, rm, reass);
 			  ip4_reass_free (rm, rt, reass);
 			  goto next_packet;
@@ -1176,20 +1204,21 @@ ip4_rehash_cb (clib_bihash_kv_16_8_t * kv, void *_ctx)
 
 static void
 ip4_reass_set_params (u32 timeout_ms, u32 max_reassemblies,
-		      u32 expire_walk_interval_ms)
+		      u32 max_reassembly_length, u32 expire_walk_interval_ms)
 {
   ip4_reass_main.timeout_ms = timeout_ms;
   ip4_reass_main.timeout = (f64) timeout_ms / (f64) MSEC_PER_SEC;
   ip4_reass_main.max_reass_n = max_reassemblies;
+  ip4_reass_main.max_reass_len = max_reassembly_length;
   ip4_reass_main.expire_walk_interval_ms = expire_walk_interval_ms;
 }
 
 vnet_api_error_t
 ip4_reass_set (u32 timeout_ms, u32 max_reassemblies,
-	       u32 expire_walk_interval_ms)
+	       u32 max_reassembly_length, u32 expire_walk_interval_ms)
 {
   u32 old_nbuckets = ip4_reass_get_nbuckets ();
-  ip4_reass_set_params (timeout_ms, max_reassemblies,
+  ip4_reass_set_params (timeout_ms, max_reassemblies, max_reassembly_length,
 			expire_walk_interval_ms);
   vlib_process_signal_event (ip4_reass_main.vlib_main,
 			     ip4_reass_main.ip4_reass_expire_node_idx,
@@ -1223,10 +1252,11 @@ ip4_reass_set (u32 timeout_ms, u32 max_reassemblies,
 
 vnet_api_error_t
 ip4_reass_get (u32 * timeout_ms, u32 * max_reassemblies,
-	       u32 * expire_walk_interval_ms)
+	       u32 * max_reassembly_length, u32 * expire_walk_interval_ms)
 {
   *timeout_ms = ip4_reass_main.timeout_ms;
   *max_reassemblies = ip4_reass_main.max_reass_n;
+  *max_reassembly_length = ip4_reass_main.max_reass_len;
   *expire_walk_interval_ms = ip4_reass_main.expire_walk_interval_ms;
   return 0;
 }
@@ -1256,6 +1286,7 @@ ip4_reass_init_function (vlib_main_t * vm)
 
   ip4_reass_set_params (IP4_REASS_TIMEOUT_DEFAULT_MS,
 			IP4_REASS_MAX_REASSEMBLIES_DEFAULT,
+			IP4_REASS_MAX_REASSEMBLY_LENGTH_DEFAULT,
 			IP4_REASS_EXPIRE_WALK_INTERVAL_DEFAULT_MS);
 
   nbuckets = ip4_reass_get_nbuckets ();
diff --git a/src/vnet/ip/ip4_reassembly.h b/src/vnet/ip/ip4_reassembly.h
index 521ca0f1998..4ceb0ab2409 100644
--- a/src/vnet/ip/ip4_reassembly.h
+++ b/src/vnet/ip/ip4_reassembly.h
@@ -30,12 +30,14 @@
  * @brief set ip4 reassembly configuration
  */
 vnet_api_error_t ip4_reass_set (u32 timeout_ms, u32 max_reassemblies,
+				u32 max_reassembly_length,
 				u32 expire_walk_interval_ms);
 
 /**
  * @brief get ip4 reassembly configuration
  */
 vnet_api_error_t ip4_reass_get (u32 * timeout_ms, u32 * max_reassemblies,
+				u32 * max_reassembly_length,
 				u32 * expire_walk_interval_ms);
 
 vnet_api_error_t ip4_reass_enable_disable (u32 sw_if_index,
diff --git a/src/vnet/ip/ip6_error.h b/src/vnet/ip/ip6_error.h
index 6a20de4f18e..3ca2be61a55 100644
--- a/src/vnet/ip/ip6_error.h
+++ b/src/vnet/ip/ip6_error.h
@@ -81,6 +81,8 @@
   _ (REASS_DUPLICATE_FRAGMENT, "duplicate fragments")                   \
   _ (REASS_OVERLAPPING_FRAGMENT, "overlapping fragments")               \
   _ (REASS_LIMIT_REACHED, "drops due to concurrent reassemblies limit") \
+  _ (REASS_FRAGMENT_CHAIN_TOO_LONG, "fragment chain too long (drop)")   \
+  _ (REASS_NO_BUF, "out of buffers (drop)")                             \
   _ (REASS_TIMEOUT, "fragments dropped due to reassembly timeout")      \
   _ (REASS_INTERNAL_ERROR, "drops due to internal reassembly error")
 
diff --git a/src/vnet/ip/ip6_reassembly.c b/src/vnet/ip/ip6_reassembly.c
index 9906250cb0f..45cd2b2eaeb 100644
--- a/src/vnet/ip/ip6_reassembly.c
+++ b/src/vnet/ip/ip6_reassembly.c
@@ -30,12 +30,14 @@
 #define IP6_REASS_TIMEOUT_DEFAULT_MS 100
 #define IP6_REASS_EXPIRE_WALK_INTERVAL_DEFAULT_MS 10000	// 10 seconds default
 #define IP6_REASS_MAX_REASSEMBLIES_DEFAULT 1024
+#define IP6_REASS_MAX_REASSEMBLY_LENGTH_DEFAULT 3
 #define IP6_REASS_HT_LOAD_FACTOR (0.75)
 
 typedef enum
 {
   IP6_REASS_RC_OK,
   IP6_REASS_RC_INTERNAL_ERROR,
+  IP6_REASS_RC_TOO_MANY_FRAGMENTS,
   IP6_REASS_RC_NO_BUF,
 } ip6_reass_rc_t;
 
@@ -112,6 +114,8 @@ typedef struct
   u8 next_index;
   // minimum fragment length for this reassembly - used to estimate MTU
   u16 min_fragment_length;
+  // number of fragments for this reassembly
+  u32 fragments_n;
 } ip6_reass_t;
 
 typedef struct
@@ -128,6 +132,9 @@ typedef struct
   u32 timeout_ms;
   f64 timeout;
   u32 expire_walk_interval_ms;
+  // maximum number of fragments in one reassembly
+  u32 max_reass_len;
+  // maximum number of reassemblies
   u32 max_reass_n;
 
   // IPv6 runtime
@@ -744,6 +751,7 @@ ip6_reass_update (vlib_main_t * vm, vlib_node_runtime_t * node,
 				       *bi0);
       reass->min_fragment_length = clib_net_to_host_u16 (fip->payload_length);
       consumed = 1;
+      reass->fragments_n = 1;
       goto check_if_done_maybe;
     }
   reass->min_fragment_length =
@@ -797,6 +805,7 @@ ip6_reass_update (vlib_main_t * vm, vlib_node_runtime_t * node,
 	}
       break;
     }
+  ++reass->fragments_n;
 check_if_done_maybe:
   if (consumed)
     {
@@ -816,6 +825,10 @@ check_if_done_maybe:
       if (consumed)
 	{
 	  *bi0 = ~0;
+	  if (reass->fragments_n > rm->max_reass_len)
+	    {
+	      return IP6_REASS_RC_TOO_MANY_FRAGMENTS;
+	    }
 	}
       else
 	{
@@ -989,10 +1002,25 @@ ip6_reassembly_inline (vlib_main_t * vm,
 		case IP6_REASS_RC_OK:
 		  /* nothing to do here */
 		  break;
+		case IP6_REASS_RC_TOO_MANY_FRAGMENTS:
+		  vlib_node_increment_counter (vm, node->node_index,
+					       IP6_ERROR_REASS_FRAGMENT_CHAIN_TOO_LONG,
+					       1);
+		  ip6_reass_drop_all (vm, rm, reass);
+		  ip6_reass_free (rm, rt, reass);
+		  goto next_packet;
+		  break;
 		case IP6_REASS_RC_NO_BUF:
-		  /* fallthrough */
+		  vlib_node_increment_counter (vm, node->node_index,
+					       IP6_ERROR_REASS_NO_BUF, 1);
+		  ip6_reass_drop_all (vm, rm, reass);
+		  ip6_reass_free (rm, rt, reass);
+		  goto next_packet;
+		  break;
 		case IP6_REASS_RC_INTERNAL_ERROR:
-		  /* drop everything and start with a clean slate */
+		  vlib_node_increment_counter (vm, node->node_index,
+					       IP6_ERROR_REASS_INTERNAL_ERROR,
+					       1);
 		  ip6_reass_drop_all (vm, rm, reass);
 		  ip6_reass_free (rm, rt, reass);
 		  goto next_packet;
@@ -1151,20 +1179,21 @@ ip6_rehash_cb (clib_bihash_kv_48_8_t * kv, void *_ctx)
 
 static void
 ip6_reass_set_params (u32 timeout_ms, u32 max_reassemblies,
-		      u32 expire_walk_interval_ms)
+		      u32 max_reassembly_length, u32 expire_walk_interval_ms)
 {
   ip6_reass_main.timeout_ms = timeout_ms;
   ip6_reass_main.timeout = (f64) timeout_ms / (f64) MSEC_PER_SEC;
   ip6_reass_main.max_reass_n = max_reassemblies;
+  ip6_reass_main.max_reass_len = max_reassembly_length;
   ip6_reass_main.expire_walk_interval_ms = expire_walk_interval_ms;
 }
 
 vnet_api_error_t
 ip6_reass_set (u32 timeout_ms, u32 max_reassemblies,
-	       u32 expire_walk_interval_ms)
+	       u32 max_reassembly_length, u32 expire_walk_interval_ms)
 {
   u32 old_nbuckets = ip6_reass_get_nbuckets ();
-  ip6_reass_set_params (timeout_ms, max_reassemblies,
+  ip6_reass_set_params (timeout_ms, max_reassemblies, max_reassembly_length,
 			expire_walk_interval_ms);
   vlib_process_signal_event (ip6_reass_main.vlib_main,
 			     ip6_reass_main.ip6_reass_expire_node_idx,
@@ -1231,6 +1260,7 @@ ip6_reass_init_function (vlib_main_t * vm)
 
   ip6_reass_set_params (IP6_REASS_TIMEOUT_DEFAULT_MS,
 			IP6_REASS_MAX_REASSEMBLIES_DEFAULT,
+			IP6_REASS_MAX_REASSEMBLY_LENGTH_DEFAULT,
 			IP6_REASS_EXPIRE_WALK_INTERVAL_DEFAULT_MS);
 
   nbuckets = ip6_reass_get_nbuckets ();
diff --git a/src/vnet/ip/ip6_reassembly.h b/src/vnet/ip/ip6_reassembly.h
index 5084edaaf8c..1ca2b20813c 100644
--- a/src/vnet/ip/ip6_reassembly.h
+++ b/src/vnet/ip/ip6_reassembly.h
@@ -30,6 +30,7 @@
  * @brief set ip6 reassembly configuration
  */
 vnet_api_error_t ip6_reass_set (u32 timeout_ms, u32 max_reassemblies,
+				u32 max_reassembly_length,
 				u32 expire_walk_interval_ms);
 
 /**
diff --git a/src/vnet/ip/ip_api.c b/src/vnet/ip/ip_api.c
index ce3456d77d9..5a6053d1f42 100644
--- a/src/vnet/ip/ip_api.c
+++ b/src/vnet/ip/ip_api.c
@@ -3328,12 +3328,14 @@ vl_api_ip_reassembly_set_t_handler (vl_api_ip_reassembly_set_t * mp)
     {
       rv = ip6_reass_set (clib_net_to_host_u32 (mp->timeout_ms),
 			  clib_net_to_host_u32 (mp->max_reassemblies),
+			  clib_net_to_host_u32 (mp->max_reassembly_length),
 			  clib_net_to_host_u32 (mp->expire_walk_interval_ms));
     }
   else
     {
       rv = ip4_reass_set (clib_net_to_host_u32 (mp->timeout_ms),
 			  clib_net_to_host_u32 (mp->max_reassemblies),
+			  clib_net_to_host_u32 (mp->max_reassembly_length),
 			  clib_net_to_host_u32 (mp->expire_walk_interval_ms));
     }
 
@@ -3364,6 +3366,7 @@ vl_api_ip_reassembly_get_t_handler (vl_api_ip_reassembly_get_t * mp)
     {
       rmp->is_ip6 = 0;
       ip4_reass_get (&rmp->timeout_ms, &rmp->max_reassemblies,
+		     &rmp->max_reassembly_length,
 		     &rmp->expire_walk_interval_ms);
     }
   rmp->timeout_ms = clib_host_to_net_u32 (rmp->timeout_ms);