ip: add extended shallow reassembly

This patch adds some fixes and improvements:

Fixes bug where save_rewrite_length gets overwritten on reassembly
handoff.

Fixes bug where duplicate fragments could cause a reassembly context
to be lost, because the race losing thread would remove bihash entry
created by winning thread.

Improves tracing by adding more events.

Adds extended shallow reassembly. This is a toggleable option, which if
turned on will cause reassembly to wait for both first and last
fragments to calculate total IP payload length. Furthermore it'll store
a local copy of first fragment and necessary data to retrieve it in
vnet_buffer2. This allows downstream features to access full L3/L4
headers when dealing with fragments.

Type: fix
Change-Id: I81695070533410c5815291dbc65ea71c87e3ae05
Signed-off-by: Klement Sekera <klement.sekera@gmail.com>

src/plugins/map/ip6_map_t.c

@@ -151,9 +151,8 @@ ip6_map_t_icmp (vlib_main_t * vm,
 			       vnet_buffer (p0)->map_t.map_domain_index);
 	  ctx0.d = d0;
 	  ctx0.sender_port = 0;
-	  if (!ip6_get_port
-	      (vm, p0, ip60, p0->current_length, NULL, &ctx0.sender_port,
-	       NULL, NULL, NULL, NULL))
+	  if (!ip6_get_port (vm, p0, ip60, p0->current_length, NULL,
+			     &ctx0.sender_port, NULL, NULL, NULL, NULL, NULL))
 	    {
 	      // In case of 1:1 mapping, we don't care about the port
 	      if (!(d0->ea_bits_len == 0 && d0->rules))

src/vnet/buffer.h

@@ -241,7 +241,8 @@ typedef struct
 		u8 ip_proto;	/* protocol in ip header */
 		u8 icmp_type_or_tcp_flags;
 		u8 is_non_first_fragment : 1;
-		u8 l4_layer_truncated : 7;
+		u8 l4_hdr_truncated : 1;
+		u8 unused : 6;
 		u32 tcp_seq_number;
 	      };
 	      /* full reassembly output variables */
@@ -492,7 +493,22 @@ typedef struct
     };
   } nat;
 
-  u32 unused[8];
+  struct
+  {
+    /*
+     * Shallow virtual reassembly output values.
+     * Only populated if extended reassembly enabled via
+     * ipX_sv_reass_enable_disable_extended().
+     */
+    struct
+    {
+      u32 thread_index;
+      u32 pool_index;
+      u32 id;
+    } reass;
+  } ip;
+
+  u32 unused[5];
 } vnet_buffer_opaque2_t;
 
 #define vnet_buffer2(b) ((vnet_buffer_opaque2_t *) (b)->opaque2)

src/vnet/ip/ip6_to_ip4.h

@@ -96,10 +96,10 @@ ip6_parse (vlib_main_t *vm, vlib_buffer_t *b, ip6_header_t *ip6, u32 buff_len,
  * @returns 1 on success, 0 otherwise.
  */
 always_inline u16
-ip6_get_port (vlib_main_t * vm, vlib_buffer_t * b, ip6_header_t * ip6,
-	      u16 buffer_len, u8 * ip_protocol, u16 * src_port,
-	      u16 * dst_port, u8 * icmp_type_or_tcp_flags,
-	      u32 * tcp_ack_number, u32 * tcp_seq_number)
+ip6_get_port (vlib_main_t *vm, vlib_buffer_t *b, ip6_header_t *ip6,
+	      u16 buffer_len, u8 *ip_protocol, u16 *src_port, u16 *dst_port,
+	      u8 *icmp_type_or_tcp_flags, u32 *tcp_ack_number,
+	      u32 *tcp_seq_number, void **l4_hdr)
 {
   u8 l4_protocol;
   u16 l4_offset;
@@ -120,8 +120,19 @@ ip6_get_port (vlib_main_t * vm, vlib_buffer_t * b, ip6_header_t * ip6,
       *ip_protocol = l4_protocol;
     }
   l4 = u8_ptr_add (ip6, l4_offset);
+  if (l4_hdr)
+    *l4_hdr = l4;
   if (l4_protocol == IP_PROTOCOL_TCP || l4_protocol == IP_PROTOCOL_UDP)
     {
+      if ((IP_PROTOCOL_UDP == l4_protocol &&
+	   u8_ptr_add (l4, sizeof (udp_header_t)) >
+	     u8_ptr_add (vlib_buffer_get_current (b), b->current_length)) ||
+	  (IP_PROTOCOL_TCP == l4_protocol &&
+	   u8_ptr_add (l4, sizeof (tcp_header_t)) >
+	     u8_ptr_add (vlib_buffer_get_current (b), b->current_length)))
+	{
+	  return 0;
+	}
       if (src_port)
 	*src_port = ((udp_header_t *) (l4))->src_port;
       if (dst_port)
@@ -135,6 +146,11 @@ ip6_get_port (vlib_main_t * vm, vlib_buffer_t * b, ip6_header_t * ip6,
     }
   else if (l4_protocol == IP_PROTOCOL_ICMP6)
     {
+      if (u8_ptr_add (l4, sizeof (icmp46_header_t)) >
+	  u8_ptr_add (vlib_buffer_get_current (b), b->current_length))
+	{
+	  return 0;
+	}
       icmp46_header_t *icmp = (icmp46_header_t *) (l4);
       if (icmp_type_or_tcp_flags)
 	*icmp_type_or_tcp_flags = ((icmp46_header_t *) (l4))->type;

src/vnet/ip/reass/ip4_sv_reass.h

@@ -23,6 +23,7 @@
 #ifndef __included_ip4_sv_reass_h__
 #define __included_ip4_sv_reass_h__
 
+#include <stdbool.h>
 #include <vnet/api_errno.h>
 #include <vnet/vnet.h>
 
@@ -48,6 +49,33 @@ int ip4_sv_reass_enable_disable_with_refcnt (u32 sw_if_index, int is_enable);
 int ip4_sv_reass_output_enable_disable_with_refcnt (u32 sw_if_index,
 						    int is_enable);
 
+/*
+ * Enable or disable extended reassembly.
+ *
+ * Extended reassembly means that fragments are cached until both first and
+ * last fragments are seen. Furthermore, first fragment buffer will be cloned
+ * and stored in reassembly context for later retrieval.
+ */
+void ip4_sv_reass_enable_disable_extended (bool is_enable);
+
+struct ip4_sv_lock_unlock_args
+{
+  u32 *total_ip_payload_length;
+  u32 *first_fragment_buffer_index;
+  u32 *first_fragment_total_ip_header_length;
+};
+
+/*
+ * Lock thread-level lock and fetch information from reassembly context.
+ * Uses vnet_buffer2 data filled by extended reassembly.
+ *
+ * Returns 0 on success, -1 otherwise.
+ */
+int ip4_sv_reass_extended_lock (vlib_buffer_t *b,
+				struct ip4_sv_lock_unlock_args *a);
+
+void ip4_sv_reass_extended_unlock (vlib_buffer_t *b);
+
 uword ip4_sv_reass_custom_register_next_node (uword node_index);
 uword ip4_sv_reass_custom_context_register_next_node (uword node_index);
 

src/vnet/ip/reass/ip6_sv_reass.h

@@ -23,6 +23,7 @@
 #ifndef __included_ip6_sv_reass_h__
 #define __included_ip6_sv_reass_h__
 
+#include <stdbool.h>
 #include <vnet/api_errno.h>
 #include <vnet/vnet.h>
 
@@ -46,6 +47,33 @@ vnet_api_error_t
 ip6_sv_reass_output_enable_disable_with_refcnt (u32 sw_if_index,
 						int is_enable);
 
+/*
+ * Enable or disable extended reassembly.
+ *
+ * Extended reassembly means that fragments are cached until both first and
+ * last fragments are seen. Furthermore, first fragment buffer will be cloned
+ * and stored in reassembly context for later retrieval.
+ */
+void ip6_sv_reass_enable_disable_extended (bool is_enable);
+
+struct ip6_sv_lock_unlock_args
+{
+  u32 *total_ip_payload_length;
+  u32 *first_fragment_buffer_index;
+  u32 *first_fragment_total_ip_header_length;
+};
+
+/*
+ * Lock thread-level lock and fetch information from reassembly context.
+ * Uses vnet_buffer2 data filled by extended reassembly.
+ *
+ * Returns 0 on success, -1 otherwise.
+ */
+int ip6_sv_reass_extended_lock (vlib_buffer_t *b,
+				struct ip6_sv_lock_unlock_args *a);
+
+void ip6_sv_reass_extended_unlock (vlib_buffer_t *b);
+
 int ip6_sv_reass_enable_disable_with_refcnt (u32 sw_if_index, int is_enable);
 uword ip6_sv_reass_custom_context_register_next_node (uword node_index);