Handling of TCPv4 fragments

With support for refragmenting too big fragments

src/nat.c

@@ -41,8 +41,14 @@ struct s_radixtree_nat4 {
 	unsigned char		zeros;		/* unused space */
 } __attribute__ ((__packed__));
 
+struct s_radixtree_fragments4 {
+	struct s_ipv4_addr	addr;
+	unsigned short		id;
+} __attribute__ ((__packed__));
+
 radixtree_t *nat6_tcp, *nat6_udp, *nat6_icmp,
-	    *nat4_tcp, *nat4_udp, *nat4_icmp;
+	    *nat4_tcp, *nat4_udp, *nat4_icmp,
+	    *nat4_tcp_fragments;
 
 void nat_init(void)
 {
@@ -53,6 +59,8 @@ void nat_init(void)
 	nat4_tcp  = radixtree_create();
 	nat4_udp  = radixtree_create();
 	nat4_icmp = radixtree_create();
+
+	nat4_tcp_fragments = radixtree_create();
 }
 
 void nat_quit(void)
@@ -66,6 +74,9 @@ void nat_quit(void)
 	radixtree_destroy(nat4_tcp,  12);
 	radixtree_destroy(nat4_udp,  12);
 	radixtree_destroy(nat4_icmp, 12);
+
+	/* 32 + 16 = 48 / 6 = 8 */
+	radixtree_destroy(nat4_tcp_fragments, 8);
 }
 
 struct s_nat *nat_out(radixtree_t *nat_proto6, radixtree_t *nat_proto4,
@@ -147,3 +158,84 @@ struct s_nat *nat_in(radixtree_t *nat_proto4, struct s_ipv4_addr ipv4_src,
 		return result;
 	}
 }
+
+/**
+ * Save and retrieve NATted connections via fragment identification.
+ *
+ * @param	nat_proto4	Radix tree of fragments
+ * @param	ipv4_src	Source IPv4 address
+ * @param	id		Fragment identification
+ * @param	nat		Connection to save
+ *
+ * @return			Connection
+ */
+struct s_nat *nat_in_fragments(radixtree_t *nat_proto4,
+			       struct s_ipv4_addr ipv4_src,
+			       unsigned short id, struct s_nat *nat)
+{
+	struct s_nat *result;
+
+	/* create structure to search in the tree */
+	struct s_radixtree_fragments4 radixsearch4;
+	radixsearch4.addr = ipv4_src;
+	radixsearch4.id = id;
+
+	if ((result = (struct s_nat *) radixtree_lookup(nat_proto4,
+	    radixtree_chunker, &radixsearch4, sizeof(radixsearch4))) == NULL) {
+		if (nat != NULL) {
+			/* when fragmentation is not found, add one */
+			radixtree_insert(nat_proto4, radixtree_chunker,
+					 &radixsearch4, sizeof(radixsearch4),
+					 nat);
+			return nat;
+		} else {
+			/* nothing found and nothing to be added */
+			return NULL;
+		}
+	} else {
+		if (nat != NULL) {
+			/* when fragmentation is found, is it the same? */
+			if (result == nat) {
+				/* OK, return */
+				return result;
+			} else {
+				/* sender determines usage of IDs, so this one
+				 * shouldn't be used anymore for that
+				 * connection */
+				radixtree_delete(nat_proto4, radixtree_chunker,
+						 &radixsearch4,
+						 sizeof(radixsearch4));
+				radixtree_insert(nat_proto4, radixtree_chunker,
+						 &radixsearch4,
+						 sizeof(radixsearch4), nat);
+				return nat;
+			}
+		} else {
+			/* refresh it's connection and return */
+			result->last_packet = time(NULL);
+			return result;
+		}
+	}
+}
+
+/**
+ * Remove one entry from "fragment NAT".
+ *
+ * @param	nat_proto4	Radix tree of fragments
+ * @param	ipv4_src	Source IPv4 address
+ * @param	id		Fragment identification
+ */
+void nat_in_fragments_clenup(radixtree_t *nat_proto4,
+			     struct s_ipv4_addr ipv4_src, unsigned short id)
+{
+	/* create structure to search in the tree */
+	struct s_radixtree_fragments4 radixsearch4;
+	radixsearch4.addr = ipv4_src;
+	radixsearch4.id = id;
+
+	if (radixtree_lookup(nat_proto4, radixtree_chunker, &radixsearch4,
+	    sizeof(radixsearch4)) != NULL) {
+		radixtree_delete(nat_proto4, radixtree_chunker, &radixsearch4,
+				 sizeof(radixsearch4));
+	}
+}

src/nat.h

@@ -38,7 +38,8 @@ struct s_nat {
 };
 
 extern radixtree_t *nat6_tcp, *nat6_udp, *nat6_icmp,
-		   *nat4_tcp, *nat4_udp, *nat4_icmp;
+		   *nat4_tcp, *nat4_udp, *nat4_icmp,
+		   *nat4_tcp_fragments;
 
 void nat_init(void);
 void nat_quit(void);
@@ -49,5 +50,10 @@ struct s_nat *nat_out(radixtree_t *nat_proto6, radixtree_t *nat_proto4,
 		      unsigned short	 port_src, unsigned short     port_dst);
 struct s_nat *nat_in(radixtree_t *nat_proto4, struct s_ipv4_addr ipv4_src,
 		     unsigned short port_src, unsigned short port_dst);
+struct s_nat *nat_in_fragments(radixtree_t *nat_proto4,
+			       struct s_ipv4_addr ipv4_src,
+			       unsigned short id, struct s_nat *nat);
+void nat_in_fragments_clenup(radixtree_t *nat_proto4,
+			     struct s_ipv4_addr ipv4_src, unsigned short id);
 
 #endif /* NAT_H */

src/tcp.c

@@ -93,6 +93,12 @@ int tcp_ipv4(struct s_ethernet *eth4, struct s_ipv4 *ip4, char *payload,
 			return 1;
 		}
 
+		/* if it's fragmented, save it to fragments table */
+		if (ip4->flags_offset & htons(IPV4_FLAG_MORE_FRAGMENTS)) {
+			nat_in_fragments(nat4_tcp_fragments, ip4->ip_src,
+					 ip4->id, connection);
+		}
+
 		/* allocate enough memory for translated packet */
 		if ((packet = (unsigned char *) malloc(
 		    payload_size > MTU - sizeof(struct s_ipv6) ?
@@ -190,11 +196,126 @@ int tcp_ipv4(struct s_ethernet *eth4, struct s_ipv4 *ip4, char *payload,
 				     sizeof(struct s_ipv6) + payload_size);
 		}
 	} else {
-		/* TODO: handle all fragments */
+		/* find connection in fragments table */
-		printf("[Debug] Can't handle fragments :c(\n");
+		connection = nat_in_fragments(nat4_tcp_fragments, ip4->ip_src,
+					      ip4->id, NULL);
+
+		if (connection == NULL) {
+			printf("[Debug] Incoming connection wasn't found in "
+			       "fragments table\n");
 			return 1;
 		}
 
+		/* allocate enough memory for translated packet */
+		if ((packet = (unsigned char *) malloc(
+		    payload_size > MTU - sizeof(struct s_ipv6) -
+		    sizeof(struct s_ipv6_fragment) ?
+		    MTU + sizeof(struct s_ethernet) :
+		    sizeof(struct s_ethernet) + sizeof(struct s_ipv6) +
+		    sizeof(struct s_ipv6_fragment) + payload_size)) == NULL) {
+			fprintf(stderr, "[Error] Lack of free memory\n");
+			return 1;
+		}
+		eth6 = (struct s_ethernet *) packet;
+		ip6 = (struct s_ipv6 *) (packet + sizeof(struct s_ethernet));
+		frag = (struct s_ipv6_fragment *) ((unsigned char *) ip6 +
+						   sizeof(struct s_ipv6));
+
+		/* build ethernet header */
+		eth6->dest		= connection->mac;
+		eth6->src		= mac;
+		eth6->type		= htons(ETHERTYPE_IPV6);
+
+		/* build IPv6 header */
+		ip6->ver		= 0x60 | (ip4->tos >> 4);
+		ip6->traffic_class	= ip4->tos << 4;
+		ip6->flow_label		= 0x0;
+		ip6->hop_limit		= ip4->ttl;
+		ip6->next_header	= IPPROTO_FRAGMENT;
+		ipv4_to_ipv6(&ip4->ip_src, &ip6->ip_src);
+		memcpy(&ip6->ip_dest, &connection->ipv6,
+		       sizeof(struct s_ipv6_addr));
+
+		/* build IPv6 fragment header */
+		frag->next_header	= IPPROTO_TCP;
+		frag->zeros		= 0x0;
+		frag->id		= htonl(htons(ip4->id));
+
+		/* fragment the fragment or not? */
+		if (payload_size > MTU - sizeof(struct s_ipv6) -
+		    sizeof(struct s_ipv6_fragment)) {
+			/* fill in missing IPv6 header fields */
+			ip6->len = htons(FRAGMENT_LEN +
+					 sizeof(struct s_ipv6_fragment));
+
+			/* fill in missing IPv6 fragment header fields */
+			frag->offset_flag = htons((htons(ip4->flags_offset) << 3) |
+						  IPV6_FLAG_MORE_FRAGMENTS);
+
+			/* copy the payload data */
+			memcpy((unsigned char *) frag +
+			       sizeof(struct s_ipv6_fragment),
+			       payload, FRAGMENT_LEN);
+
+			/* send translated packet */
+			transmit_raw(packet, sizeof(struct s_ethernet) + MTU);
+
+			/* create the second fragment */
+			ip6->len = htons(payload_size +
+					 sizeof(struct s_ipv6_fragment) -
+					 FRAGMENT_LEN);
+			frag->offset_flag = htons((htons(ip4->flags_offset) +
+						  FRAGMENT_LEN / 8) << 3);
+			if (ip4->flags_offset &
+			    htons(IPV4_FLAG_MORE_FRAGMENTS)) {
+				frag->offset_flag |=
+					htons(IPV6_FLAG_MORE_FRAGMENTS);
+			}
+
+			/* copy the payload data */
+			memcpy((unsigned char *) frag +
+			       sizeof(struct s_ipv6_fragment),
+			       payload + FRAGMENT_LEN,
+			       payload_size - FRAGMENT_LEN);
+
+			/* send translated packet */
+			transmit_raw(packet, sizeof(struct s_ethernet) +
+					     sizeof(struct s_ipv6) +
+					     sizeof(struct s_ipv6_fragment) -
+					     FRAGMENT_LEN + payload_size);
+		} else {
+			/* fill in missing IPv6 header fields */
+			ip6->len = htons(payload_size +
+					 sizeof(struct s_ipv6_fragment));
+
+			/* fill in missing IPv6 fragment header fields */
+			frag->offset_flag = htons(htons(ip4->flags_offset) << 3);
+			if (ip4->flags_offset &
+			    htons(IPV4_FLAG_MORE_FRAGMENTS)) {
+				frag->offset_flag |=
+					htons(IPV6_FLAG_MORE_FRAGMENTS);
+			}
+
+			/* copy the payload data */
+			memcpy((unsigned char *) ip6 + sizeof(struct s_ipv6) +
+			       sizeof(struct s_ipv6_fragment),
+			       payload, payload_size);
+
+			/* send translated packet */
+			transmit_raw(packet, sizeof(struct s_ethernet) +
+				     sizeof(struct s_ipv6) +
+				     sizeof(struct s_ipv6_fragment) +
+				     payload_size);
+		}
+
+		/* if this is the last fragment, remove the entry from table */
+		if (!(ip4->flags_offset & htons(IPV4_FLAG_MORE_FRAGMENTS))) {
+			printf("[Debug] Removing fragment entry\n");
+			nat_in_fragments_clenup(nat4_tcp_fragments,
+						ip4->ip_src, ip4->id);
+		}
+	}
+
 	/* clean-up */
 	free(packet);