3 files changed, 48 insertions, 1 deletions
diff --git a/proto/ospf/iface.c b/proto/ospf/iface.c
index b30a8112..83d591b4 100644
--- a/proto/ospf/iface.c
+++ b/proto/ospf/iface.c
@@ -97,6 +97,27 @@ ospf_sk_open(struct ospf_iface *ifa)
     goto err;
 #endif
 
+  /*
+   * For OSPFv2: When sending a packet, it is important to have a
+   * proper source address. We expect that when we send one-hop
+   * unicast packets, OS chooses a source address according to the
+   * destination address (to be in the same prefix). We also expect
+   * that when we send multicast packets, OS uses the source address
+   * from sk->saddr registered to OS by sk_setup_multicast(). This
+   * behavior is needed to implement multiple virtual ifaces (struct
+   * ospf_iface) on one physical iface and is signalized by
+   * CONFIG_MC_PROPER_SRC.
+   *
+   * If this behavior is not available (for example on BSD), we create
+   * non-stub iface just for the primary IP address (see
+   * ospf_iface_stubby()) and we expect OS to use primary IP address
+   * as a source address for both unicast and multicast packets.
+   *
+   * FIXME: the primary IP address is currently just the
+   * lexicographically smallest address on an interface, it should be
+   * signalized by sysdep code which one is really the primary.
+   */
+
   sk->saddr = ifa->addr->ip;
   if (sk_setup_multicast(sk) < 0)
     goto err;
diff --git a/proto/ospf/ospf.c b/proto/ospf/ospf.c
index e8245110..d345e49b 100644
--- a/proto/ospf/ospf.c
+++ b/proto/ospf/ospf.c
@@ -42,6 +42,32 @@
  * and deletion. Each LSA is kept in two pieces: header and body. Both of them are
  * kept in the endianity of the CPU.
  *
+ * In OSPFv2 specification, it is implied that there is one IP prefix
+ * for each physical network/interface (unless it is an ptp link). But
+ * in modern systems, there might be more independent IP prefixes
+ * associated with an interface.  To handle this situation, we have
+ * one &ospf_iface for each active IP prefix (instead for each active
+ * iface); This behaves like virtual interface for the purpose of OSPF.
+ * If we receive packet, we associate it with a proper virtual interface
+ * mainly according to its source address.
+ *
+ * OSPF keeps one socket per &ospf_iface. This allows us (compared to
+ * one socket approach) to evade problems with a limit of multicast
+ * groups per socket and with sending multicast packets to appropriate
+ * interface in a portable way. The socket is associated with
+ * underlying physical iface and should not receive packets received
+ * on other ifaces (unfortunately, this is not true on
+ * BSD). Generally, one packet can be received by more sockets (for
+ * example, if there are more &ospf_iface on one physical iface),
+ * therefore we explicitly filter received packets according to
+ * src/dst IP address and received iface.
+ *
+ * Vlinks are implemented using particularly degenerate form of
+ * &ospf_iface, which has several exceptions: it does not have its
+ * iface or socket (it copies these from 'parent' &ospf_iface) and it
+ * is present in iface list even when down (it is not freed in
+ * ospf_iface_down()).
+ *
  * The heart beat of ospf is ospf_disp(). It is called at regular intervals
  * (&proto_ospf->tick). It is responsible for aging and flushing of LSAs in
  * the database, for routing table calculaction and it call area_disp() of every
diff --git a/proto/ospf/ospf.h b/proto/ospf/ospf.h
index ef51b9fd..bb0b6afc 100644
--- a/proto/ospf/ospf.h
+++ b/proto/ospf/ospf.h
@@ -159,7 +159,7 @@ struct ospf_iface
 {
   node n;
   struct iface *iface;		/* Nest's iface */
-  struct ifa *addr;
+  struct ifa *addr;		/* IP prefix associated with that OSPF iface */
   struct ospf_area *oa;
   struct object_lock *lock;
   sock *sk;			/* IP socket (for DD ...) */