Implements hostcache and recursive next hops.

Hostcache is a structure for monitoring changes in a routing table that
is used for routes with dynamic/recursive next hops. This is needed for
proper iBGP next hop handling.

14 files changed, 511 insertions, 81 deletions

diff --git a/nest/proto.c b/nest/proto.c
index c9e2f5c7..16ec3f9b 100644
--- a/nest/proto.c
+++ b/nest/proto.c
@@ -577,10 +577,8 @@ proto_fell_down(struct proto *p)
   bzero(&p->stats, sizeof(struct proto_stats));
   rt_unlock_table(p->table);
 
-#ifdef CONFIG_PIPE
-  if (proto_is_pipe(p))
-    rt_unlock_table(pipe_get_peer_table(p));
-#endif
+  if (p->proto->cleanup)
+    p->proto->cleanup(p);
 
   proto_rethink_goal(p);
 }
diff --git a/nest/protocol.h b/nest/protocol.h
index 5dac2a9b..70999f0e 100644
--- a/nest/protocol.h
+++ b/nest/protocol.h
@@ -48,6 +48,7 @@ struct protocol {
   void (*dump_attrs)(struct rte *);		/* Dump protocol-dependent attributes */
   int (*start)(struct proto *);			/* Start the instance */
   int (*shutdown)(struct proto *);		/* Stop the instance */
+  void (*cleanup)(struct proto *);		/* Called after shutdown when protocol became hungry/down */
   void (*get_status)(struct proto *, byte *buf); /* Get instance status (for `show protocols' command) */
   void (*get_route_info)(struct rte *, byte *buf, struct ea_list *attrs); /* Get route information (for `show route' command) */
   int (*get_attr)(struct eattr *, byte *buf, int buflen);	/* ASCIIfy dynamic attribute (returns GA_*) */
diff --git a/nest/route.h b/nest/route.h
index 99803e7c..e599f5b9 100644
--- a/nest/route.h
+++ b/nest/route.h
@@ -129,14 +129,19 @@ typedef struct rtable {
   list hooks;				/* List of announcement hooks */
   int pipe_busy;			/* Pipe loop detection */
   int use_count;			/* Number of protocols using this table */
+  struct hostcache *hostcache;
   struct rtable_config *config;		/* Configuration of this table */
   struct config *deleted;		/* Table doesn't exist in current configuration,
 					 * delete as soon as use_count becomes 0 and remove
 					 * obstacle from this routing table.
 					 */
-  struct event *gc_event;		/* Garbage collector event */
+  struct event *rt_event;		/* Routing table event */
   int gc_counter;			/* Number of operations since last GC */
   bird_clock_t gc_time;			/* Time of last GC */
+  byte gc_scheduled;			/* GC is scheduled */
+  byte hcu_scheduled;			/* Hostcache update is scheduled */
+  byte nhu_state;			/* Next Hop Update state */
+  struct fib_iterator nhu_fit;		/* Next Hop Update FIB iterator */
 } rtable;
 
 typedef struct network {
@@ -144,6 +149,23 @@ typedef struct network {
   struct rte *routes;			/* Available routes for this network */
 } net;
 
+struct hostcache {
+  struct fib htable;
+  list hostentries;
+  byte update_hostcache;
+};
+
+struct hostentry {
+  struct fib_node fn;
+  node ln;
+  unsigned uc;				/* Use count */
+  struct iface *iface;			/* Chosen outgoing interface */
+  ip_addr gw;				/* Chosen next hop */
+  byte dest;				/* Chosen route destination type (RTD_...) */
+  byte pxlen;				/* Pxlen from net that matches route */
+  struct rtable *tab;
+};
+
 typedef struct rte {
   struct rte *next;
   net *net;				/* Network this RTE belongs to */
@@ -207,7 +229,6 @@ void rt_dump(rtable *);
 void rt_dump_all(void);
 int rt_feed_baby(struct proto *p);
 void rt_feed_baby_abort(struct proto *p);
-void rt_prune(rtable *tab);
 void rt_prune_all(void);
 struct rtable_config *rt_new_table(struct symbol *s);
 
@@ -248,6 +269,7 @@ typedef struct rta {
   u16 hash_key;				/* Hash over important fields */
   ip_addr gw;				/* Next hop */
   ip_addr from;				/* Advertising router */
+  struct hostentry *hostentry;		/* Hostentry for recursive next-hops */
   struct iface *iface;			/* Outgoing interface */
   struct ea_list *eattrs;		/* Extended Attribute chain */
 } rta;
@@ -357,6 +379,25 @@ static inline void rta_free(rta *r) { if (r && !--r->uc) rta__free(r); }
 void rta_dump(rta *);
 void rta_dump_all(void);
 void rta_show(struct cli *, rta *, ea_list *);
+void rta_set_recursive_next_hop(rtable *dep, rta *a, rtable *tab, ip_addr *gw);
+
+/*
+ * rta_set_recursive_next_hop() acquires hostentry from hostcache and
+ * fills rta->hostentry field.  New hostentry has zero use
+ * count. Cached rta locks its hostentry (increases its use count),
+ * uncached rta does not lock it. Hostentry with zero use count is
+ * removed asynchronously during host cache update, therefore it is
+ * safe to hold such hostentry temorarily. There is no need to hold
+ * a lock for hostentry->dep table, because that table contains routes
+ * responsible for that hostentry, and therefore is non-empty if given
+ * hostentry has non-zero use count. The protocol responsible for routes
+ * with recursive next hops should also hold a lock for a table governing
+ * that routes (argument tab to rta_set_recursive_next_hop()).
+ */
+
+static inline void rt_lock_hostentry(struct hostentry *he) { if (he) he->uc++; }
+static inline void rt_unlock_hostentry(struct hostentry *he) { if (he) he->uc--; }
+
 
 extern struct protocol *attr_class_to_protocol[EAP_MAX];
 
diff --git a/nest/rt-attr.c b/nest/rt-attr.c
index abd49c70..9caee8d5 100644
--- a/nest/rt-attr.c
+++ b/nest/rt-attr.c
@@ -671,6 +671,7 @@ rta_lookup(rta *o)
   r = rta_copy(o);
   r->hash_key = h;
   r->aflags = RTAF_CACHED;
+  rt_lock_hostentry(r->hostentry);
   rta_insert(r);
 
   if (++rta_cache_count > rta_cache_limit)
@@ -688,6 +689,7 @@ rta__free(rta *a)
   if (a->next)
     a->next->pprev = a->pprev;
   a->aflags = 0;		/* Poison the entry */
+  rt_unlock_hostentry(a->hostentry);
   ea_free(a->eattrs);
   sl_free(rta_slab, a);
 }
diff --git a/nest/rt-table.c b/nest/rt-table.c
index 8cca42a7..b5256945 100644
--- a/nest/rt-table.c
+++ b/nest/rt-table.c
@@ -51,6 +51,13 @@ static linpool *rte_update_pool;
 static list routing_tables;
 
 static void rt_format_via(rte *e, byte *via);
+static void rt_free_hostcache(rtable *tab);
+static void rt_notify_hostcache(rtable *tab, net *net);
+static void rt_update_hostcache(rtable *tab);
+static void rt_next_hop_update(rtable *tab);
+static void rt_prune(rtable *tab);
+
+static inline void rt_schedule_gc(rtable *tab);
 
 static void
 rte_init(struct fib_node *N)
@@ -210,7 +217,7 @@ do_rte_announce(struct announce_hook *a, int type UNUSED, net *net, rte *new, rt
    * This is a tricky part - we don't know whether route 'old' was
    * exported to protocol 'p' or was filtered by the export filter.
    * We try tu run the export filter to know this to have a correct
-   * value in 'old' argument of rt_update (and proper filter value)
+   * value in 'old' argument of rte_update (and proper filter value)
    *
    * FIXME - this is broken because 'configure soft' may change
    * filters but keep routes. Refeed is expected to be called after
@@ -327,6 +334,9 @@ rte_announce(rtable *tab, unsigned type, net *net, rte *new, rte *old, ea_list *
 	new->attrs->proto->stats.pref_routes++;
       if (old)
 	old->attrs->proto->stats.pref_routes--;
+
+      if (tab->hostcache)
+	rt_notify_hostcache(tab, net);
     }
 
   WALK_LIST(a, tab->hooks)
@@ -337,6 +347,7 @@ rte_announce(rtable *tab, unsigned type, net *net, rte *new, rte *old, ea_list *
     }
 }
 
+
 static inline int
 rte_validate(rte *e)
 {
@@ -469,7 +480,6 @@ rte_recalculate(rtable *table, net *net, struct proto *p, struct proto *src, rte
 
   rte_announce(table, RA_ANY, net, new, old, tmpa);
 
-  
   if (new && rte_better(new, old_best))
     {
       /* The first case - the new route is cleary optimal, we link it
@@ -523,7 +533,7 @@ rte_recalculate(rtable *table, net *net, struct proto *p, struct proto *src, rte
 	}
       else if (table->gc_counter++ >= table->config->gc_max_ops &&
 	       table->gc_time + table->config->gc_min_time <= now)
-	ev_schedule(table->gc_event);
+	rt_schedule_gc(table);
     }
   else if (new)
     {
@@ -688,6 +698,21 @@ drop:
   rte_update_unlock();
 }
 
+/* Independent call to rte_announce(), used from next hop
+   recalculation, outside of rte_update(). new must be non-NULL */
+static inline void 
+rte_announce_i(rtable *tab, unsigned type, net *n, rte *new, rte *old)
+{
+  struct proto *src;
+  ea_list *tmpa;
+
+  rte_update_lock();
+  src = new->attrs->proto;
+  tmpa = src->make_tmp_attrs ? src->make_tmp_attrs(new, rte_update_pool) : NULL;
+  rte_announce(tab, type, n, new, old, tmpa);
+  rte_update_unlock();
+}
+
 void
 rte_discard(rtable *t, rte *old)	/* Non-filtered route deletion, used during garbage collection */
 {
@@ -760,14 +785,49 @@ rt_dump_all(void)
     rt_dump(t);
 }
 
+static inline void
+rt_schedule_gc(rtable *tab)
+{
+  if (tab->gc_scheduled)
+    return;
+
+  tab->gc_scheduled = 1;
+  ev_schedule(tab->rt_event);
+}
+
+static inline void
+rt_schedule_hcu(rtable *tab)
+{
+  if (tab->hcu_scheduled)
+    return;
+
+  tab->hcu_scheduled = 1;
+  ev_schedule(tab->rt_event);
+}
+
+static inline void
+rt_schedule_nhu(rtable *tab)
+{
+  if (tab->nhu_state == 0)
+    ev_schedule(tab->rt_event);
+
+  /* state change 0->1, 2->3 */
+  tab->nhu_state |= 1;
+}
+
 static void
-rt_gc(void *tab)
+rt_event(void *ptr)
 {
-  rtable *t = tab;
+  rtable *tab = ptr;
+
+  if (tab->hcu_scheduled)
+    rt_update_hostcache(tab);
+
+  if (tab->nhu_state)
+    rt_next_hop_update(tab);
 
-  DBG("Entered routing table garbage collector for %s after %d seconds and %d deletes\n",
-      t->name, (int)(now - t->gc_time), t->gc_counter);
-  rt_prune(t);
+  if (tab->gc_scheduled)
+    rt_prune(tab);
 }
 
 void
@@ -780,9 +840,9 @@ rt_setup(pool *p, rtable *t, char *name, struct rtable_config *cf)
   init_list(&t->hooks);
   if (cf)
     {
-      t->gc_event = ev_new(p);
-      t->gc_event->hook = rt_gc;
-      t->gc_event->data = t;
+      t->rt_event = ev_new(p);
+      t->rt_event->hook = rt_event;
+      t->rt_event->data = t;
       t->gc_time = now;
     }
 }
@@ -811,7 +871,7 @@ rt_init(void)
  * the routing table and removes all routes belonging to inactive
  * protocols and also stale network entries.
  */
-void
+static void
 rt_prune(rtable *tab)
 {
   struct fib_iterator fit;
@@ -852,6 +912,7 @@ again:
 #endif
   tab->gc_counter = 0;
   tab->gc_time = now;
+  tab->gc_scheduled = 0;
 }
 
 /**
@@ -868,6 +929,151 @@ rt_prune_all(void)
     rt_prune(t);
 }
 
+void
+rt_preconfig(struct config *c)
+{
+  struct symbol *s = cf_find_symbol("master");
+
+  init_list(&c->tables);
+  c->master_rtc = rt_new_table(s);
+}
+
+
+/* 
+ * Some functions for handing internal next hop updates
+ * triggered by rt_schedule_nhu().
+ */
+
+static inline int
+hostentry_diff(struct hostentry *he, struct iface *iface, ip_addr gw, byte dest)
+{
+  return (he->iface != iface) || !ipa_equal(he->gw, gw) || (he->dest != dest);
+}
+
+static inline int
+rta_next_hop_outdated(rta *a)
+{
+  struct hostentry *he = a->hostentry;
+  return he && hostentry_diff(he, a->iface, a->gw, a->dest);
+}
+
+static inline void
+rta_apply_hostentry(rta *a, struct hostentry *he)
+{
+  a->hostentry = he;
+  a->iface = he->iface;
+  a->gw = he->gw;
+  a->dest = he->dest;
+}
+
+static inline rte *
+rt_next_hop_update_rte(rtable *tab, rte *old)
+{
+  rta a;
+  memcpy(&a, old->attrs, sizeof(rta));
+  rta_apply_hostentry(&a, old->attrs->hostentry);
+  a.aflags = 0;
+
+  rte *e = sl_alloc(rte_slab);
+  memcpy(e, old, sizeof(rte));
+  e->attrs = rta_lookup(&a);
+
+  return e;
+}
+
+static inline int
+rt_next_hop_update_net(rtable *tab, net *n)
+{
+  rte **k, *e, *new, *old_best, **new_best;
+  int count = 0;
+  int free_old_best = 0;
+
+  old_best = n->routes;
+  if (!old_best)
+    return 0;
+
+  new_best = NULL;
+
+  for (k = &n->routes; e = *k; k = &e->next)
+    {
+      if (rta_next_hop_outdated(e->attrs))
+	{
+	  new = rt_next_hop_update_rte(tab, e);
+	  *k = new;
+
+	  rte_announce_i(tab, RA_ANY, n, new, e);
+	  rte_trace_in(D_ROUTES, new->sender, new, "updated");
+
+	  if (e != old_best)
+	    rte_free_quick(e);
+	  else /* Freeing of the old best rte is postponed */
+	    free_old_best = 1;
+
+	  e = new;
+	  count++;
+	}
+
+      if (!new_best || rte_better(e, *new_best))
+	new_best = k;
+    }
+
+  /* Relink the new best route to the first position */
+  new = *new_best;
+  if (new != n->routes)
+    {
+      *new_best = new->next;
+      new->next = n->routes;
+      n->routes = new;
+    }
+
+  /* Announce the new best route */
+  if (new != old_best)
+    {
+      rte_announce_i(tab, RA_OPTIMAL, n, new, old_best);
+      rte_trace_in(D_ROUTES, new->sender, new, "updated [best]");
+    }
+
+   if (free_old_best)
+    rte_free_quick(old_best);
+
+  return count;
+}
+
+static void
+rt_next_hop_update(rtable *tab)
+{
+  struct fib_iterator *fit = &tab->nhu_fit;
+  int max_feed = 32;
+
+  if (tab->nhu_state == 0)
+    return;
+
+  if (tab->nhu_state == 1)
+    {
+      FIB_ITERATE_INIT(fit, &tab->fib);
+      tab->nhu_state = 2;
+    }
+
+  FIB_ITERATE_START(&tab->fib, fit, fn)
+    {
+      if (max_feed <= 0)
+	{
+	  FIB_ITERATE_PUT(fit, fn);
+	  ev_schedule(tab->rt_event);
+	  return;
+	}
+      max_feed -= rt_next_hop_update_net(tab, (net *) fn);
+    }
+  FIB_ITERATE_END(fn);
+
+  /* state change 2->0, 3->1 */
+  tab->nhu_state &= 1;
+
+  if (tab->nhu_state > 0)
+    ev_schedule(tab->rt_event);
+}
+
+
 struct rtable_config *
 rt_new_table(struct symbol *s)
 {
@@ -881,15 +1087,6 @@ rt_new_table(struct symbol *s)
   return c;
 }
 
-void
-rt_preconfig(struct config *c)
-{
-  struct symbol *s = cf_find_symbol("master");
-
-  init_list(&c->tables);
-  c->master_rtc = rt_new_table(s);
-}
-
 /**
  * rt_lock_table - lock a routing table
  * @r: routing table to be locked
@@ -919,8 +1116,11 @@ rt_unlock_table(rtable *r)
     {
       struct config *conf = r->deleted;
       DBG("Deleting routing table %s\n", r->name);
+      if (r->hostcache)
+	rt_free_hostcache(r);
       rem_node(&r->n);
       fib_free(&r->fib);
+      rfree(r->rt_event);
       mb_free(r);
       config_del_obstacle(conf);
     }
@@ -1087,6 +1287,178 @@ rt_feed_baby_abort(struct proto *p)
     }
 }
 
+static void
+hostentry_init(struct fib_node *fn)
+{
+  ((struct hostentry *) fn)->uc = 0;
+  ((struct hostentry *) fn)->tab = NULL;
+}
+
+static void
+rt_init_hostcache(rtable *tab)
+{
+  struct hostcache *hc = mb_allocz(rt_table_pool, sizeof(struct hostcache));
+  init_list(&hc->hostentries);
+  fib_init(&hc->htable, rt_table_pool, sizeof(struct hostentry), 0, hostentry_init);
+  tab->hostcache = hc;
+}
+
+static void
+rt_free_hostcache(rtable *tab)
+{
+  struct hostcache *hc = tab->hostcache;
+
+  node *n;
+  WALK_LIST(n, hc->hostentries)
+    {
+      struct hostentry *he = SKIP_BACK(struct hostentry, ln, n);
+      if (he->uc)
+	log(L_ERR "Hostcache is not empty in table %s", tab->name);
+    }
+
+  fib_free(&hc->htable);
+  mb_free(hc);
+}
+
+static void
+rt_notify_hostcache(rtable *tab, net *net)
+{
+  struct hostcache *hc = tab->hostcache;
+
+  if (tab->hcu_scheduled)
+    return;
+
+  node *n;
+  WALK_LIST(n, hc->hostentries)
+    {
+      struct hostentry *he = SKIP_BACK(struct hostentry, ln, n);
+      if (ipa_in_net(he->fn.prefix, net->n.prefix, net->n.pxlen) &&
+	  (he->pxlen <= net->n.pxlen))
+	{
+	  rt_schedule_hcu(tab);
+	  return;
+	}
+    }
+}
+
+static int
+if_local_addr(ip_addr a, struct iface *i)
+{
+  struct ifa *b;
+
+  WALK_LIST(b, i->addrs)
+    if (ipa_equal(a, b->ip))
+      return 1;
+
+  return 0;
+}
+
+static int
+rt_update_hostentry(rtable *tab, struct hostentry *he)
+{
+  struct iface *old_iface = he->iface;
+  ip_addr old_gw = he->gw;
+  byte old_dest = he->dest;
+
+  net *n = fib_route(&tab->fib, he->fn.prefix, MAX_PREFIX_LENGTH);
+  if (n && n->routes)
+    {
+      rta *a = n->routes->attrs;
+
+      if (a->dest == RTD_DEVICE)
+	{
+	  if (if_local_addr(he->fn.prefix, a->iface))
+	    {
+	      /* The host address is a local address, this is not valid */
+	      log(L_WARN "Next hop address %I is a local address of iface %s",
+		  he->fn.prefix, a->iface->name);
+	      he->iface = NULL;
+	      he->gw = IPA_NONE;
+	      he->dest = RTD_UNREACHABLE;
+      	    }
+	  else
+	    {
+	      /* The host is directly reachable, us it as a gateway */
+	      he->iface = a->iface;
+	      he->gw = he->fn.prefix;
+	      he->dest = RTD_ROUTER;
+	    }
+	}
+      else
+	{
+	  /* The host is reachable through some route entry */
+	  he->iface = a->iface;
+	  he->gw = a->gw;
+	  he->dest = a->dest;
+	}
+
+      he->pxlen = n->n.pxlen;
+    }
+  else
+    {
+      /* The host is unreachable */
+      he->iface = NULL;
+      he->gw = IPA_NONE;
+      he->dest = RTD_UNREACHABLE;
+
+      he->pxlen = 0;
+    }
+
+  return hostentry_diff(he, old_iface, old_gw, old_dest);
+}
+
+static void
+rt_update_hostcache(rtable *tab)
+{
+  struct hostcache *hc = tab->hostcache;
+  struct hostentry *he;
+  node *n, *x;
+
+  WALK_LIST_DELSAFE(n, x, hc->hostentries)
+    {
+      he = SKIP_BACK(struct hostentry, ln, n);
+      if (!he->uc)
+	{
+	  /* Delete a hostentry */
+	  rem_node(&he->ln);
+	  fib_delete(&hc->htable, he);
+	  continue;
+	}
+
+      if (rt_update_hostentry(tab, he))
+	rt_schedule_nhu(he->tab);
+    }
+
+  tab->hcu_scheduled = 0;
+}
+
+static struct hostentry *
+rt_find_hostentry(rtable *tab, ip_addr *a, rtable *dep)
+{
+  struct hostentry *he;
+
+  if (!tab->hostcache)
+    rt_init_hostcache(tab);
+
+  he = fib_get(&tab->hostcache->htable, a, MAX_PREFIX_LENGTH);
+  if (!he->tab)
+    {
+      /* New entry */
+      add_tail(&tab->hostcache->hostentries, &he->ln);
+      he->tab = dep;
+
+      rt_update_hostentry(tab, he);
+    }
+
+  return he;
+}
+
+void
+rta_set_recursive_next_hop(rtable *dep, rta *a, rtable *tab, ip_addr *gw)
+{
+  rta_apply_hostentry(a, rt_find_hostentry(tab, gw, dep));
+}
+
 /*
  *  CLI commands
  */
diff --git a/proto/bgp/attrs.c b/proto/bgp/attrs.c
index c1b8fd1e..8743358e 100644
--- a/proto/bgp/attrs.c
+++ b/proto/bgp/attrs.c
@@ -1265,15 +1265,13 @@ bgp_decode_attrs(struct bgp_conn *conn, byte *attr, unsigned int len, struct lin
   ea_list *ea;
   struct adata *ad;
 
+  bzero(a, sizeof(rta));
   a->proto = &bgp->p;
   a->source = RTS_BGP;
   a->scope = SCOPE_UNIVERSE;
   a->cast = RTC_UNICAST;
-  a->dest = RTD_ROUTER;
-  a->flags = 0;
-  a->aflags = 0;
+  /* a->dest = RTD_ROUTER;  -- set in bgp_set_next_hop() */
   a->from = bgp->cf->remote_ip;
-  a->eattrs = NULL;
 
   /* Parse the attributes */
   bzero(seen, sizeof(seen));
diff --git a/proto/bgp/bgp.c b/proto/bgp/bgp.c
index 932c8aa3..b36c4a3d 100644
--- a/proto/bgp/bgp.c
+++ b/proto/bgp/bgp.c
@@ -786,6 +786,8 @@ bgp_start(struct proto *P)
   p->incoming_conn.state = BS_IDLE;
   p->neigh = NULL;
 
+  rt_lock_table(p->igp_table);
+
   p->event = ev_new(p->p.pool);
   p->event->hook = bgp_decision;
   p->event->data = p;
@@ -837,6 +839,19 @@ bgp_shutdown(struct proto *P)
   return p->p.proto_state;
 }
 
+static void
+bgp_cleanup(struct proto *P)
+{
+  struct bgp_proto *p = (struct bgp_proto *) P;
+  rt_unlock_table(p->igp_table);
+}
+
+static rtable *
+get_igp_table(struct bgp_config *cf)
+{
+  return cf->igp_table ? cf->igp_table->table : cf->c.table->table;
+}
+
 static struct proto *
 bgp_init(struct proto_config *C)
 {
@@ -854,6 +869,7 @@ bgp_init(struct proto_config *C)
   p->local_as = c->local_as;
   p->remote_as = c->remote_as;
   p->is_internal = (c->local_as == c->remote_as);
+  p->igp_table = get_igp_table(c);
   return P;
 }
 
@@ -1065,7 +1081,8 @@ bgp_reconfigure(struct proto *P, struct proto_config *C)
 		     // password item is last and must be checked separately
 		     OFFSETOF(struct bgp_config, password) - sizeof(struct proto_config))
     && ((!old->password && !new->password)
-	|| (old->password && new->password && !strcmp(old->password, new->password)));
+	|| (old->password && new->password && !strcmp(old->password, new->password)))
+    && (get_igp_table(old) == get_igp_table(new));
 
   /* We should update our copy of configuration ptr as old configuration will be freed */
   if (same)
@@ -1081,6 +1098,7 @@ struct protocol proto_bgp = {
   init:			bgp_init,
   start:		bgp_start,
   shutdown:		bgp_shutdown,
+  cleanup:		bgp_cleanup,
   reconfigure:		bgp_reconfigure,
   get_status:		bgp_get_status,
   get_attr:		bgp_get_attr,
diff --git a/proto/bgp/bgp.h b/proto/bgp/bgp.h
index c81fe624..76844af3 100644
--- a/proto/bgp/bgp.h
+++ b/proto/bgp/bgp.h
@@ -47,6 +47,7 @@ struct bgp_config {
   unsigned error_delay_time_max;
   unsigned disable_after_error;		/* Disable the protocol when error is detected */
   char *password;			/* Password used for MD5 authentication */
+  struct rtable_config *igp_table;	/* Table used for recursive next hop lookups */
 };
 
 #define MLL_SELF 1
@@ -92,6 +93,7 @@ struct bgp_proto {
   struct neighbor *neigh;		/* Neighbor entry corresponding to next_hop */
   ip_addr local_addr;			/* Address of the local end of the link to next_hop */
   ip_addr source_addr;			/* Address used as advertised next hop, usually local_addr */
+  rtable *igp_table;			/* Table used for recursive next hop lookups */
   struct event *event;			/* Event for respawning and shutting process */
   struct timer *startup_timer;		/* Timer used to delay protocol startup due to previous errors (startup_delay) */
   struct bgp_bucket **bucket_hash;	/* Hash table of attribute buckets */
diff --git a/proto/bgp/config.Y b/proto/bgp/config.Y
index c4ed1032..a46431cb 100644
--- a/proto/bgp/config.Y
+++ b/proto/bgp/config.Y
@@ -24,7 +24,7 @@ CF_KEYWORDS(BGP, LOCAL, NEIGHBOR, AS, HOLD, TIME, CONNECT, RETRY, KEEPALIVE,
 	PASSWORD, RR, RS, CLIENT, CLUSTER, ID, AS4, ADVERTISE, IPV4,
 	CAPABILITIES, LIMIT, PASSIVE, PREFER, OLDER, MISSING, LLADDR,
 	DROP, IGNORE, ROUTE, REFRESH, INTERPRET, COMMUNITIES,
-	BGP_ORIGINATOR_ID, BGP_CLUSTER_LIST)
+	BGP_ORIGINATOR_ID, BGP_CLUSTER_LIST, IGP, TABLE)
 
 CF_GRAMMAR
 
@@ -89,6 +89,7 @@ bgp_proto:
  | bgp_proto ROUTE LIMIT expr ';' { BGP_CFG->route_limit = $4; }
  | bgp_proto PASSIVE bool ';' { BGP_CFG->passive = $3; }
  | bgp_proto INTERPRET COMMUNITIES bool ';' { BGP_CFG->interpret_communities = $4; }
+ | bgp_proto IGP TABLE rtable ';' { BGP_CFG->igp_table = $4; }
  ;
 
 CF_ADDTO(dynamic_attr, BGP_ORIGIN
diff --git a/proto/bgp/packets.c b/proto/bgp/packets.c
index 1e9d6465..514e878d 100644
--- a/proto/bgp/packets.c
+++ b/proto/bgp/packets.c
@@ -802,26 +802,26 @@ bgp_rx_open(struct bgp_conn *conn, byte *pkt, int len)
 } while (0)
 
 static inline int
-bgp_get_nexthop(struct bgp_proto *bgp, rta *a)
+bgp_set_next_hop(struct bgp_proto *p, rta *a)
 {
-  neighbor *neigh;
-  ip_addr nexthop;
   struct eattr *nh = ea_find(a->eattrs, EA_CODE(EAP_BGP, BA_NEXT_HOP));
-  ASSERT(nh);
-  nexthop = *(ip_addr *) nh->u.ptr->data;
-  neigh = neigh_find(&bgp->p, &nexthop, 0);
-  if (neigh)
+  ip_addr nexthop = *(ip_addr *) nh->u.ptr->data;
+
+  if (!p->is_internal) /* FIXME better option
+ */
     {
-      if (neigh->scope == SCOPE_HOST)
-	{
-	  DBG("BGP: Loop!\n");
-	  return 0;
-	}
+      neighbor *ng = neigh_find(&p->p, &nexthop, 0) ? : p->neigh;
+      if (ng->scope == SCOPE_HOST)
+	return 0;
+
+      a->dest = RTD_ROUTER;
+      a->gw = ng->addr;
+      a->iface = ng->iface;
+      a->hostentry = NULL;
     }
   else
-    neigh = bgp->neigh;
-  a->gw = neigh->addr;
-  a->iface = neigh->iface;
+    rta_set_recursive_next_hop(p->p.table, a, p->igp_table, &nexthop);
+
   return 1;
 }
 
@@ -853,7 +853,7 @@ bgp_do_rx_update(struct bgp_conn *conn,
     return;
 
   a0 = bgp_decode_attrs(conn, attrs, attr_len, bgp_linpool, nlri_len);
-  if (a0 && nlri_len && bgp_get_nexthop(p, a0))
+  if (a0 && nlri_len && bgp_set_next_hop(p, a0))
     {
       a = rta_lookup(a0);
       while (nlri_len)
diff --git a/proto/pipe/pipe.c b/proto/pipe/pipe.c
index 7fdf2733..e557097d 100644
--- a/proto/pipe/pipe.c
+++ b/proto/pipe/pipe.c
@@ -63,6 +63,7 @@ pipe_rt_notify(struct proto *P, rtable *src_table, net *n, rte *new, rte *old, e
 
       a.aflags = 0;
       a.eattrs = attrs;
+      a.hostentry = NULL;
       e = rte_get_temp(&a);
       e->net = nn;
       e->pflags = 0;
@@ -120,7 +121,7 @@ pipe_start(struct proto *P)
   /* Clean up the secondary stats */
   bzero(&p->peer_stats, sizeof(struct proto_stats));
 
-  /* Lock the peer table, unlock is handled in proto_fell_down() */
+  /* Lock the peer table, unlock is handled in pipe_cleanup() */
   rt_lock_table(p->peer);
 
   /* Connect the protocol also to the peer routing table. */
@@ -129,6 +130,13 @@ pipe_start(struct proto *P)
   return PS_UP;
 }
 
+static void
+pipe_cleanup(struct proto *P)
+{
+  struct pipe_proto *p = (struct pipe_proto *) P;
+  rt_unlock_table(p->peer);
+}
+
 static struct proto *
 pipe_init(struct proto_config *C)
 {
@@ -185,6 +193,7 @@ struct protocol proto_pipe = {
   postconfig:	pipe_postconfig,
   init:		pipe_init,
   start:	pipe_start,
+  cleanup:	pipe_cleanup,
   reconfigure:	pipe_reconfigure,
   get_status:	pipe_get_status,
 };
diff --git a/sysdep/bsd/krt-sock.c b/sysdep/bsd/krt-sock.c
index 4c8676ef..9ae658d6 100644
--- a/sysdep/bsd/krt-sock.c
+++ b/sysdep/bsd/krt-sock.c
@@ -244,7 +244,6 @@ krt_set_start(struct krt_proto *x, int first UNUSED)
 static void
 krt_read_rt(struct ks_msg *msg, struct krt_proto *p, int scan)
 {
-  rta a;
   rte *e;
   net *net;
   sockaddr dst, gate, mask;
@@ -329,17 +328,12 @@ krt_read_rt(struct ks_msg *msg, struct krt_proto *p, int scan)
 
   net = net_get(p->p.table, idst, pxlen);
 
-  bzero(&a, sizeof(a));
-
-  a.proto = &p->p;
-  a.source = RTS_INHERIT;
-  a.scope = SCOPE_UNIVERSE;
-  a.cast = RTC_UNICAST;
-  a.flags = a.aflags = 0;
-  a.from = IPA_NONE;
-  a.gw = IPA_NONE;
-  a.iface = NULL;
-  a.eattrs = NULL;
+  rta a = {
+    .proto = &p->p,
+    .source = RTS_INHERIT,
+    .scope = SCOPE_UNIVERSE,
+    .cast = RTC_UNICAST
+  };
 
   /* reject/blackhole routes have also set RTF_GATEWAY,
      we wil check them first. */
diff --git a/sysdep/linux/krt-scan.c b/sysdep/linux/krt-scan.c
index feb128ef..8591607e 100644
--- a/sysdep/linux/krt-scan.c
+++ b/sysdep/linux/krt-scan.c
@@ -48,7 +48,6 @@ krt_parse_entry(byte *ent, struct krt_proto *p)
   int masklen;
   net *net;
   byte *iface = ent;
-  rta a;
   rte *e;
 
   if (sscanf(ent, "%*s\t%x\t%x\t%x\t%*d\t%*d\t%*d\t%x\t", &dest0, &gw0, &flags, &mask0) != 4)
@@ -88,14 +87,12 @@ krt_parse_entry(byte *ent, struct krt_proto *p)
 
   net = net_get(p->p.table, dest, masklen);
 
-  a.proto = &p->p;
-  a.source = RTS_INHERIT;
-  a.scope = SCOPE_UNIVERSE;
-  a.cast = RTC_UNICAST;
-  a.flags = a.aflags = 0;
-  a.from = IPA_NONE;
-  a.iface = NULL;
-  a.eattrs = NULL;
+  rta a = {
+    .proto = &p->p,
+    .source = RTS_INHERIT,
+    .scope = SCOPE_UNIVERSE,
+    .cast = RTC_UNICAST
+  };
 
   if (flags & RTF_GATEWAY)
     {
diff --git a/sysdep/linux/netlink/netlink.c b/sysdep/linux/netlink/netlink.c
index a10a2e9f..2dd0359b 100644
--- a/sysdep/linux/netlink/netlink.c
+++ b/sysdep/linux/netlink/netlink.c
@@ -570,7 +570,6 @@ nl_parse_route(struct nlmsghdr *h, int scan)
   struct rtattr *a[RTA_CACHEINFO+1];
   int new = h->nlmsg_type == RTM_NEWROUTE;
   ip_addr dst;
-  rta ra;
   rte *e;
   net *net;
   u32 oif;
@@ -655,15 +654,13 @@ nl_parse_route(struct nlmsghdr *h, int scan)
     }
 
   net = net_get(p->p.table, dst, i->rtm_dst_len);
-  ra.proto = &p->p;
-  ra.source = RTS_INHERIT;
-  ra.scope = SCOPE_UNIVERSE;
-  ra.cast = RTC_UNICAST;
-  ra.flags = ra.aflags = 0;
-  ra.from = IPA_NONE;
-  ra.gw = IPA_NONE;
-  ra.iface = NULL;
-  ra.eattrs = NULL;
+
+  rta ra = {
+    .proto = &p->p,
+    .source = RTS_INHERIT,
+    .scope = SCOPE_UNIVERSE,
+    .cast = RTC_UNICAST
+  };
 
   switch (i->rtm_type)
     {