Implements deterministic MED handling.

Thanks to Alexander V. Chernikov for many suggestions.

doc/bird.sgml

@@ -1316,6 +1316,21 @@ for each neighbor using the following configuration parameters:
 	received from the same AS (which is the standard behavior).
 	Default: off.
 
+	<tag>deterministic med <m/switch/</tag> BGP route selection
+	algorithm is often viewed as a comparison between individual
+	routes (e.g. if a new route appears and is better than the
+	current best one, it is chosen as the new best one). But the
+	proper route selection, as specified by RFC 4271, cannot be
+	fully implemented in that way. The problem is mainly in
+	handling the MED attribute. BIRD, by default, uses an
+	simplification based on individual route comparison, which in
+	some cases may lead to temporally dependent behavior (i.e. the
+	selection is dependent on the order in which routes appeared).
+	This option enables a different (and slower) algorithm
+	implementing proper RFC 4271 route selection, which is
+	deterministic. Alternative way how to get deterministic
+	behavior is to use <cf/med metric/ option. Default: off.
+
 	<tag>igp metric <m/switch/</tag> Enable comparison of internal
  	distances to boundary routers during best route selection. Default: on.
 

nest/proto-hooks.c

@@ -267,6 +267,26 @@ void store_tmp_attrs(rte *e, ea_list *attrs)
 int import_control(struct proto *p, rte **e, ea_list **attrs, struct linpool *pool)
 { DUMMY; }
 
+/**
+ * rte_recalculate - prepare routes for comparison
+ * @table: a routing table 
+ * @net: a network entry
+ * @new: new route for the network
+ * @old: old route for the network
+ * @old_best: old best route for the network (may be NULL)
+ *
+ * This hook is called when a route change (from @old to @new for a
+ * @net entry) is propagated to a @table. It may be used to prepare
+ * routes for comparison by rte_better() in the best route
+ * selection. @new may or may not be in @net->routes list,
+ * @old is not there.
+ *
+ * Result: 1 if the ordering implied by rte_better() changes enough
+ * that full best route calculation have to be done, 0 otherwise.
+ */
+int rte_recalculate(struct rtable *table, struct network *net, struct rte *new, struct rte *old, struct rte *old_best)
+{ DUMMY; }
+
 /**
  * rte_better - compare metrics of two routes
  * @new: the new route

nest/protocol.h

@@ -178,12 +178,14 @@ struct proto {
   /*
    *	Routing entry hooks (called only for rte's belonging to this protocol):
    *
+   *	   rte_recalculate Called at the beginning of the best route selection  
    *	   rte_better	Compare two rte's and decide which one is better (1=first, 0=second).
    *       rte_same	Compare two rte's and decide whether they are identical (1=yes, 0=no).
    *	   rte_insert	Called whenever a rte is inserted to a routing table.
    *	   rte_remove	Called whenever a rte is removed from the routing table.
    */
 
+  int (*rte_recalculate)(struct rtable *, struct network *, struct rte *, struct rte *, struct rte *);
   int (*rte_better)(struct rte *, struct rte *);
   int (*rte_same)(struct rte *, struct rte *);
   void (*rte_insert)(struct network *, struct rte *);

nest/route.h

@@ -200,6 +200,11 @@ typedef struct rte {
       u32 tag;				/* External route tag */
       u32 router_id;			/* Router that originated this route */
     } ospf;
+#endif
+#ifdef CONFIG_BGP
+    struct {
+      u8 suppressed;			/* Used for deterministic MED comparison */
+    } bgp;
 #endif
     struct {				/* Routes generated by krt sync (both temporary and inherited ones) */
       s8 src;				/* Alleged route source (see krt.h) */

nest/rt-table.c

@@ -498,6 +498,9 @@ rte_recalculate(rtable *table, net *net, struct proto *p, struct proto *src, rte
 
   rte_announce(table, RA_ANY, net, new, old, tmpa);
 
+  if (src->rte_recalculate && src->rte_recalculate(table, net, new, old, old_best))
+    goto do_recalculate;
+
   if (new && rte_better(new, old_best))
     {
       /* The first case - the new route is cleary optimal, we link it
@@ -516,6 +519,7 @@ rte_recalculate(rtable *table, net *net, struct proto *p, struct proto *src, rte
 	 that route at the first position and announce it. New optimal
 	 route might be NULL if there is no more routes */
 
+    do_recalculate:
       /* Add the new route to the list */
       if (new)
 	{
@@ -1015,10 +1019,7 @@ rt_next_hop_update_net(rtable *tab, net *n)
   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);
@@ -1027,6 +1028,11 @@ rt_next_hop_update_net(rtable *tab, net *n)
 	rte_announce_i(tab, RA_ANY, n, new, e);
 	rte_trace_in(D_ROUTES, new->sender, new, "updated");
 
+	/* Call a pre-comparison hook */
+	/* Not really an efficient way to compute this */
+	if (e->attrs->proto->rte_recalculate)
+	  e->attrs->proto->rte_recalculate(tab, n, new, e, NULL);
+
 	if (e != old_best)
 	  rte_free_quick(e);
 	else /* Freeing of the old best rte is postponed */
@@ -1036,6 +1042,13 @@ rt_next_hop_update_net(rtable *tab, net *n)
 	count++;
       }
 
+  if (!count)
+    return 0;
+
+  /* Find the new best route */
+  new_best = NULL;
+  for (k = &n->routes; e = *k; k = &e->next)
+    {
       if (!new_best || rte_better(e, *new_best))
 	new_best = k;
     }

proto/bgp/attrs.c

@@ -1125,6 +1125,14 @@ bgp_rte_better(rte *new, rte *old)
   eattr *x, *y;
   u32 n, o;
 
+  /* Skip suppressed routes (see bgp_rte_recalculate()) */
+  n = new->u.bgp.suppressed;
+  o = old->u.bgp.suppressed;
+  if (n > o)
+    return 0;
+  if (n < o)
+    return 1;
+
   /* RFC 4271 9.1.2.1. Route resolvability test */
   n = rte_resolvable(new);
   o = rte_resolvable(old);
@@ -1167,14 +1175,15 @@ bgp_rte_better(rte *new, rte *old)
     return 0;
 
   /* RFC 4271 9.1.2.2. c) Compare MED's */
-  /* This is noncompliant. Proper RFC 4271 path selection cannot be
+  /* Proper RFC 4271 path selection cannot be interpreted as finding
-   * interpreted as finding the best path in some ordering.
+   * the best path in some ordering. It is implemented partially in
-   * Therefore, it cannot be implemented in BIRD without some ugly
+   * bgp_rte_recalculate() when deterministic_med option is
-   * hacks. This is just an approximation, which in specific
+   * active. Without that option, the behavior is just an
-   * situations may lead to persistent routing loops, because it is
+   * approximation, which in specific situations may lead to
-   * nondeterministic - it depends on the order in which routes
+   * persistent routing loops, because it is nondeterministic - it
-   * appeared. But it is also the same behavior as used by default in
+   * depends on the order in which routes appeared. But it is also the
-   * Cisco routers, so it is probably not a big issue.
+   * same behavior as used by default in Cisco routers, so it is
+   * probably not a big issue.
    */
   if (new_bgp->cf->med_metric || old_bgp->cf->med_metric ||
       (bgp_get_neighbor(new) == bgp_get_neighbor(old)))
@@ -1236,6 +1245,148 @@ bgp_rte_better(rte *new, rte *old)
   return (ipa_compare(new_bgp->cf->remote_ip, old_bgp->cf->remote_ip) < 0);
 }
 
+
+static inline int
+same_group(rte *r, u32 lpref, u32 lasn)
+{
+  return (r->pref == lpref) && (bgp_get_neighbor(r) == lasn);
+}
+
+static inline int
+use_deterministic_med(rte *r)
+{
+  return ((struct bgp_proto *) r->attrs->proto)->cf->deterministic_med;
+}
+
+int
+bgp_rte_recalculate(rtable *table, net *net, rte *new, rte *old, rte *old_best)
+{
+  rte *r, *s;
+  rte *key = new ? new : old;
+  u32 lpref = key->pref;
+  u32 lasn = bgp_get_neighbor(key);
+  int old_is_group_best = 0;
+
+  /*
+   * Proper RFC 4271 path selection is a bit complicated, it cannot be
+   * implemented just by rte_better(), because it is not a linear
+   * ordering. But it can be splitted to two levels, where the lower
+   * level chooses the best routes in each group of routes from the
+   * same neighboring AS and higher level chooses the best route (with
+   * a slightly different ordering) between the best-in-group routes.
+   *
+   * When deterministic_med is disabled, we just ignore this issue and
+   * choose the best route by bgp_rte_better() alone. If enabled, the
+   * lower level of the route selection is done here (for the group
+   * to which the changed route belongs), all routes in group are
+   * marked as suppressed, just chosen best-in-group is not.
+   *
+   * Global best route selection then implements higher level by
+   * choosing between non-suppressed routes (as they are always
+   * preferred over suppressed routes). Routes from BGP protocols
+   * that do not set deterministic_med are just never suppressed. As
+   * they do not participate in the lower level selection, it is OK
+   * that this fn is not called for them.
+   *
+   * The idea is simple, the implementation is more problematic,
+   * mostly because of optimizations in rte_recalculate() that 
+   * avoids full recalculation in most cases.
+   *
+   * We can assume that at least one of new, old is non-NULL and both
+   * are from the same protocol with enabled deterministic_med. We
+   * group routes by both neighbor AS (lasn) and preference (lpref),
+   * because bgp_rte_better() does not handle preference itself.
+   */
+
+  /* If new and old are from different groups, we just process that
+     as two independent events */
+  if (new && old && !same_group(old, lpref, lasn))
+    {
+      int i1, i2;
+      i1 = bgp_rte_recalculate(table, net, NULL, old, old_best);
+      i2 = bgp_rte_recalculate(table, net, new, NULL, old_best);
+      return i1 || i2;
+    }
+
+  /* 
+   * We could find the best-in-group and then make some shortcuts like
+   * in rte_recalculate, but as we would have to walk through all
+   * net->routes just to find it, it is probably not worth. So we
+   * just have two simpler fast cases that use just the old route.
+   * We also set suppressed flag to avoid using it in bgp_rte_better().
+   */
+
+  if (new)
+    new->u.bgp.suppressed = 1;
+
+  if (old)
+    {
+      old_is_group_best = !old->u.bgp.suppressed;
+      old->u.bgp.suppressed = 1;
+      int new_is_better = new && bgp_rte_better(new, old);
+
+      /* The first case - replace not best with worse (or remove not best) */
+      if (!old_is_group_best && !new_is_better)
+	return 0;
+
+      /* The second case - replace the best with better */
+      if (old_is_group_best && new_is_better)
+	{
+	  /* new is best-in-group, the see discussion below - this is
+	     a special variant of NBG && OBG. From OBG we can deduce
+	     that same_group(old_best) iff (old == old_best)  */
+	  new->u.bgp.suppressed = 0;
+	  return (old == old_best);
+	}
+    }
+
+  /* The default case - find a new best-in-group route */
+  r = new; /* new may not be in the list */
+  for (s=net->routes; s; s=s->next)
+    if (use_deterministic_med(s) && same_group(s, lpref, lasn))
+      {
+	s->u.bgp.suppressed = 1;
+	if (!r || bgp_rte_better(s, r))
+	  r = s;
+      }
+
+  /* Simple case - the last route in group disappears */
+  if (!r)
+    return 0;
+
+  /* Found best-in-group */
+  r->u.bgp.suppressed = 0;
+
+  /*
+   * There are generally two reasons why we have to force
+   * recalculation (return 1): First, the new route may be wrongfully
+   * chosen to be the best in the first case check in
+   * rte_recalculate(), this may happen only if old_best is from the
+   * same group. Second, another (different than new route)
+   * best-in-group is chosen and that may be the proper best (although
+   * rte_recalculate() without ignore that possibility).
+   *
+   * There are three possible cases according to whether the old route
+   * was the best in group (OBG, stored in old_is_group_best) and
+   * whether the new route is the best in group (NBG, tested by r == new).
+   * These cases work even if old or new is NULL.
+   *
+   * NBG -> new is a possible candidate for the best route, so we just
+   *        check for the first reason using same_group().
+   *
+   * !NBG && OBG -> Second reason applies, return 1
+   *
+   * !NBG && !OBG -> Best in group does not change, old != old_best,
+   *                 rte_better(new, old_best) is false and therefore
+   *                 the first reason does not apply, return 0
+   */
+
+  if (r == new)
+    return old_best && same_group(old_best, lpref, lasn);
+  else
+    return old_is_group_best;
+}
+
 static struct adata *
 bgp_aggregator_convert_to_new(struct adata *old, struct linpool *pool)
 {
@@ -1614,6 +1765,11 @@ bgp_get_route_info(rte *e, byte *buf, ea_list *attrs)
   eattr *o = ea_find(attrs, EA_CODE(EAP_BGP, BA_ORIGIN));
   u32 origas;
 
+  /*
+  if (e->u.bgp.suppressed)
+    buf += bsprintf(buf, " -");
+  */
+
   buf += bsprintf(buf, " (%d", e->pref);
   if (e->attrs->hostentry)
     {

proto/bgp/bgp.c

@@ -908,6 +908,10 @@ bgp_init(struct proto_config *C)
   P->import_control = bgp_import_control;
   P->neigh_notify = bgp_neigh_notify;
   P->reload_routes = bgp_reload_routes;
+
+  if (c->deterministic_med)
+    P->rte_recalculate = bgp_rte_recalculate;
+
   p->cf = c;
   p->local_as = c->local_as;
   p->remote_as = c->remote_as;

proto/bgp/bgp.h

@@ -29,6 +29,7 @@ struct bgp_config {
   int med_metric;			/* Compare MULTI_EXIT_DISC even between routes from differen ASes */
   int igp_metric;			/* Use IGP metrics when selecting best route */
   int prefer_older;			/* Prefer older routes according to RFC 5004 */
+  int deterministic_med;		/* Use more complicated algo to have strict RFC 4271 MED comparison */
   u32 default_local_pref;		/* Default value for LOCAL_PREF attribute */
   u32 default_med;			/* Default value for MULTI_EXIT_DISC attribute */
   int capabilities;			/* Enable capability handshake [RFC3392] */
@@ -185,6 +186,7 @@ byte *bgp_attach_attr_wa(struct ea_list **to, struct linpool *pool, unsigned att
 struct rta *bgp_decode_attrs(struct bgp_conn *conn, byte *a, unsigned int len, struct linpool *pool, int mandatory);
 int bgp_get_attr(struct eattr *e, byte *buf, int buflen);
 int bgp_rte_better(struct rte *, struct rte *);
+int bgp_rte_recalculate(rtable *table, net *net, rte *new, rte *old, rte *old_best);
 void bgp_rt_notify(struct proto *P, rtable *tbl UNUSED, net *n, rte *new, rte *old UNUSED, ea_list *attrs);
 int bgp_import_control(struct proto *, struct rte **, struct ea_list **, struct linpool *);
 void bgp_attr_init(struct bgp_proto *);

proto/bgp/config.Y

@@ -25,7 +25,7 @@ CF_KEYWORDS(BGP, LOCAL, NEIGHBOR, AS, HOLD, TIME, CONNECT, RETRY,
 	CLUSTER, ID, AS4, ADVERTISE, IPV4, CAPABILITIES, LIMIT, PASSIVE,
 	PREFER, OLDER, MISSING, LLADDR, DROP, IGNORE, ROUTE, REFRESH,
 	INTERPRET, COMMUNITIES, BGP_ORIGINATOR_ID, BGP_CLUSTER_LIST, IGP,
-	TABLE, GATEWAY, DIRECT, RECURSIVE, MED, TTL, SECURITY)
+	TABLE, GATEWAY, DIRECT, RECURSIVE, MED, TTL, SECURITY, DETERMINISTIC)
 
 CF_GRAMMAR
 
@@ -82,6 +82,7 @@ bgp_proto:
  | bgp_proto MED METRIC bool ';' { BGP_CFG->med_metric = $4; }
  | bgp_proto IGP METRIC bool ';' { BGP_CFG->igp_metric = $4; }
  | bgp_proto PREFER OLDER bool ';' { BGP_CFG->prefer_older = $4; }
+ | bgp_proto DETERMINISTIC MED bool ';' { BGP_CFG->deterministic_med = $4; }
  | bgp_proto DEFAULT BGP_MED expr ';' { BGP_CFG->default_med = $4; }
  | bgp_proto DEFAULT BGP_LOCAL_PREF expr ';' { BGP_CFG->default_local_pref = $4; }
  | bgp_proto SOURCE ADDRESS ipa ';' { BGP_CFG->source_addr = $4; }