Extends multipath support for OSPF.

Fixes cases where the same network or external route are propagated by
several OSPF routes and some other corner cases in next hop construction
and ECMP. Allows to specify whether external routes should be merged.

Thanks to Peter Christensen for the original patch.

5 files changed, 372 insertions, 199 deletions

diff --git a/proto/ospf/config.Y b/proto/ospf/config.Y
index 90f289d0..478529bc 100644
--- a/proto/ospf/config.Y
+++ b/proto/ospf/config.Y
@@ -132,7 +132,7 @@ CF_KEYWORDS(ELIGIBLE, POLL, NETWORKS, HIDDEN, VIRTUAL, CHECK, LINK, ONLY, BFD)
 CF_KEYWORDS(RX, BUFFER, LARGE, NORMAL, STUBNET, HIDDEN, SUMMARY, TAG, EXTERNAL)
 CF_KEYWORDS(WAIT, DELAY, LSADB, ECMP, LIMIT, WEIGHT, NSSA, TRANSLATOR, STABILITY)
 CF_KEYWORDS(GLOBAL, LSID, ROUTER, SELF, INSTANCE, REAL, NETMASK, TX, PRIORITY, LENGTH)
-CF_KEYWORDS(SECONDARY)
+CF_KEYWORDS(SECONDARY, MERGE)
 
 %type <t> opttext
 %type <ld> lsadb_args
@@ -162,6 +162,7 @@ ospf_proto_item:
  | STUB ROUTER bool { OSPF_CFG->stub_router = $3; }
  | ECMP bool { OSPF_CFG->ecmp = $2 ? DEFAULT_ECMP_LIMIT : 0; }
  | ECMP bool LIMIT expr { OSPF_CFG->ecmp = $2 ? $4 : 0; if ($4 < 0) cf_error("ECMP limit cannot be negative"); }
+ | MERGE EXTERNAL bool { OSPF_CFG->merge_external = $3; }
  | TICK expr { OSPF_CFG->tick = $2; if($2<=0) cf_error("Tick must be greater than zero"); }
  | ospf_area
  ;
diff --git a/proto/ospf/ospf.c b/proto/ospf/ospf.c
index e751f7ca..6756ff49 100644
--- a/proto/ospf/ospf.c
+++ b/proto/ospf/ospf.c
@@ -234,6 +234,7 @@ ospf_start(struct proto *p)
   po->router_id = proto_get_router_id(p->cf);
   po->rfc1583 = c->rfc1583;
   po->stub_router = c->stub_router;
+  po->merge_external = c->merge_external;
   po->ebit = 0;
   po->ecmp = c->ecmp;
   po->tick = c->tick;
@@ -742,6 +743,7 @@ ospf_reconfigure(struct proto *p, struct proto_config *c)
     return 0;
 
   po->stub_router = new->stub_router;
+  po->merge_external = new->merge_external;
   po->ecmp = new->ecmp;
   po->tick = new->tick;
   po->disp_timer->recurrent = po->tick;
diff --git a/proto/ospf/ospf.h b/proto/ospf/ospf.h
index 66719e30..e705b88b 100644
--- a/proto/ospf/ospf.h
+++ b/proto/ospf/ospf.h
@@ -81,6 +81,7 @@ struct ospf_config
   unsigned tick;
   byte rfc1583;
   byte stub_router;
+  byte merge_external;
   byte abr;
   int ecmp;
   list area_list;		/* list of struct ospf_area_config */
@@ -777,6 +778,7 @@ struct proto_ospf
   struct fib rtf;		/* Routing table */
   byte rfc1583;			/* RFC1583 compatibility */
   byte stub_router;		/* Do not forward transit traffic */
+  byte merge_external;		/* Should i merge external routes? */
   byte ebit;			/* Did I originate any ext lsa? */
   byte ecmp;			/* Maximal number of nexthops in ECMP route, or 0 */
   struct ospf_area *backbone;	/* If exists */
diff --git a/proto/ospf/rt.c b/proto/ospf/rt.c
index 1b39bda0..2a879c05 100644
--- a/proto/ospf/rt.c
+++ b/proto/ospf/rt.c
@@ -37,9 +37,15 @@ ospf_rt_initort(struct fib_node *fn)
 }
 
 static inline int
-unresolved_vlink(struct mpnh *nhs)
+nh_is_vlink(struct mpnh *nhs)
 {
-  return nhs && !nhs->iface;
+  return !nhs->iface;
+}
+
+static inline int
+unresolved_vlink(ort *ort)
+{
+  return ort->n.nhs && nh_is_vlink(ort->n.nhs);
 }
 
 static inline struct mpnh *
@@ -54,7 +60,7 @@ new_nexthop(struct proto_ospf *po, ip_addr gw, struct iface *iface, unsigned cha
 }
 
 static inline struct mpnh *
-copy_nexthop(struct proto_ospf *po, struct mpnh *src)
+copy_nexthop(struct proto_ospf *po, const struct mpnh *src)
 {
   struct mpnh *nh = lp_alloc(po->nhpool, sizeof(struct mpnh));
   nh->gw = src->gw;
@@ -64,72 +70,138 @@ copy_nexthop(struct proto_ospf *po, struct mpnh *src)
   return nh;
 }
 
-
-/* If new is better return 1 */
+/* Compare nexthops during merge.
+   We need to maintain nhs sorted to eliminate duplicities */
 static int
-ri_better(struct proto_ospf *po, orta *new, orta *old)
+cmp_nhs(struct mpnh *s1, struct mpnh *s2)
 {
-  if (old->type == RTS_DUMMY)
-    return 1;
+  int r;
 
-  if (new->type < old->type)
+  if (!s1)
     return 1;
 
-  if (new->type > old->type)
-    return 0;
+  if (!s2)
+    return -1;
 
-  if (new->metric1 < old->metric1)
-    return 1;
+  r = ((int) s2->weight) - ((int) s1->weight);
+  if (r)
+    return r;
 
-  if (new->metric1 > old->metric1)
-    return 0;
+  r = ipa_compare(s1->gw, s2->gw);
+  if (r)
+    return r;
 
-  return 0;
+  return ((int) s1->iface->index) - ((int) s2->iface->index);
 }
 
+static struct mpnh *
+merge_nexthops(struct proto_ospf *po, struct mpnh *s1, struct mpnh *s2, int r1, int r2)
+{
+  struct mpnh *root = NULL;
+  struct mpnh **n = &root;
+  int count = po->ecmp;
 
-/* Whether the ASBR or the forward address destination is preferred
-   in AS external route selection according to 16.4.1. */
+  /*
+   * r1, r2 signalize whether we can reuse nexthops from s1, s2.
+   * New nexthops (s2, new) can be reused if they are not inherited
+   * from the parent (i.e. it is allocated in calc_next_hop()).
+   * Current nexthops (s1, en->nhs) can be reused if they weren't
+   * inherited in previous steps (that is stored in nhs_reuse,
+   * i.e. created by merging or allocalted in calc_next_hop()).
+   *
+   * Generally, a node first inherits shared nexthops from its
+   * parent and later possibly gets reusable copy during merging.
+   */
+
+  while ((s1 || s2) && count--)
+  {
+    int cmp = cmp_nhs(s1, s2);
+    if (cmp < 0)
+    {
+      *n = r1 ? s1 : copy_nexthop(po, s1);
+      s1 = s1->next;
+    }
+    else if (cmp > 0)
+    {
+      *n = r2 ? s2 : copy_nexthop(po, s2);
+      s2 = s2->next;
+    }
+    else
+    {
+      *n = r1 ? s1 : (r2 ? s2 : copy_nexthop(po, s1));
+      s1 = s1->next;
+      s2 = s2->next;
+    }
+    n = &((*n)->next);
+  }
+  *n = NULL;
+
+  return root;
+}
+
+/* Returns true if there are device nexthops in n */
 static inline int
-epath_preferred(orta *ep)
+has_device_nexthops(const struct mpnh *n)
 {
-  return (ep->type == RTS_OSPF) && (ep->oa->areaid != 0);
+  for (; n; n = n->next)
+    if (ipa_zero(n->gw))
+      return 1;
+
+  return 0;
 }
 
-/* 16.4. (3), return 1 if new is better */
-static int
-ri_better_asbr(struct proto_ospf *po, orta *new, orta *old)
+/* Replace device nexthops with nexthops to gw */
+static struct mpnh *
+fix_device_nexthops(struct proto_ospf *po, const struct mpnh *n, ip_addr gw)
 {
-  if (old->type == RTS_DUMMY)
-    return 1;
+  struct mpnh *root1 = NULL;
+  struct mpnh *root2 = NULL;
+  struct mpnh **nn1 = &root1;
+  struct mpnh **nn2 = &root2;
 
-  if (!po->rfc1583)
-  {
-    int new_pref = epath_preferred(new);
-    int old_pref = epath_preferred(old);
+  /* This is a bit tricky. We cannot just copy the list and update n->gw,
+     because the list should stay sorted, so we create two lists, one with new
+     gateways and one with old ones, and then merge them. */
 
-    if (new_pref > old_pref)
-      return 1;
+  for (; n; n = n->next)
+  {
+    struct mpnh *nn = new_nexthop(po, ipa_zero(n->gw) ? gw : n->gw, n->iface, n->weight);
 
-    if (new_pref < old_pref)
-      return 0;
+    if (ipa_zero(n->gw))
+    {
+      *nn1 = nn;
+      nn1 = &(nn->next);
+    }
+    else
+    {
+      *nn2 = nn;
+      nn2 = &(nn->next);
+    }
   }
 
-  if (new->metric1 < old->metric1)
-    return 1;
+  return merge_nexthops(po, root1, root2, 1, 1);
+}
 
-  if (new->metric1 > old->metric1)
-    return 0;
 
-  /* Larger area ID is preferred */
-  if (new->oa->areaid > old->oa->areaid)
-    return 1;
+/* Whether the ASBR or the forward address destination is preferred
+   in AS external route selection according to 16.4.1. */
+static inline int
+epath_preferred(const orta *ep)
+{
+  return (ep->type == RTS_OSPF) && (ep->oa->areaid != 0);
+}
 
-  return 0;
+/* Whether the ext route has ASBR/next_hop marked as preferred. */
+static inline int
+orta_pref(const orta *nf)
+{
+  return !!(nf->options & ORTA_PREF);
 }
 
+/* Classify orta entries according to RFC 3101 2.5 (6e) priorities:
+   Type-7 LSA with P-bit, Type-5 LSA, Type-7 LSA without P-bit */
 static int
-orta_prio(orta *nf)
+orta_prio(const orta *nf)
 {
   /* RFC 3103 2.5 (6e) priorities */
   u32 opts = nf->options & (ORTA_NSSA | ORTA_PROP);
@@ -145,98 +217,246 @@ orta_prio(orta *nf)
   return 0;
 }
 
-/* 16.4. (6), return 1 if new is better */
+/* Whether the route is better according to RFC 3101 2.5 (6e):
+   Prioritize Type-7 LSA with P-bit, then Type-5 LSA, then higher router ID */
+static int
+orta_prefer_lsa(const orta *new, const orta *old)
+{
+  int pn = orta_prio(new);
+  int po = orta_prio(old);
+
+  return (pn > po) || ((pn == po) && (new->en->lsa.rt > old->en->lsa.rt));
+}
+
+/*
+ * Compare an existing routing table entry with a new one. Applicable for
+ * intra-area routes, inter-area routes and router entries. Returns integer
+ * <, = or > than 0 if the new orta is less, equal or more preferred than
+ * the old orta.
+ */
 static int
-ri_better_ext(struct proto_ospf *po, orta *new, orta *old)
+orta_compare(const struct proto_ospf *po, const orta *new, const orta *old)
 {
+  int r;
+
   if (old->type == RTS_DUMMY)
     return 1;
 
-  /* 16.4. (6a) */
-  if (new->type < old->type)
+  /* Prefer intra-area to inter-area to externals */
+  r = ((int) old->type) - ((int) new->type);
+  if (r) return r;
+
+  /* Prefer lowest type 1 metric */
+  r = ((int) old->metric1) - ((int) new->metric1);
+  if (r) return r;
+
+
+  /* Rest is BIRD-specific */
+
+  /* Area-wide routes should not mix next-hops from different areas.
+     This generally should not happen unless there is some misconfiguration. */
+  if (new->oa->areaid != old->oa->areaid)
+    return (new->oa->areaid > old->oa->areaid) ? 1 : -1;
+
+  /* Prefer routes for configured stubnets (!nhs) to regular routes to dummy
+     vlink nexthops. We intentionally return -1 if both are stubnets or vlinks. */
+  if (!old->nhs)
+    return -1;
+  if (!new->nhs)
     return 1;
+  if (nh_is_vlink(new->nhs))
+    return -1;
+  if (nh_is_vlink(old->nhs))
+    return 1;
+
 
-  if (new->type > old->type)
+  if (po->ecmp)
     return 0;
 
-  /* 16.4. (6b), same type */
-  if (new->type == RTS_OSPF_EXT2)
-  {
-    if (new->metric2 < old->metric2)
-      return 1;
+  /* Prefer routes with higher Router ID, just to be more deterministic */
+  if (new->rid > old->rid)
+    return 1;
+  
+  return -1;
+}
 
-    if (new->metric2 > old->metric2)
-      return 0;
-  }
+/*
+ * Compare ASBR routing table entry with a new one, used for precompute ASBRs
+ * for AS external route selection (RFC 2328 16.4 (3)), Returns integer < or >
+ * than 0 if the new ASBR is less or more preferred than the old ASBR.
+ */
+static int
+orta_compare_asbr(const struct proto_ospf *po, const orta *new, const orta *old)
+{
+  int r;
+
+  if (old->type == RTS_DUMMY)
+    return 1;
 
-  /* 16.4. (6c) */
   if (!po->rfc1583)
   {
-    u32 new_pref = new->options & ORTA_PREF;
-    u32 old_pref = old->options & ORTA_PREF;
-
-    if (new_pref > old_pref)
-      return 1;
-
-    if (new_pref < old_pref)
-      return 0;
+    r = epath_preferred(new) - epath_preferred(old);
+    if (r) return r;
   }
 
-  /* 16.4. (6d) */
-  if (new->metric1 < old->metric1)
+  r = ((int) old->metric1) - ((int) new->metric1);
+  if (r) return r;
+
+  /* Larger area ID is preferred */
+  if (new->oa->areaid > old->oa->areaid)
     return 1;
 
-  if (new->metric1 > old->metric1)
-    return 0;
+  /* There is just one ASBR of that RID per area, so tie is not possible */
+  return -1;
+}
 
-  /* RFC 3103, 2.5. (6e) */
-  int new_prio = orta_prio(new);
-  int old_prio = orta_prio(old);
+/*
+ * Compare a routing table entry with a new one, for AS external routes
+ * (RFC 2328 16.4) and NSSA routes (RFC 3103 2.5), Returns integer <, = or >
+ * than 0 if the new orta is less, equal or more preferred than the old orta.
+ */
+static int
+orta_compare_ext(const struct proto_ospf *po, const orta *new, const orta *old)
+{
+  int r;
 
-  if (new_prio > old_prio)
+  if (old->type == RTS_DUMMY)
     return 1;
 
-  if (old_prio > new_prio)
+  /* 16.4 (6a) - prefer routes with lower type */
+  r = ((int) old->type) - ((int) new->type);
+  if (r) return r;
+
+  /* 16.4 (6b) - prefer routes with lower type 2 metric */
+  if (new->type == RTS_OSPF_EXT2)
+  {
+    r = ((int) old->metric2) - ((int) new->metric2);
+    if (r) return r;
+  }
+
+  /* 16.4 (6c) - if not RFC1583, prefer routes with preferred ASBR/next_hop */
+  if (!po->rfc1583)
+  {
+    r = orta_pref(new) - orta_pref(old);
+    if (r) return r;
+  }
+
+  /* 16.4 (6d) - prefer routes with lower type 1 metric */
+  r = ((int) old->metric1) - ((int) new->metric1);
+  if (r) return r;
+
+
+  if (po->ecmp && po->merge_external)
     return 0;
 
-  /* make it more deterministic */
-  if (new->rid > old->rid)
+  /*
+   * RFC 3101 2.5 (6e) - prioritize Type-7 LSA with P-bit, then Type-5 LSA, then
+   * LSA with higher router ID. Although this should apply just to functionally
+   * equivalent LSAs (i.e. ones with the same non-zero forwarding address), we
+   * use it also to disambiguate otherwise equally preferred nexthops.
+   */
+  if (orta_prefer_lsa(new, old))
     return 1;
 
-  return 0;
+  return -1;
+}
+
+
+static inline void
+ort_replace(ort *o, const orta *new)
+{
+  memcpy(&o->n, new, sizeof(orta));
+}
+
+static void
+ort_merge(struct proto_ospf *po, ort *o, const orta *new)
+{
+  orta *old = &o->n;
+
+  if (old->nhs != new->nhs)
+  {
+    old->nhs = merge_nexthops(po, old->nhs, new->nhs, old->nhs_reuse, new->nhs_reuse);
+    old->nhs_reuse = 1;
+  }
+
+  if (old->rid < new->rid)
+    old->rid = new->rid;
 }
 
+static void
+ort_merge_ext(struct proto_ospf *po, ort *o, const orta *new)
+{
+  orta *old = &o->n;
+
+  if (old->nhs != new->nhs)
+  {
+    old->nhs = merge_nexthops(po, old->nhs, new->nhs, old->nhs_reuse, new->nhs_reuse);
+    old->nhs_reuse = 1;
+  }
+
+  if (old->tag != new->tag)
+    old->tag = 0;
+
+  /*
+   * Even with multipath, we store only one LSA in orta.en for the purpose of
+   * NSSA/ext translation. Therefore, we apply procedures from RFC 3101 2.5 (6e)
+   * to all chosen LSAs for given network, not just to functionally equivalent
+   * ones (i.e. ones with the same non-zero forwarding address).
+   */
+  if (orta_prefer_lsa(new, old))
+  {
+    old->options = new->options;
+    old->rid = new->rid;
+    old->oa = new->oa;
+    old->en = new->en;
+  }
+}
+
+
+
 static inline void
-ri_install_net(struct proto_ospf *po, ip_addr prefix, int pxlen, orta *new)
+ri_install_net(struct proto_ospf *po, ip_addr prefix, int pxlen, const orta *new)
 {
   ort *old = (ort *) fib_get(&po->rtf, &prefix, pxlen);
-  if (ri_better(po, new, &old->n))
-    memcpy(&old->n, new, sizeof(orta));
+  int cmp = orta_compare(po, new, &old->n);
+
+  if (cmp > 0)
+    ort_replace(old, new);
+  else if (cmp == 0)
+    ort_merge(po, old, new);
 }
 
 static inline void
-ri_install_rt(struct ospf_area *oa, u32 rid, orta *new)
+ri_install_rt(struct ospf_area *oa, u32 rid, const orta *new)
 {
   ip_addr addr = ipa_from_rid(rid);
   ort *old = (ort *) fib_get(&oa->rtr, &addr, MAX_PREFIX_LENGTH);
-  if (ri_better(oa->po, new, &old->n))
-    memcpy(&old->n, new, sizeof(orta));
+  int cmp = orta_compare(oa->po, new, &old->n);
+
+  if (cmp > 0)
+    ort_replace(old, new);
+  else if (cmp == 0)
+    ort_merge(oa->po, old, new);
 }
 
 static inline void
-ri_install_asbr(struct proto_ospf *po, ip_addr *addr, orta *new)
+ri_install_asbr(struct proto_ospf *po, ip_addr *addr, const orta *new)
 {
   ort *old = (ort *) fib_get(&po->backbone->rtr, addr, MAX_PREFIX_LENGTH);
-  if (ri_better_asbr(po, new, &old->n))
-    memcpy(&old->n, new, sizeof(orta));
+  if (orta_compare_asbr(po, new, &old->n) > 0)
+    ort_replace(old, new);
 }
 
 static inline void
-ri_install_ext(struct proto_ospf *po, ip_addr prefix, int pxlen, orta *new)
+ri_install_ext(struct proto_ospf *po, ip_addr prefix, int pxlen, const orta *new)
 {
   ort *old = (ort *) fib_get(&po->rtf, &prefix, pxlen);
-  if (ri_better_ext(po, new, &old->n))
-    memcpy(&old->n, new, sizeof(orta));
+  int cmp = orta_compare_ext(po, new, &old->n);
+
+  if (cmp > 0)
+    ort_replace(old, new);
+  else if (cmp == 0)
+    ort_merge_ext(po, old, new);
 }
 
 static inline struct ospf_iface *
@@ -842,7 +1062,7 @@ ospf_rt_sum_tr(struct ospf_area *oa)
 
     /* 16.3. (5) */
     if ((metric < re->n.metric1) || 
-	((metric == re->n.metric1) && unresolved_vlink(re->n.nhs)))
+	((metric == re->n.metric1) && unresolved_vlink(re)))
     {
       /* We want to replace the next-hop even if the metric is equal
 	 to replace a virtual next-hop through vlink with a real one.
@@ -1129,13 +1349,24 @@ ospf_rt_abr1(struct proto_ospf *po)
   struct area_net *anet;
   ort *nf, *default_nf;
 
+  /* RFC 2328 G.3 - incomplete resolution of virtual next hops - routers */
+  FIB_WALK(&po->backbone->rtr, nftmp)
+  {
+    nf = (ort *) nftmp;
+
+    if (nf->n.type && unresolved_vlink(nf))
+      reset_ri(nf);
+  }
+  FIB_WALK_END;
+
+
   FIB_WALK(&po->rtf, nftmp)
   {
     nf = (ort *) nftmp;
 
 
-    /* RFC 2328 G.3 - incomplete resolution of virtual next hops */
-    if (nf->n.type && unresolved_vlink(nf->n.nhs))
+    /* RFC 2328 G.3 - incomplete resolution of virtual next hops - networks */
+    if (nf->n.type && unresolved_vlink(nf))
       reset_ri(nf);
 
 
@@ -1361,7 +1592,7 @@ static void
 ospf_ext_spf(struct proto_ospf *po)
 {
   ort *nf1, *nf2;
-  orta nfa;
+  orta nfa = {};
   struct top_hash_entry *en;
   struct proto *p = &po->proto;
   struct ospf_lsa_ext *le;
@@ -1369,7 +1600,6 @@ ospf_ext_spf(struct proto_ospf *po)
   ip_addr ip, rtid, rt_fwaddr;
   u32 br_metric, rt_metric, rt_tag;
   struct ospf_area *atmp;
-  struct mpnh* nhs = NULL;
 
   OSPF_TRACE(D_EVENTS, "Starting routing table calculation for ext routes");
 
@@ -1465,7 +1695,7 @@ ospf_ext_spf(struct proto_ospf *po)
     if (!rt_fwaddr_valid)
     {
       nf2 = nf1;
-      nhs = nf1->n.nhs;
+      nfa.nhs = nf1->n.nhs;
       br_metric = nf1->n.metric1;
     }
     else
@@ -1491,11 +1721,15 @@ ospf_ext_spf(struct proto_ospf *po)
       if (!nf2->n.nhs)
 	continue;
 
-      nhs = nf2->n.nhs;
-      /* If gw is zero, it is a device route */
-      if (ipa_zero(nhs->gw))
-	nhs = new_nexthop(po, rt_fwaddr, nhs->iface, nhs->weight);
+      nfa.nhs = nf2->n.nhs;
       br_metric = nf2->n.metric1;
+
+      /* Replace device nexthops with nexthops to forwarding address from LSA */
+      if (has_device_nexthops(nfa.nhs))
+      {
+	nfa.nhs = fix_device_nexthops(po, nfa.nhs, rt_fwaddr);
+	nfa.nhs_reuse = 1;
+      }
     }
 
     if (ebit)
@@ -1526,8 +1760,6 @@ ospf_ext_spf(struct proto_ospf *po)
     nfa.tag = rt_tag;
     nfa.rid = en->lsa.rt;
     nfa.oa = atmp; /* undefined in RFC 2328 */
-    nfa.voa = NULL;
-    nfa.nhs = nhs;
     nfa.en = en; /* store LSA for later (NSSA processing) */
 
     ri_install_ext(po, ip, pxlen, &nfa);
@@ -1745,81 +1977,6 @@ calc_next_hop(struct ospf_area *oa, struct top_hash_entry *en,
   return NULL;
 }
 
-/* Compare nexthops during merge.
-   We need to maintain nhs sorted to eliminate duplicities */
-static int
-cmp_nhs(struct mpnh *s1, struct mpnh *s2)
-{
-  int r;
-
-  if (!s1)
-    return 1;
-
-  if (!s2)
-    return -1;
-
-  r = ((int) s2->weight) - ((int) s1->weight);
-  if (r)
-    return r;
-
-  r = ipa_compare(s1->gw, s2->gw);
-  if (r)
-    return r;
-
-  return ((int) s1->iface->index) - ((int) s2->iface->index);
-}
-
-static void
-merge_nexthops(struct proto_ospf *po, struct top_hash_entry *en,
-	       struct top_hash_entry *par, struct mpnh *new)
-{
-  if (en->nhs == new)
-    return;
-
-  int r1 = en->nhs_reuse;
-  int r2 = (par->nhs != new);
-  int count = po->ecmp;
-  struct mpnh *s1 = en->nhs;
-  struct mpnh *s2 = new;
-  struct mpnh **n = &(en->nhs);
-
-  /*
-   * r1, r2 signalize whether we can reuse nexthops from s1, s2.
-   * New nexthops (s2, new) can be reused if they are not inherited
-   * from the parent (i.e. it is allocated in calc_next_hop()).
-   * Current nexthops (s1, en->nhs) can be reused if they weren't
-   * inherited in previous steps (that is stored in nhs_reuse,
-   * i.e. created by merging or allocalted in calc_next_hop()).
-   *
-   * Generally, a node first inherits shared nexthops from its
-   * parent and later possibly gets reusable copy during merging.
-   */
-
-  while ((s1 || s2) && count--)
-  {
-    int cmp = cmp_nhs(s1, s2);
-    if (cmp < 0)
-    {
-      *n = r1 ? s1 : copy_nexthop(po, s1);
-      s1 = s1->next;
-    }
-    else if (cmp > 0)
-    {
-      *n = r2 ? s2 : copy_nexthop(po, s2);
-      s2 = s2->next;
-    }
-    else
-    {
-      *n = r1 ? s1 : (r2 ? s2 : copy_nexthop(po, s1));
-      s1 = s1->next;
-      s2 = s2->next;
-    }
-    n = &((*n)->next);
-  }
-  *n = NULL;
-
-  en->nhs_reuse=1;
-}
 
 /* Add LSA into list of candidates in Dijkstra's algorithm */
 static void
@@ -1866,31 +2023,33 @@ add_cand(list * l, struct top_hash_entry *en, struct top_hash_entry *par,
     return;
   }
 
-  if (dist == en->dist)
+  /* We know that en->color == CANDIDATE and en->nhs is defined. */
+
+  if ((dist == en->dist) && !nh_is_vlink(en->nhs))
   {
     /*
-     * For multipath, we should merge nexthops. We do not mix dummy
-     * vlink nexthops, device nexthops and gateway nexthops. We merge
-     * gateway nexthops only. We prefer device nexthops over gateway
-     * nexthops and gateway nexthops over vlink nexthops. We either
-     * keep old nexthops, merge old and new, or replace old with new.
-     * 
-     * We know that en->color == CANDIDATE and en->nhs is defined.
+     * For multipath, we should merge nexthops. We merge regular nexthops only.
+     * Dummy vlink nexthops are less preferred and handled as a special case.
+     *
+     * During merging, new nexthops (nhs) can be reused if they are not
+     * inherited from the parent (i.e. they are allocated in calc_next_hop()).
+     * Current nexthops (en->nhs) can be reused if they weren't inherited in
+     * previous steps (that is stored in nhs_reuse, i.e. created by merging or
+     * allocated in calc_next_hop()).
+     *
+     * Generally, a node first inherits shared nexthops from its parent and
+     * later possibly gets reusable copy during merging.
      */
-    struct mpnh *onhs = en->nhs;
 
     /* Keep old ones */
-    if (!po->ecmp || !nhs->iface || (onhs->iface && ipa_zero(onhs->gw)))
+    if (!po->ecmp || nh_is_vlink(nhs) || (nhs == en->nhs))
       return;
 
     /* Merge old and new */
-    if (ipa_nonzero(nhs->gw) && ipa_nonzero(onhs->gw))
-    {
-      merge_nexthops(po, en, par, nhs);
-      return;
-    }
-
-    /* Fallback to replace old ones */
+    int new_reuse = (par->nhs != nhs);
+    en->nhs = merge_nexthops(po, en->nhs, nhs, en->nhs_reuse, new_reuse);
+    en->nhs_reuse = 1;
+    return;
   }
 
   DBG("     Adding candidate: rt: %R, id: %R, type: %u\n",
diff --git a/proto/ospf/rt.h b/proto/ospf/rt.h
index 1ba76e3c..a11748fc 100644
--- a/proto/ospf/rt.h
+++ b/proto/ospf/rt.h
@@ -16,7 +16,8 @@
 
 typedef struct orta
 {
-  int type;
+  u8 type;			/* RTS_OSPF_* */
+  u8 nhs_reuse;			/* Whether nhs nodes can be reused during merging */
   u32 options;
   /*
    * For ORT_ROUTER routes, options field are router-LSA style
@@ -93,16 +94,24 @@ static inline int rt_is_nssa(ort *nf)
  * - n.metric1 may be at most a small multiple of LSINFINITY,
  *   therefore sums do not overflow
  * - n.oa is always non-NULL
- * - n.nhs is always non-NULL with one exception - configured stubnet
- *   nodes (in po->rtf).
+ * - n.nhs is always non-NULL unless it is configured stubnet
+ * - n.en is non-NULL for external routes, NULL for intra/inter area routes.
  * - oa->rtr does not contain calculating router itself
  *
- * There are three types of nexthops in nhs fields:
+ * There are four types of nexthops in nhs fields:
  * - gateway nexthops (non-NULL iface, gw != IPA_NONE)
  * - device nexthops (non-NULL iface, gw == IPA_NONE)
  * - dummy vlink nexthops (NULL iface, gw == IPA_NONE)
- * These three types don't mix, nhs field contains either
- * one device, one vlink node, or one/more gateway nodes.
+ * - configured stubnets (nhs is NULL, only RTS_OSPF orta nodes in po->rtf)
+ *
+ * Dummy vlink nexthops and configured stubnets cannot be mixed with
+ * regular ones, nhs field contains either list of gateway+device nodes,
+ * one vlink node, or NULL for configured stubnet.
+ *
+ * Dummy vlink nexthops can appear in both network (rtf) and backbone area router
+ * (rtr) tables for regular and inter-area routes, but only if areano > 1. They are
+ * replaced in ospf_rt_sum_tr() and removed in ospf_rt_abr1(), therefore cannot
+ * appear in ASBR pre-selection and external routes processing.
  */
 
 void ospf_rt_spf(struct proto_ospf *po);