zebra: implement per-route mtu handling

This commits allow overriding MTU using netlink attributes on
per-route basis. This is useful for routing protocols that can
advertice prefix specific MTUs between routers (e.g. NHRP).

Signed-off-by: Timo Teräs <timo.teras@iki.fi>
(cherry picked from commit b11f3b54c842117e22e2f5cf1561ea34eee8dfcc)

10 files changed, 110 insertions, 17 deletions

diff --git a/lib/zclient.c b/lib/zclient.c
index a28db083cb..0259af5971 100644
--- a/lib/zclient.c
+++ b/lib/zclient.c
@@ -709,6 +709,8 @@ zclient_connect (struct thread *t)
   *
   * If ZAPI_MESSAGE_TAG is set, the tag value is written as a 2 byte value
   *
+  * If ZAPI_MESSAGE_MTU is set, the mtu value is written as a 4 byte value
+  *
   * XXX: No attention paid to alignment.
   */ 
 int
@@ -769,6 +771,8 @@ zapi_ipv4_route (u_char cmd, struct zclient *zclient, struct prefix_ipv4 *p,
     stream_putl (s, api->metric);
   if (CHECK_FLAG (api->message, ZAPI_MESSAGE_TAG))
     stream_putw (s, api->tag);
+  if (CHECK_FLAG (api->message, ZAPI_MESSAGE_MTU))
+    stream_putl (s, api->mtu);
 
   /* Put length at the first point of the stream. */
   stream_putw_at (s, 0, stream_get_endp (s));
@@ -834,6 +838,8 @@ zapi_ipv4_route_ipv6_nexthop (u_char cmd, struct zclient *zclient,
     stream_putl (s, api->metric);
   if (CHECK_FLAG (api->message, ZAPI_MESSAGE_TAG))
     stream_putw (s, api->tag);
+  if (CHECK_FLAG (api->message, ZAPI_MESSAGE_MTU))
+    stream_putl (s, api->mtu);
 
   /* Put length at the first point of the stream. */
   stream_putw_at (s, 0, stream_get_endp (s));
@@ -898,6 +904,8 @@ zapi_ipv6_route (u_char cmd, struct zclient *zclient, struct prefix_ipv6 *p,
     stream_putl (s, api->metric);
   if (CHECK_FLAG (api->message, ZAPI_MESSAGE_TAG))
     stream_putw (s, api->tag);
+  if (CHECK_FLAG (api->message, ZAPI_MESSAGE_MTU))
+    stream_putl (s, api->mtu);
 
   /* Put length at the first point of the stream. */
   stream_putw_at (s, 0, stream_get_endp (s));
diff --git a/lib/zclient.h b/lib/zclient.h
index 8926bbbe63..c42d8c9aa3 100644
--- a/lib/zclient.h
+++ b/lib/zclient.h
@@ -116,6 +116,7 @@ struct zclient
 #define ZAPI_MESSAGE_DISTANCE 0x04
 #define ZAPI_MESSAGE_METRIC   0x08
 #define ZAPI_MESSAGE_TAG      0x10
+#define ZAPI_MESSAGE_MTU      0x20
 
 /* Zserv protocol message header */
 struct zserv_header
@@ -154,6 +155,8 @@ struct zapi_ipv4
 
   u_short tag;
 
+  u_int32_t mtu;
+
   vrf_id_t vrf_id;
 };
 
@@ -237,6 +240,8 @@ struct zapi_ipv6
 
   u_short tag;
 
+  u_int32_t mtu;
+
   vrf_id_t vrf_id;
 };
 
diff --git a/zebra/connected.c b/zebra/connected.c
index 42b5a8ab9c..3077272221 100644
--- a/zebra/connected.c
+++ b/zebra/connected.c
@@ -201,10 +201,10 @@ connected_up_ipv4 (struct interface *ifp, struct connected *ifc)
     return;
 
   rib_add_ipv4 (ZEBRA_ROUTE_CONNECT, 0, 0, &p, NULL, NULL, ifp->ifindex,
-	ifp->vrf_id, RT_TABLE_MAIN, ifp->metric, 0, SAFI_UNICAST);
+		ifp->vrf_id, RT_TABLE_MAIN, ifp->metric, 0, 0, SAFI_UNICAST);
 
   rib_add_ipv4 (ZEBRA_ROUTE_CONNECT, 0, 0, &p, NULL, NULL, ifp->ifindex,
-	ifp->vrf_id, RT_TABLE_MAIN, ifp->metric, 0, SAFI_MULTICAST);
+		ifp->vrf_id, RT_TABLE_MAIN, ifp->metric, 0, 0, SAFI_MULTICAST);
 
   if (IS_ZEBRA_DEBUG_RIB_DETAILED)
     zlog_debug ("%u: IF %s IPv4 address add/up, scheduling RIB processing",
@@ -379,7 +379,7 @@ connected_up_ipv6 (struct interface *ifp, struct connected *ifc)
 #endif
 
   rib_add_ipv6 (ZEBRA_ROUTE_CONNECT, 0, 0, &p, NULL, ifp->ifindex, ifp->vrf_id,
-                RT_TABLE_MAIN, ifp->metric, 0, SAFI_UNICAST);
+                RT_TABLE_MAIN, ifp->metric, 0, 0, SAFI_UNICAST);
 
   if (IS_ZEBRA_DEBUG_RIB_DETAILED)
     zlog_debug ("%u: IF %s IPv6 address down, scheduling RIB processing",
diff --git a/zebra/kernel_socket.c b/zebra/kernel_socket.c
index 638062d5d6..c5763dd46f 100644
--- a/zebra/kernel_socket.c
+++ b/zebra/kernel_socket.c
@@ -974,7 +974,8 @@ rtm_read (struct rt_msghdr *rtm)
           || rtm->rtm_type == RTM_ADD
           || rtm->rtm_type == RTM_CHANGE)
 	rib_add_ipv4 (ZEBRA_ROUTE_KERNEL, 0, zebra_flags,
-		      &p, &gate.sin.sin_addr, NULL, 0, VRF_DEFAULT, 0, 0, 0, SAFI_UNICAST);
+		      &p, &gate.sin.sin_addr, NULL, 0, VRF_DEFAULT,
+		      0, 0, 0, 0, SAFI_UNICAST);
       else
 	rib_delete_ipv4 (ZEBRA_ROUTE_KERNEL, 0, zebra_flags,
 		      &p, &gate.sin.sin_addr, 0, VRF_DEFAULT, 0, SAFI_UNICAST);
@@ -1016,7 +1017,8 @@ rtm_read (struct rt_msghdr *rtm)
           || rtm->rtm_type == RTM_ADD
           || rtm->rtm_type == RTM_CHANGE)
 	rib_add_ipv6 (ZEBRA_ROUTE_KERNEL, 0, zebra_flags,
-		      &p, &gate.sin6.sin6_addr, ifindex, VRF_DEFAULT, 0, 0, 0, SAFI_UNICAST);
+		      &p, &gate.sin6.sin6_addr, ifindex, VRF_DEFAULT,
+		      0, 0, 0, 0, SAFI_UNICAST);
       else
 	rib_delete_ipv6 (ZEBRA_ROUTE_KERNEL, 0, zebra_flags,
 			 &p, &gate.sin6.sin6_addr, ifindex, VRF_DEFAULT, 0, SAFI_UNICAST);
diff --git a/zebra/redistribute.c b/zebra/redistribute.c
index 4238d8329e..892bc2af2b 100644
--- a/zebra/redistribute.c
+++ b/zebra/redistribute.c
@@ -578,7 +578,7 @@ zebra_add_import_table_entry (struct route_node *rn, struct rib *rib, const char
 	      rib_add_ipv4(ZEBRA_ROUTE_TABLE, rib->table, 0, &p4,
                            gate, &nhop->src.ipv4,
 		           nhop->ifindex, rib->vrf_id, zebrad.rtm_table_default,
-		           rib->metric,
+		           rib->metric, rib->mtu,
 		           zebra_import_table_distance[AFI_IP][rib->table],
 		           SAFI_UNICAST);
 	    }
@@ -589,6 +589,7 @@ zebra_add_import_table_entry (struct route_node *rn, struct rib *rib, const char
 	      newrib->distance = zebra_import_table_distance[AFI_IP][rib->table];
 	      newrib->flags = rib->flags;
 	      newrib->metric = rib->metric;
+	      newrib->mtu = rib->mtu;
 	      newrib->table = zebrad.rtm_table_default;
 	      newrib->nexthop_num = 0;
 	      newrib->uptime = time(NULL);
diff --git a/zebra/rib.h b/zebra/rib.h
index a2109d8c17..095b9c6f8f 100644
--- a/zebra/rib.h
+++ b/zebra/rib.h
@@ -64,6 +64,10 @@ struct rib
   /* Metric */
   u_int32_t metric;
 
+  /* MTU */
+  u_int32_t mtu;
+  u_int32_t nexthop_mtu;
+
   /* Distance. */
   u_char distance;
 
@@ -376,7 +380,7 @@ extern int zebra_check_addr (struct prefix *p);
 extern int rib_add_ipv4 (int type, u_short instance, int flags, struct prefix_ipv4 *p,
 			 struct in_addr *gate, struct in_addr *src,
 			 ifindex_t ifindex, vrf_id_t vrf_id, u_int32_t table_id,
-			 u_int32_t, u_char, safi_t);
+			 u_int32_t, u_int32_t, u_char, safi_t);
 
 extern int rib_add_ipv4_multipath (struct prefix_ipv4 *, struct rib *, safi_t);
 
@@ -417,7 +421,8 @@ static_delete_ipv4 (safi_t safi, struct prefix *p, struct in_addr *gate, ifindex
 extern int
 rib_add_ipv6 (int type, u_short instance, int flags, struct prefix_ipv6 *p,
 	      struct in6_addr *gate, ifindex_t ifindex, vrf_id_t vrf_id,
-              u_int32_t table_id, u_int32_t metric, u_char distance, safi_t safi);
+              u_int32_t table_id, u_int32_t metric, u_int32_t mtu,
+	      u_char distance, safi_t safi);
 
 extern int
 rib_delete_ipv6 (int type, u_short instance, int flags, struct prefix_ipv6 *p,
diff --git a/zebra/rt_netlink.c b/zebra/rt_netlink.c
index d93a530317..49dcd1e6de 100644
--- a/zebra/rt_netlink.c
+++ b/zebra/rt_netlink.c
@@ -866,6 +866,7 @@ netlink_routing_table (struct sockaddr_nl *snl, struct nlmsghdr *h,
   int index;
   int table;
   int metric;
+  u_int32_t mtu = 0;
 
   void *dest;
   void *gate;
@@ -937,6 +938,18 @@ netlink_routing_table (struct sockaddr_nl *snl, struct nlmsghdr *h,
   if (tb[RTA_PRIORITY])
     metric = *(int *) RTA_DATA(tb[RTA_PRIORITY]);
 
+  if (tb[RTA_METRICS])
+    {
+      struct rtattr *mxrta[RTAX_MAX+1];
+
+      memset (mxrta, 0, sizeof mxrta);
+      netlink_parse_rtattr (mxrta, RTAX_MAX, RTA_DATA(tb[RTA_METRICS]),
+                            RTA_PAYLOAD(tb[RTA_METRICS]));
+
+      if (mxrta[RTAX_MTU])
+        mtu = *(u_int32_t *) RTA_DATA(mxrta[RTAX_MTU]);
+    }
+
   if (rtm->rtm_family == AF_INET)
     {
       struct prefix_ipv4 p;
@@ -946,7 +959,7 @@ netlink_routing_table (struct sockaddr_nl *snl, struct nlmsghdr *h,
 
       if (!tb[RTA_MULTIPATH])
           rib_add_ipv4 (ZEBRA_ROUTE_KERNEL, 0, flags, &p, gate, src, index,
-                        vrf_id, table, metric, 0, SAFI_UNICAST);
+                        vrf_id, table, metric, mtu, 0, SAFI_UNICAST);
       else
         {
           /* This is a multipath route */
@@ -962,6 +975,7 @@ netlink_routing_table (struct sockaddr_nl *snl, struct nlmsghdr *h,
           rib->distance = 0;
           rib->flags = flags;
           rib->metric = metric;
+          rib->mtu = mtu;
           rib->vrf_id = vrf_id;
           rib->table = table;
           rib->nexthop_num = 0;
@@ -1014,7 +1028,7 @@ netlink_routing_table (struct sockaddr_nl *snl, struct nlmsghdr *h,
       p.prefixlen = rtm->rtm_dst_len;
 
       rib_add_ipv6 (ZEBRA_ROUTE_KERNEL, 0, flags, &p, gate, index, vrf_id,
-                    table, metric, 0, SAFI_UNICAST);
+                    table, metric, mtu, 0, SAFI_UNICAST);
     }
 #endif /* HAVE_IPV6 */
 
@@ -1051,6 +1065,7 @@ netlink_route_change (struct sockaddr_nl *snl, struct nlmsghdr *h,
   int index;
   int table;
   int metric;
+  u_int32_t mtu = 0;
 
   void *dest;
   void *gate;
@@ -1142,8 +1157,23 @@ netlink_route_change (struct sockaddr_nl *snl, struct nlmsghdr *h,
   if (tb[RTA_PREFSRC])
     src = RTA_DATA (tb[RTA_PREFSRC]);
 
-  if (h->nlmsg_type == RTM_NEWROUTE && tb[RTA_PRIORITY])
-    metric = *(int *) RTA_DATA(tb[RTA_PRIORITY]);
+  if (h->nlmsg_type == RTM_NEWROUTE)
+    {
+      if (tb[RTA_PRIORITY])
+        metric = *(int *) RTA_DATA(tb[RTA_PRIORITY]);
+
+      if (tb[RTA_METRICS])
+        {
+          struct rtattr *mxrta[RTAX_MAX+1];
+
+          memset (mxrta, 0, sizeof mxrta);
+          netlink_parse_rtattr (mxrta, RTAX_MAX, RTA_DATA(tb[RTA_METRICS]),
+                                RTA_PAYLOAD(tb[RTA_METRICS]));
+
+          if (mxrta[RTAX_MTU])
+            mtu = *(u_int32_t *) RTA_DATA(mxrta[RTAX_MTU]);
+        }
+    }
 
   if (rtm->rtm_family == AF_INET)
     {
@@ -1164,7 +1194,7 @@ netlink_route_change (struct sockaddr_nl *snl, struct nlmsghdr *h,
         {
           if (!tb[RTA_MULTIPATH])
             rib_add_ipv4 (ZEBRA_ROUTE_KERNEL, 0, 0, &p, gate, src, index, vrf_id,
-                          table, metric, 0, SAFI_UNICAST);
+                          table, metric, mtu, 0, SAFI_UNICAST);
           else
             {
               /* This is a multipath route */
@@ -1180,6 +1210,7 @@ netlink_route_change (struct sockaddr_nl *snl, struct nlmsghdr *h,
               rib->distance = 0;
               rib->flags = 0;
               rib->metric = metric;
+              rib->mtu = mtu;
               rib->vrf_id = vrf_id;
               rib->table = table;
               rib->nexthop_num = 0;
@@ -1248,7 +1279,7 @@ netlink_route_change (struct sockaddr_nl *snl, struct nlmsghdr *h,
 
       if (h->nlmsg_type == RTM_NEWROUTE)
         rib_add_ipv6 (ZEBRA_ROUTE_KERNEL, 0, 0, &p, gate, index, vrf_id,
-                      table, metric, 0, SAFI_UNICAST);
+                      table, metric, mtu, 0, SAFI_UNICAST);
       else
         rib_delete_ipv6 (ZEBRA_ROUTE_KERNEL, 0, zebra_flags, &p, gate, index,
                          vrf_id, table, SAFI_UNICAST);
@@ -2079,6 +2110,20 @@ netlink_route_multipath (int cmd, struct prefix *p, struct rib *rib,
       addattr32(&req.n, sizeof req, RTA_TABLE, rib->table);
     }
 
+  if (rib->mtu || rib->nexthop_mtu)
+    {
+      char buf[NL_PKT_BUF_SIZE];
+      struct rtattr *rta = (void *) buf;
+      u_int32_t mtu = rib->mtu;
+      if (!mtu || (rib->nexthop_mtu && rib->nexthop_mtu < mtu))
+        mtu = rib->nexthop_mtu;
+      rta->rta_type = RTA_METRICS;
+      rta->rta_len = RTA_LENGTH(0);
+      rta_addattr_l (rta, NL_PKT_BUF_SIZE, RTAX_MTU, &mtu, sizeof mtu);
+      addattr_l (&req.n, NL_PKT_BUF_SIZE, RTA_METRICS, RTA_DATA (rta),
+                 RTA_PAYLOAD (rta));
+    }
+
   if (discard)
     {
       if (cmd == RTM_NEWROUTE)
diff --git a/zebra/zebra_rib.c b/zebra/zebra_rib.c
index 0dc8086906..195d3633e1 100644
--- a/zebra/zebra_rib.c
+++ b/zebra/zebra_rib.c
@@ -352,6 +352,7 @@ nexthop_active_ipv4 (struct rib *rib, struct nexthop *nexthop, int set,
       zebra_deregister_rnh_static_nexthops(rib->vrf_id, nexthop->resolved, top);
       nexthops_free(nexthop->resolved);
       nexthop->resolved = NULL;
+      rib->nexthop_mtu = 0;
     }
 
   /* Skip nexthops that have been filtered out due to route-map */
@@ -545,6 +546,8 @@ nexthop_active_ipv4 (struct rib *rib, struct nexthop *nexthop, int set,
 		      }
 		    resolved = 1;
 		  }
+              if (resolved && set)
+                rib->nexthop_mtu = match->mtu;
 	      return resolved;
 	    }
 	  else
@@ -2414,7 +2417,7 @@ int
 rib_add_ipv4 (int type, u_short instance, int flags, struct prefix_ipv4 *p,
 	      struct in_addr *gate, struct in_addr *src,
 	      ifindex_t ifindex, vrf_id_t vrf_id, u_int32_t table_id,
-	      u_int32_t metric, u_char distance, safi_t safi)
+	      u_int32_t metric, u_int32_t mtu, u_char distance, safi_t safi)
 {
   struct rib *rib;
   struct rib *same = NULL;
@@ -2481,6 +2484,7 @@ rib_add_ipv4 (int type, u_short instance, int flags, struct prefix_ipv4 *p,
   rib->distance = distance;
   rib->flags = flags;
   rib->metric = metric;
+  rib->mtu = mtu;
   rib->table = table_id;
   rib->vrf_id = vrf_id;
   rib->nexthop_num = 0;
@@ -2554,9 +2558,10 @@ void _rib_dump (const char * func,
   );
   zlog_debug
   (
-    "%s: metric == %u, distance == %u, flags == %u, status == %u",
+    "%s: metric == %u, mtu == %u, distance == %u, flags == %u, status == %u",
     func,
     rib->metric,
+    rib->mtu,
     rib->distance,
     rib->flags,
     rib->status
@@ -2992,6 +2997,7 @@ static_install_route (afi_t afi, safi_t safi, struct prefix *p, struct static_ro
       rib->instance = 0;
       rib->distance = si->distance;
       rib->metric = 0;
+      rib->mtu = 0;
       rib->vrf_id = si->vrf_id;
       rib->table =  si->vrf_id ? (zebra_vrf_lookup(si->vrf_id))->table_id : zebrad.rtm_table_default;
       rib->nexthop_num = 0;
@@ -3342,7 +3348,8 @@ static_delete_ipv4 (safi_t safi, struct prefix *p, struct in_addr *gate, ifindex
 int
 rib_add_ipv6 (int type, u_short instance, int flags, struct prefix_ipv6 *p,
 	      struct in6_addr *gate, ifindex_t ifindex, vrf_id_t vrf_id,
-              u_int32_t table_id, u_int32_t metric, u_char distance, safi_t safi)
+              u_int32_t table_id, u_int32_t metric, u_int32_t mtu,
+	      u_char distance, safi_t safi)
 {
   struct rib *rib;
   struct rib *same = NULL;
@@ -3401,6 +3408,7 @@ rib_add_ipv6 (int type, u_short instance, int flags, struct prefix_ipv6 *p,
   rib->distance = distance;
   rib->flags = flags;
   rib->metric = metric;
+  rib->mtu = mtu;
   rib->table = table_id;
   rib->vrf_id = vrf_id;
   rib->nexthop_num = 0;
diff --git a/zebra/zebra_vty.c b/zebra/zebra_vty.c
index e9090cbed2..ff2f7c7bc4 100644
--- a/zebra/zebra_vty.c
+++ b/zebra/zebra_vty.c
@@ -1929,6 +1929,8 @@ vty_show_ip_route_detail (struct vty *vty, struct route_node *rn, int mcast)
       vty_out (vty, ", distance %u, metric %u", rib->distance, rib->metric);
       if (rib->tag)
 	vty_out (vty, ", tag %d", rib->tag);
+       if (rib->mtu)
+        vty_out (vty, ", mtu %u", rib->mtu);
       if (rib->vrf_id != VRF_DEFAULT)
         {
           zvrf = vrf_info_lookup(rib->vrf_id);
diff --git a/zebra/zserv.c b/zebra/zserv.c
index a1fe9bc188..84881e14a4 100644
--- a/zebra/zserv.c
+++ b/zebra/zserv.c
@@ -695,6 +695,8 @@ zsend_redistribute_route (int cmd, struct zserv *client, struct prefix *p,
           SET_FLAG(zapi_flags, ZAPI_MESSAGE_TAG);
           stream_putw(s, rib->tag);
         }
+      SET_FLAG (zapi_flags, ZAPI_MESSAGE_MTU);
+      stream_putl (s, rib->mtu);
     }
 
   /* write real message flags value */
@@ -1264,6 +1266,11 @@ zread_ipv4_add (struct zserv *client, u_short length, struct zebra_vrf *zvrf)
   else
     rib->tag = 0;
 
+  if (CHECK_FLAG (message, ZAPI_MESSAGE_MTU))
+    rib->mtu = stream_getl (s);
+  else
+    rib->mtu = 0;
+
   /* Table */
   rib->table = zvrf->table_id;
 
@@ -1518,6 +1525,11 @@ zread_ipv4_route_ipv6_nexthop_add (struct zserv *client, u_short length, struct
   else
     rib->tag = 0;
 
+  if (CHECK_FLAG (message, ZAPI_MESSAGE_MTU))
+    rib->mtu = stream_getl (s);
+  else
+    rib->mtu = 0;
+
   /* Table */
   rib->table = zvrf->table_id;
 
@@ -1636,6 +1648,11 @@ zread_ipv6_add (struct zserv *client, u_short length, struct zebra_vrf *zvrf)
   else
     rib->tag = 0;
 
+  if (CHECK_FLAG (message, ZAPI_MESSAGE_MTU))
+    rib->mtu = stream_getl (s);
+  else
+    rib->mtu = 0;
+
   /* VRF ID */
   rib->vrf_id = zvrf->vrf_id;
   rib->table = zvrf->table_id;