1 files changed, 63 insertions, 226 deletions

diff --git a/sysdep/linux/netlink.c b/sysdep/linux/netlink.c
index 099ae6e9..3443957b 100644
--- a/sysdep/linux/netlink.c
+++ b/sysdep/linux/netlink.c
@@ -75,51 +75,16 @@
 #endif
 
 #define krt_ipv4(p) ((p)->af == AF_INET)
-#define krt_ecmp6(p) ((p)->af == AF_INET6)
 
 const int rt_default_ecmp = 16;
 
-/*
- * Structure nl_parse_state keeps state of received route processing. Ideally,
- * we could just independently parse received Netlink messages and immediately
- * propagate received routes to the rest of BIRD, but older Linux kernel (before
- * version 4.11) represents and announces IPv6 ECMP routes not as one route with
- * multiple next hops (like RTA_MULTIPATH in IPv4 ECMP), but as a sequence of
- * routes with the same prefix. More recent kernels work as with IPv4.
- *
- * Therefore, BIRD keeps currently processed route in nl_parse_state structure
- * and postpones its propagation until we expect it to be final; i.e., when
- * non-matching route is received or when the scan ends. When another matching
- * route is received, it is merged with the already processed route to form an
- * ECMP route. Note that merging is done only for IPv6 (merge == 1), but the
- * postponing is done in both cases (for simplicity). All IPv4 routes or IPv6
- * routes with RTA_MULTIPATH set are just considered non-matching.
- *
- * This is ignored for asynchronous notifications (every notification is handled
- * as a separate route). It is not an issue for our routes, as we ignore such
- * notifications anyways. But importing alien IPv6 ECMP routes does not work
- * properly with older kernels.
- *
- * Whatever the kernel version is, IPv6 ECMP routes are sent as multiple routes
- * for the same prefix.
- */
-
 struct nl_parse_state
 {
+  struct krt_proto *proto;
   struct linpool *pool;
   int scan;
-  int merge;
 
-  net_addr *net;
-  ea_list *attrs;
-  struct krt_proto *proto;
-  s8 new;
-  s8 krt_src;
-  u8 krt_type;
-  u8 krt_proto;
-  u32 krt_metric;
-
-  u32 rta_flow;		/* Used during parsing */
+  u32 rta_flow;
 };
 
 /*
@@ -1453,10 +1418,13 @@ nh_bufsize(struct nexthop_adata *nhad)
 }
 
 static int
-nl_send_route(struct krt_proto *p, const rte *e, int op, int dest, struct nexthop_adata *nh)
+nl_send_route(struct krt_proto *p, const rte *e, int op)
 {
   eattr *ea;
   ea_list *eattrs = e->attrs;
+  eattr *nhea = ea_find(eattrs, &ea_gen_nexthop);
+  struct nexthop_adata *nh = nhea ? (struct nexthop_adata *) nhea->u.ptr : NULL;
+  int dest = nhea_dest(nhea);
 
   int bufsize = 128 + KRT_METRICS_MAX*8 + (nh ? nh_bufsize(nh) : 0);
   u32 priority = 0;
@@ -1534,15 +1502,17 @@ nl_send_route(struct krt_proto *p, const rte *e, int op, int dest, struct nextho
 
   /* For route delete, we do not specify remaining route attributes */
   if (op == NL_OP_DELETE)
-    goto dest;
+    goto done;
 
   /* Default scope is LINK for device routes, UNIVERSE otherwise */
   if (p->af == AF_MPLS)
     r->r.rtm_scope = RT_SCOPE_UNIVERSE;
   else if (ea = ea_find(eattrs, &ea_krt_scope))
     r->r.rtm_scope = ea->u.data;
+  else if (dest == RTD_UNICAST && ipa_zero(nh->nh.gw))
+    r->r.rtm_scope = RT_SCOPE_LINK;
   else
-    r->r.rtm_scope = (dest == RTD_UNICAST && ipa_zero(nh->nh.gw)) ? RT_SCOPE_LINK : RT_SCOPE_UNIVERSE;
+    r->r.rtm_scope = RT_SCOPE_UNIVERSE;
 
   if (ea = ea_find(eattrs, &ea_krt_prefsrc))
     nl_add_attr_ipa(&r->h, rsize, RTA_PREFSRC, *(ip_addr *)ea->u.ptr->data);
@@ -1565,13 +1535,11 @@ nl_send_route(struct krt_proto *p, const rte *e, int op, int dest, struct nextho
   if (metrics[0])
     nl_add_metrics(&r->h, rsize, metrics, KRT_METRICS_MAX);
 
-
-dest:
   switch (dest)
     {
     case RTD_UNICAST:
       r->r.rtm_type = RTN_UNICAST;
-      if (!NEXTHOP_ONE(nh) && !krt_ecmp6(p))
+      if (!NEXTHOP_ONE(nh))
 	nl_add_multipath(&r->h, rsize, nh, p->af, eattrs);
       else
       {
@@ -1597,99 +1565,56 @@ dest:
       bug("krt_capable inconsistent with nl_send_route");
     }
 
+done:
   /* Ignore missing for DELETE */
   return nl_exchange(&r->h, (op == NL_OP_DELETE));
 }
 
 static inline int
-nl_add_rte(struct krt_proto *p, rte *e)
+nl_allow_replace(struct krt_proto *p, rte *new)
 {
-  ea_list *ea = e->attrs;
-  int err = 0;
-
-  eattr *nhea = ea_find(ea, &ea_gen_nexthop);
-  struct nexthop_adata *nhad = nhea ? (struct nexthop_adata *) nhea->u.ptr : NULL;
-
-  if (krt_ecmp6(p) && nhad && NEXTHOP_IS_REACHABLE(nhad) && !NEXTHOP_ONE(nhad))
-  {
-    uint cnt = 0;
-    NEXTHOP_WALK(nh, nhad)
-    {
-      struct {
-	struct nexthop_adata nhad;
-	u32 labels[MPLS_MAX_LABEL_STACK];
-      } nhx;
-      memcpy(&nhx.nhad.nh, nh, NEXTHOP_SIZE(nh));
-      nhx.nhad.ad.length = (void *) NEXTHOP_NEXT(&nhx.nhad.nh) - (void *) nhx.nhad.ad.data;
-
-      if (!cnt++)
-      {
-	err = nl_send_route(p, e, NL_OP_ADD, RTD_UNICAST, &nhx.nhad);
-	if (err < 0)
-	  return err;
-      }
-      else
-	err += nl_send_route(p, e, NL_OP_APPEND, RTD_UNICAST, &nhx.nhad);
-    }
-
-    return err;
-  }
-
-  return nl_send_route(p, e, NL_OP_ADD,
-      NEXTHOP_IS_REACHABLE(nhad) ? RTD_UNICAST : nhad->dest, nhad);
-}
-
-static inline int
-nl_delete_rte(struct krt_proto *p, const rte *e)
-{
-  int err = 0;
+  /*
+   * We use NL_OP_REPLACE for IPv4, it has an issue with not checking for
+   * matching rtm_protocol, but that is OK when dedicated priority is used.
+   *
+   * For IPv6, the NL_OP_REPLACE is still broken even in Linux 4.19 LTS
+   * (although it seems to be fixed in Linux 5.10 LTS) for sequence:
+   *
+   * ip route add 2001:db8::/32 via fe80::1 dev eth0
+   * ip route replace 2001:db8::/32 dev eth0
+   *
+   * (it ends with two routes instead of replacing the first by the second one)
+   *
+   * Replacing with direct and special type (e.g. unreachable) routes does not
+   * work, but replacing with regular routes work reliably
+   */
 
-  /* For IPv6, we just repeatedly request DELETE until we get error */
-  do
-    err = nl_send_route(p, e, NL_OP_DELETE, RTD_NONE, NULL);
-  while (krt_ecmp6(p) && !err);
+  if (krt_ipv4(p))
+    return 1;
 
-  return err;
-}
+  eattr *nhea = ea_find(new->attrs, &ea_gen_nexthop);
+  struct nexthop_adata *nh = nhea ? (struct nexthop_adata *) nhea->u.ptr : NULL;
+  int dest = nhea_dest(nhea);
 
-static inline int
-nl_replace_rte(struct krt_proto *p, rte *e)
-{
-  eattr *nhea = ea_find(e->attrs, &ea_gen_nexthop);
-  struct nexthop_adata *nhad = nhea ? (struct nexthop_adata *) nhea->u.ptr : NULL;
-  return nl_send_route(p, e, NL_OP_REPLACE,
-      NEXTHOP_IS_REACHABLE(nhad) ? RTD_UNICAST : nhad->dest, nhad);
+  return (dest == RTD_UNICAST) && ipa_nonzero(nh->nh.gw);
 }
 
-
 void
 krt_replace_rte(struct krt_proto *p, const net_addr *n UNUSED, rte *new, const rte *old)
 {
   int err = 0;
 
-  /*
-   * We use NL_OP_REPLACE for IPv4, it has an issue with not checking for
-   * matching rtm_protocol, but that is OK when dedicated priority is used.
-   *
-   * We do not use NL_OP_REPLACE for IPv6, as it has broken semantics for ECMP
-   * and with some kernel versions ECMP replace crashes kernel. Would need more
-   * testing and checks for kernel versions.
-   *
-   * For IPv6, we use NL_OP_DELETE and then NL_OP_ADD. We also do not trust the
-   * old route value, so we do not try to optimize IPv6 ECMP reconfigurations.
-   */
-
-  if (krt_ipv4(p) && old && new)
+  if (old && new && nl_allow_replace(p, new))
   {
-    err = nl_replace_rte(p, new);
+    err = nl_send_route(p, new, NL_OP_REPLACE);
   }
   else
   {
     if (old)
-      nl_delete_rte(p, old);
+      nl_send_route(p, old, NL_OP_DELETE);
 
     if (new)
-      err = nl_add_rte(p, new);
+      err = nl_send_route(p, new, NL_OP_ADD);
   }
 
   if (new)
@@ -1701,68 +1626,6 @@ krt_replace_rte(struct krt_proto *p, const net_addr *n UNUSED, rte *new, const r
   }
 }
 
-static int
-nl_mergable_route(struct nl_parse_state *s, const net_addr *net, struct krt_proto *p, uint priority, uint krt_type, uint rtm_family)
-{
-  /* Route merging is used for IPv6 scans */
-  if (!s->scan || (rtm_family != AF_INET6))
-    return 0;
-
-  /* Saved and new route must have same network, proto/table, and priority */
-  if ((s->net != net) || (s->proto != p) || (s->krt_metric != priority))
-    return 0;
-
-  /* Both must be regular unicast routes */
-  if ((s->krt_type != RTN_UNICAST) || (krt_type != RTN_UNICAST))
-    return 0;
-
-  return 1;
-}
-
-static void
-nl_announce_route(struct nl_parse_state *s)
-{
-  rte e0 = {
-    .attrs = s->attrs,
-    .net = s->net,
-  };
-
-  EA_LOCAL_LIST(2) ea = {
-    .l = { .count = 2, .next = e0.attrs },
-    .a = {
-      EA_LITERAL_EMBEDDED(&ea_krt_source, 0, s->krt_proto),
-      EA_LITERAL_EMBEDDED(&ea_krt_metric, 0, s->krt_metric),
-    },
-  };
-
-  e0.attrs = &ea.l;
-
-  if (s->scan)
-    krt_got_route(s->proto, &e0, s->krt_src);
-  else
-    krt_got_route_async(s->proto, &e0, s->new, s->krt_src);
-
-  s->net = NULL;
-  s->attrs = NULL;
-  s->proto = NULL;
-  lp_flush(s->pool);
-}
-
-static inline void
-nl_parse_begin(struct nl_parse_state *s, int scan)
-{
-  memset(s, 0, sizeof (struct nl_parse_state));
-  s->pool = nl_linpool;
-  s->scan = scan;
-}
-
-static inline void
-nl_parse_end(struct nl_parse_state *s)
-{
-  if (s->net)
-    nl_announce_route(s);
-}
-
 
 #define SKIP0(ARG, ...) do { DBG("KRT: Ignoring route - " ARG, ##__VA_ARGS__); return; } while(0)
 #define SKIP(ARG, ...)  do { DBG("KRT: Ignoring route %N - " ARG, &dst, ##__VA_ARGS__); return; } while(0)
@@ -1898,11 +1761,10 @@ nl_parse_route(struct nl_parse_state *s, struct nlmsghdr *h)
 		      net6_prefix(&src), net6_pxlen(&src));
   }
 
-  if (s->net && !nl_mergable_route(s, net, p, priority, i->rtm_type, i->rtm_family))
-    nl_announce_route(s);
-
   ea_list *ra = NULL;
   ea_set_attr_u32(&ra, &ea_gen_source, 0, RTS_INHERIT);
+  ea_set_attr_u32(&ra, &ea_krt_source, 0, i->rtm_protocol);
+  ea_set_attr_u32(&ra, &ea_krt_metric, 0, priority);
 
   if (a[RTA_FLOW])
     s->rta_flow = rta_get_u32(a[RTA_FLOW]);
@@ -2012,6 +1874,9 @@ nl_parse_route(struct nl_parse_state *s, struct nlmsghdr *h)
       return;
     }
 
+  if (nhad.ad.length)
+    ea_set_attr(&ra, EA_LITERAL_DIRECT_ADATA(&ea_gen_nexthop, 0, &nhad.ad));
+
   if (i->rtm_scope != def_scope)
     ea_set_attr(&ra,
 	EA_LITERAL_EMBEDDED(&ea_krt_scope, 0, i->rtm_scope));
@@ -2044,56 +1909,27 @@ nl_parse_route(struct nl_parse_state *s, struct nlmsghdr *h)
 	      EA_LITERAL_EMBEDDED(&ea_krt_metrics[t], 0, metrics[t]));
     }
 
-  /*
-   * Ideally, now we would send the received route to the rest of kernel code.
-   * But IPv6 ECMP routes before 4.11 are sent as a sequence of routes, so we
-   * postpone it and merge next hops until the end of the sequence. Note that
-   * when doing merging of next hops, we expect the new route to be unipath.
-   * Otherwise, we ignore additional next hops in nexthop_insert().
-   */
+  rte e0 = {
+    .net = net,
+    .attrs = ra,
+  };
 
-  if (!s->net)
-  {
-    /* Store the new route */
-    s->net = lp_alloc(s->pool, net->length);
-    net_copy(s->net, net);
-
-    ea_set_attr_data(&ra, &ea_gen_nexthop, 0,
-	nhad.ad.data, nhad.ad.length);
-
-    s->attrs = ra;
-    s->proto = p;
-    s->new = new;
-    s->krt_src = krt_src;
-    s->krt_type = i->rtm_type;
-    s->krt_proto = i->rtm_protocol;
-    s->krt_metric = priority;
-  }
+  if (s->scan)
+    krt_got_route(p, &e0, krt_src);
   else
-  {
-    /* Merge next hops with the stored route */
-    eattr *nhea = ea_find(s->attrs, &ea_gen_nexthop);
-    struct nexthop_adata *nhad_old = nhea ? (struct nexthop_adata *) nhea->u.ptr : NULL;
-
-    if (nhad_old)
-      ea_set_attr(&s->attrs,
-	  EA_LITERAL_DIRECT_ADATA(&ea_gen_nexthop, 0, 
-	    &(nexthop_merge(nhad_old, &nhad.nhad,
-		KRT_CF->merge_paths, s->pool)->ad)
-	  ));
-    else
-      ea_set_attr_data(&s->attrs, &ea_gen_nexthop, 0,
-	  nhad.ad.data, nhad.ad.length);
-  }
+    krt_got_route_async(p, &e0, new, krt_src);
+
+  lp_flush(s->pool);
 }
 
 void
 krt_do_scan(struct krt_proto *p)
 {
-  struct nlmsghdr *h;
-  struct nl_parse_state s;
-
-  nl_parse_begin(&s, 1);
+  struct nl_parse_state s = {
+    .proto = p,
+    .pool = nl_linpool,
+    .scan = 1,
+  };
 
   /* Table-specific scan or shared scan */
   if (p)
@@ -2101,6 +1937,7 @@ krt_do_scan(struct krt_proto *p)
   else
     nl_request_dump_route(AF_UNSPEC, 0);
 
+  struct nlmsghdr *h;
   while (h = nl_get_scan())
   {
     if (h->nlmsg_type == RTM_NEWROUTE || h->nlmsg_type == RTM_DELROUTE)
@@ -2108,8 +1945,6 @@ krt_do_scan(struct krt_proto *p)
     else
       log(L_DEBUG "nl_scan_fire: Unknown packet received (type=%d)", h->nlmsg_type);
   }
-
-  nl_parse_end(&s);
 }
 
 /*
@@ -2124,16 +1959,18 @@ static struct config *nl_last_config;	/* For tracking changes to nl_async_bufsiz
 static void
 nl_async_msg(struct nlmsghdr *h)
 {
-  struct nl_parse_state s;
+  struct nl_parse_state s = {
+    .proto = NULL,
+    .pool = nl_linpool,
+    .scan = 0,
+  };
 
   switch (h->nlmsg_type)
     {
     case RTM_NEWROUTE:
     case RTM_DELROUTE:
       DBG("KRT: Received async route notification (%d)\n", h->nlmsg_type);
-      nl_parse_begin(&s, 0);
       nl_parse_route(&s, h);
-      nl_parse_end(&s);
       break;
     case RTM_NEWLINK:
     case RTM_DELLINK: