Lockless hostentry resolution

Now the hostentry doesn't need to lock table, instead it tracks the
hostentry version and retries if the hostentry changed while updating.

nest/route.h

@@ -502,9 +502,10 @@ struct hostentry {
   struct hostentry *next;		/* Next in hash chain */
   unsigned hash_key;			/* Hash key */
   u32 igp_metric;			/* Chosen route IGP metric */
+  _Atomic u32 version;			/* Bumped on update */
+  byte nexthop_linkable;		/* Nexthop list is completely non-device */
   ea_list *src;				/* Source attributes */
   struct lfuc uc;			/* Use count */
-  byte nexthop_linkable;		/* Nexthop list is completely non-device */
 };
 
 struct hostcache {
@@ -669,6 +670,7 @@ struct rt_show_data_rtable * rt_show_add_table(struct rt_show_data *d, rtable *t
 
 /* Host entry: Resolve hook for recursive nexthops */
 extern struct ea_class ea_gen_hostentry;
+extern struct ea_class ea_gen_hostentry_version;
 struct hostentry_adata {
   adata ad;
   struct hostentry *he;

nest/rt-attr.c

@@ -161,6 +161,13 @@ struct ea_class ea_gen_hostentry = {
   .freed = ea_gen_hostentry_freed,
 };
 
+struct ea_class ea_gen_hostentry_version = {
+  .name = "hostentry version",
+  .type = T_INT,
+  .readonly = 1,
+  .hidden = 1,
+};
+
 const char * rta_dest_names[RTD_MAX] = {
   [RTD_NONE]		= "",
   [RTD_UNICAST]		= "unicast",
@@ -1821,6 +1828,7 @@ rta_init(void)
   /* These attributes are required to be first for nice "show route" output */
   ea_register_init(&ea_gen_nexthop);
   ea_register_init(&ea_gen_hostentry);
+  ea_register_init(&ea_gen_hostentry_version);
 
   /* Other generic route attributes */
   ea_register_init(&ea_gen_preference);

nest/rt-table.c

@@ -3001,18 +3001,32 @@ ea_set_hostentry(ea_list **to, rtable *dep, rtable *src, ip_addr gw, ip_addr ll,
 
 
 static void
-rta_apply_hostentry(struct rtable_private *tab UNUSED, ea_list **to, struct hostentry_adata *head)
+rta_apply_hostentry(ea_list **to, struct hostentry_adata *head)
 {
-  struct hostentry *he = head->he;
   u32 *labels = head->labels;
   u32 lnum = (u32 *) (head->ad.data + head->ad.length) - labels;
+  struct hostentry *he = head->he;
+
+  rcu_read_lock();
+  u32 version = atomic_load_explicit(&he->version, memory_order_acquire);
+
+  while (1)
+  {
+    if (version & 1)
+    {
+      rcu_read_unlock();
+      birdloop_yield();
+      rcu_read_lock();
+      version = atomic_load_explicit(&he->version, memory_order_acquire);
+      continue;
+    }
 
     ea_set_attr_u32(to, &ea_gen_igp_metric, 0, he->igp_metric);
 
     if (!he->src)
     {
       ea_set_dest(to, 0, RTD_UNREACHABLE);
-    return;
+      break;
     }
 
     eattr *he_nh_ea = ea_find(he->src, &ea_gen_nexthop);
@@ -3023,9 +3037,10 @@ rta_apply_hostentry(struct rtable_private *tab UNUSED, ea_list **to, struct host
 
     if ((idest != RTD_UNICAST) ||
 	!lnum && he->nexthop_linkable)
-  { /* Just link the nexthop chain, no label append happens. */
+    {
+      /* Just link the nexthop chain, no label append happens. */
       ea_copy_attr(to, he->src, &ea_gen_nexthop);
-    return;
+      break;
     }
 
     uint total_size = OFFSETOF(struct nexthop_adata, nh);
@@ -3052,7 +3067,7 @@ rta_apply_hostentry(struct rtable_private *tab UNUSED, ea_list **to, struct host
       };
 
       ea_set_attr_data(to, &ea_gen_nexthop, 0, &nha.ad.data, nha.ad.length);
-    return;
+      break;
     }
 
     struct nexthop_adata *new = (struct nexthop_adata *) tmp_alloc_adata(total_size);
@@ -3087,53 +3102,42 @@ rta_apply_hostentry(struct rtable_private *tab UNUSED, ea_list **to, struct host
     new->ad.length = (void *) dest - (void *) new->ad.data;
     ea_set_attr(to, EA_LITERAL_DIRECT_ADATA(
 	  &ea_gen_nexthop, 0, &new->ad));
+
+    /* Has the HE version changed? */
+    u32 end_version = atomic_load_explicit(&he->version, memory_order_acquire);
+
+    /* Stayed stable, we can finalize the route */
+    if (end_version == version)
+      break;
+
+    /* No, retry once again */
+    version = end_version;
   }
 
-static inline struct hostentry_adata *
+  rcu_read_unlock();
-rta_next_hop_outdated(ea_list *a)
-{
-  /* First retrieve the hostentry */
-  eattr *heea = ea_find(a, &ea_gen_hostentry);
-  if (!heea)
-    return NULL;
 
-  struct hostentry_adata *head = (struct hostentry_adata *) heea->u.ptr;
+  ea_set_attr_u32(to, &ea_gen_hostentry_version, 0, version);
-
-  /* If no nexthop is present, we have to create one */
-  eattr *a_nh_ea = ea_find(a, &ea_gen_nexthop);
-  if (!a_nh_ea)
-    return head;
-
-  struct nexthop_adata *nhad = (struct nexthop_adata *) a_nh_ea->u.ptr;
-
-  /* Shortcut for unresolvable hostentry */
-  if (!head->he->src)
-    return NEXTHOP_IS_REACHABLE(nhad) ? head : NULL;
-
-  /* Comparing our nexthop with the hostentry nexthop */
-  eattr *he_nh_ea = ea_find(head->he->src, &ea_gen_nexthop);
-
-  return (
-      (ea_get_int(a, &ea_gen_igp_metric, IGP_METRIC_UNKNOWN) != head->he->igp_metric) ||
-      (!head->he->nexthop_linkable) ||
-      (!he_nh_ea != !a_nh_ea) ||
-      (he_nh_ea && a_nh_ea && !adata_same(he_nh_ea->u.ptr, a_nh_ea->u.ptr)))
-    ? head : NULL;
 }
 
 static inline int
 rt_next_hop_update_rte(const rte *old, rte *new)
 {
-  struct hostentry_adata *head = rta_next_hop_outdated(old->attrs);
+  eattr *hev = ea_find(old->attrs, &ea_gen_hostentry_version);
-  if (!head)
+  if (!hev)
+    return 0;
+  u32 last_version = hev->u.data;
+
+  eattr *heea = ea_find(old->attrs, &ea_gen_hostentry);
+  ASSERT_DIE(heea);
+  struct hostentry_adata *head = (struct hostentry_adata *) heea->u.ptr;
+
+  u32 current_version = atomic_load_explicit(&head->he->version, memory_order_acquire);
+  if (current_version == last_version)
     return 0;
 
-  /* Get the state of the route just before nexthop was resolved */
   *new = *old;
   new->attrs = ea_strip_to(new->attrs, BIT32_ALL(EALS_PREIMPORT, EALS_FILTERED));
-
+  rta_apply_hostentry(&new->attrs, head);
-  RT_LOCKED(head->he->owner, tab)
-    rta_apply_hostentry(tab, &new->attrs, head);
   return 1;
 }
 
@@ -3144,10 +3148,7 @@ rt_next_hop_resolve_rte(rte *r)
   if (!heea)
     return;
 
-  struct hostentry_adata *head = (struct hostentry_adata *) heea->u.ptr;
+  rta_apply_hostentry(&r->attrs, (struct hostentry_adata *) heea->u.ptr);
-
-  RT_LOCKED(head->he->owner, tab)
-    rta_apply_hostentry(tab, &r->attrs, head);
 }
 
 #ifdef CONFIG_BGP
@@ -4374,6 +4375,9 @@ rt_update_hostentry(struct rtable_private *tab, struct hostentry *he)
   int direct = 0;
   int pxlen = 0;
 
+  /* Signalize work in progress */
+  ASSERT_DIE((atomic_fetch_add_explicit(&he->version, 1, memory_order_acq_rel) & 1) == 0);
+
   /* Reset the hostentry */
   he->src = NULL;
   he->nexthop_linkable = 0;
@@ -4426,6 +4430,10 @@ rt_update_hostentry(struct rtable_private *tab, struct hostentry *he)
     }
 
 done:
+  /* Signalize work done and wait for readers */
+  ASSERT_DIE((atomic_fetch_add_explicit(&he->version, 1, memory_order_acq_rel) & 1) == 1);
+  synchronize_rcu();
+
   /* Add a prefix range to the trie */
   trie_add_prefix(tab->hostcache->trie, &he_addr, pxlen, he_addr.pxlen);