Nest: Improve GC strategy for rtables

Use timer (configurable as 'gc period') to schedule routing table
GC/pruning to ensure that prune is done on time but not too often.

Randomize GC timers to avoid concentration of GC events from different
tables in one loop cycle.

Fix a bug that caused minimum inter-GC interval be 5 us instead of 5 s.

Make default 'gc period' adaptive based on number of routing tables,
from 10 s for small setups to 600 s for large ones.

In marge multi-table RS setup, the patch improved time of flushing
a downed peer from 20-30 min to <2 min and removed 40s latencies.

conf/conf.c

@@ -140,6 +140,7 @@ config_parse(struct config *c)
   protos_preconfig(c);
   rt_preconfig(c);
   cf_parse();
+  rt_postconfig(c);
 
   if (EMPTY_LIST(c->protos))
     cf_error("No protocol is specified in the config file");

doc/bird.sgml

@@ -684,6 +684,21 @@ to set options.
 	limit to the settle time from the initial ROA table change even if
 	there are consecutive updates gradually renewing the settle time.
 	Default: 20 s.
+
+	<tag><label id="rtable-gc-threshold">gc threshold <m/number/</tag>
+	Specify a minimum amount of removed networks that triggers a garbage
+	collection (GC) cycle. Default: 1000.
+
+	<tag><label id="rtable-gc-period">gc period <m/time/</tag>
+	Specify a period of time between consecutive GC cycles. When there is a
+	significant amount of route withdraws, GC cycles are executed repeatedly
+	with given period time (with some random factor). When there is just
+	small amount of changes, GC cycles are not executed. In extensive route
+	server setups, running GC on hundreds of full BGP routing tables can
+	take significant amount of time, therefore they should use higher GC
+	periods. Default: adaptive, based on number of routing tables in the
+	configuration. From 10 s (with <= 25 routing tables) up to 600 s (with
+	>= 1500 routing tables).
 </descrip>
 
 

nest/config.Y

@@ -125,7 +125,7 @@ CF_KEYWORDS(TIMEFORMAT, ISO, SHORT, LONG, ROUTE, PROTOCOL, BASE, LOG, S, MS, US)
 CF_KEYWORDS(GRACEFUL, RESTART, WAIT, MAX, FLUSH, AS)
 CF_KEYWORDS(MIN, IDLE, RX, TX, INTERVAL, MULTIPLIER, PASSIVE)
 CF_KEYWORDS(CHECK, LINK)
-CF_KEYWORDS(SORTED, TRIE, MIN, MAX, SETTLE, TIME)
+CF_KEYWORDS(SORTED, TRIE, MIN, MAX, SETTLE, TIME, GC, THRESHOLD, PERIOD)
 
 /* For r_args_channel */
 CF_KEYWORDS(IPV4, IPV4_MC, IPV4_MPLS, IPV6, IPV6_MC, IPV6_MPLS, IPV6_SADR, VPN4, VPN4_MC, VPN4_MPLS, VPN6, VPN6_MC, VPN6_MPLS, ROA4, ROA6, FLOW4, FLOW6, MPLS, PRI, SEC)
@@ -229,6 +229,8 @@ table_opt:
    }
  | MIN SETTLE TIME expr_us { this_table->min_settle_time = $4; }
  | MAX SETTLE TIME expr_us { this_table->max_settle_time = $4; }
+ | GC THRESHOLD expr { this_table->gc_threshold = $3; }
+ | GC PERIOD expr_us { this_table->gc_period = (uint) $3; if ($3 > 3600 S_) cf_error("GC period must be at most 3600 s"); }
  ;
 
 table_opts:

nest/route.h

@@ -148,8 +148,8 @@ struct rtable_config {
   struct rtable *table;
   struct proto_config *krt_attached;	/* Kernel syncer attached to this table */
   uint addr_type;			/* Type of address data stored in table (NET_*) */
-  int gc_max_ops;			/* Maximum number of operations before GC is run */
+  uint gc_threshold;			/* Maximum number of operations before GC is run */
-  int gc_min_time;			/* Minimum time between two consecutive GC runs */
+  uint gc_period;			/* Approximate time between two consecutive GC runs */
   byte sorted;				/* Routes of network are sorted according to rte_better() */
   byte internal;			/* Internal table of a protocol */
   byte trie_used;			/* Rtable has attached trie */
@@ -180,10 +180,11 @@ typedef struct rtable {
 					 * obstacle from this routing table.
 					 */
   struct event *rt_event;		/* Routing table event */
+  struct timer *prune_timer;		/* Timer for periodic pruning / GC */
   btime last_rt_change;			/* Last time when route changed */
   btime base_settle_time;		/* Start time of rtable settling interval */
   btime gc_time;			/* Time of last GC */
-  int gc_counter;			/* Number of operations since last GC */
+  uint gc_counter;			/* Number of operations since last GC */
   byte prune_state;			/* Table prune state, 1 -> scheduled, 2-> running */
   byte prune_trie;			/* Prune prefix trie during next table prune */
   byte hcu_scheduled;			/* Hostcache update is scheduled */
@@ -332,6 +333,7 @@ struct config;
 
 void rt_init(void);
 void rt_preconfig(struct config *);
+void rt_postconfig(struct config *);
 void rt_commit(struct config *new, struct config *old);
 void rt_lock_table(rtable *);
 void rt_unlock_table(rtable *);

nest/rt-table.c

@@ -124,6 +124,7 @@ static void rt_next_hop_update(rtable *tab);
 static inline void rt_prune_table(rtable *tab);
 static inline void rt_schedule_notify(rtable *tab);
 static void rt_flowspec_notify(rtable *tab, net *net);
+static void rt_kick_prune_timer(rtable *tab);
 
 
 static void
@@ -1641,9 +1642,8 @@ rte_recalculate(struct channel *c, net *net, rte *new, struct rte_src *src)
   rte_announce(table, RA_UNDEF, net, new, old, net->routes, old_best);
 
   if (!net->routes &&
-      (table->gc_counter++ >= table->config->gc_max_ops) &&
+      (table->gc_counter++ >= table->config->gc_threshold))
-      (table->gc_time + table->config->gc_min_time <= current_time()))
+    rt_kick_prune_timer(table);
-    rt_schedule_prune(table);
 
   if (old_ok && p->rte_remove)
     p->rte_remove(net, old);
@@ -2098,6 +2098,29 @@ rt_event(void *ptr)
 }
 
 
+static void
+rt_prune_timer(timer *t)
+{
+  rtable *tab = t->data;
+
+  if (tab->gc_counter >= tab->config->gc_threshold)
+    rt_schedule_prune(tab);
+}
+
+static void
+rt_kick_prune_timer(rtable *tab)
+{
+  /* Return if prune is already scheduled */
+  if (tm_active(tab->prune_timer) || (tab->prune_state & 1))
+    return;
+
+  /* Randomize GC period to +/- 50% */
+  btime gc_period = tab->config->gc_period;
+  gc_period = (gc_period / 2) + (random_u32() % (uint) gc_period);
+  tm_start(tab->prune_timer, gc_period);
+}
+
+
 static inline btime
 rt_settled_time(rtable *tab)
 {
@@ -2333,6 +2356,7 @@ rt_setup(pool *pp, struct rtable_config *cf)
     hmap_set(&t->id_map, 0);
 
     t->rt_event = ev_new_init(p, rt_event, t);
+    t->prune_timer = tm_new_init(p, rt_prune_timer, t, 0, 0);
     t->last_rt_change = t->gc_time = current_time();
 
     if (rt_is_flow(t))
@@ -2403,6 +2427,9 @@ rt_prune_table(rtable *tab)
     FIB_ITERATE_INIT(fit, &tab->fib);
     tab->prune_state = 2;
 
+    tab->gc_counter = 0;
+    tab->gc_time = current_time();
+
     if (tab->prune_trie)
     {
       /* Init prefix trie pruning */
@@ -2462,9 +2489,6 @@ again:
   fib_check(&tab->fib);
 #endif
 
-  tab->gc_counter = 0;
-  tab->gc_time = current_time();
-
   /* state change 2->0, 3->1 */
   tab->prune_state &= 1;
 
@@ -2591,6 +2615,20 @@ rt_preconfig(struct config *c)
   rt_new_table(cf_get_symbol("master6"), NET_IP6);
 }
 
+void
+rt_postconfig(struct config *c)
+{
+  uint num_tables = list_length(&c->tables);
+  btime def_gc_period = 400 MS * num_tables;
+  def_gc_period = MAX(def_gc_period, 10 S);
+  def_gc_period = MIN(def_gc_period, 600 S);
+
+  struct rtable_config *rc;
+  WALK_LIST(rc, c->tables)
+    if (rc->gc_period == (uint) -1)
+      rc->gc_period = (uint) def_gc_period;
+}
+
 
 /*
  * Some functions for handing internal next hop updates
@@ -2999,8 +3037,8 @@ rt_new_table(struct symbol *s, uint addr_type)
   cf_define_symbol(s, SYM_TABLE, table, c);
   c->name = s->name;
   c->addr_type = addr_type;
-  c->gc_max_ops = 1000;
+  c->gc_threshold = 1000;
-  c->gc_min_time = 5;
+  c->gc_period = (uint) -1;	/* set in rt_postconfig() */
   c->min_settle_time = 1 S;
   c->max_settle_time = 20 S;