From 7be3af7fa662958782d2e23989d79cc2c652b6bf Mon Sep 17 00:00:00 2001
From: "Ondrej Zajicek (work)" <santiago@crfreenet.org>
Date: Fri, 12 Mar 2021 15:35:56 +0100
Subject: [PATCH] Rate-limit scheduling of work-events

In general, events are code handling some some condition, which is
scheduled when such condition happened and executed independently from
I/O loop. Work-events are a subgroup of events that are scheduled
repeatedly until some (often significant) work is done (e.g. feeding
routes to protocol). All scheduled events are executed during each
I/O loop iteration.

Separate work-events from regular events to a separate queue and
rate limit their execution to a fixed number per I/O loop iteration.
That should prevent excess latency when many work-events are
scheduled at one time (e.g. simultaneous reload of many BGP sessions).
---
 lib/event.c      | 56 +++++++++++++++++++++++++++++++++++++++++++++++-
 lib/event.h      |  3 +++
 nest/proto.c     | 10 ++++-----
 sysdep/unix/io.c |  3 +++
 4 files changed, 66 insertions(+), 6 deletions(-)

diff --git a/lib/event.c b/lib/event.c
index c33e0ffc..273447e0 100644
--- a/lib/event.c
+++ b/lib/event.c
@@ -23,6 +23,7 @@
 #include "lib/event.h"
 
 event_list global_event_list;
+event_list global_work_list;
 
 inline void
 ev_postpone(event *e)
@@ -114,6 +115,22 @@ ev_schedule(event *e)
   ev_enqueue(&global_event_list, e);
 }
 
+/**
+ * ev_schedule_work - schedule a work-event.
+ * @e: an event
+ *
+ * This function schedules an event by enqueueing it to a system-wide work-event
+ * list which is run by the platform dependent code whenever appropriate. This
+ * is designated for work-events instead of regular events. They are executed
+ * less often in order to not clog I/O loop.
+ */
+void
+ev_schedule_work(event *e)
+{
+  if (!ev_active(e))
+    add_tail(&global_work_list, &e->n);
+}
+
 void io_log_event(void *hook, void *data);
 
 /**
@@ -136,10 +153,47 @@ ev_run_list(event_list *l)
       event *e = SKIP_BACK(event, n, n);
 
       /* This is ugly hack, we want to log just events executed from the main I/O loop */
-      if (l == &global_event_list)
+      if ((l == &global_event_list) || (l == &global_work_list))
 	io_log_event(e->hook, e->data);
 
       ev_run(e);
     }
+
+  return !EMPTY_LIST(*l);
+}
+
+int
+ev_run_list_limited(event_list *l, uint limit)
+{
+  node *n;
+  list tmp_list;
+
+  init_list(&tmp_list);
+  add_tail_list(&tmp_list, l);
+  init_list(l);
+
+  WALK_LIST_FIRST(n, tmp_list)
+    {
+      event *e = SKIP_BACK(event, n, n);
+
+      if (!limit)
+	break;
+
+      /* This is ugly hack, we want to log just events executed from the main I/O loop */
+      if ((l == &global_event_list) || (l == &global_work_list))
+	io_log_event(e->hook, e->data);
+
+      ev_run(e);
+      limit--;
+    }
+
+  if (!EMPTY_LIST(tmp_list))
+  {
+    /* Attach new items after the unprocessed old items */
+    add_tail_list(&tmp_list, l);
+    init_list(l);
+    add_tail_list(l, &tmp_list);
+  }
+
   return !EMPTY_LIST(*l);
 }
diff --git a/lib/event.h b/lib/event.h
index 03735c3f..5f3b78d8 100644
--- a/lib/event.h
+++ b/lib/event.h
@@ -21,14 +21,17 @@ typedef struct event {
 typedef list event_list;
 
 extern event_list global_event_list;
+extern event_list global_work_list;
 
 event *ev_new(pool *);
 void ev_run(event *);
 #define ev_init_list(el) init_list(el)
 void ev_enqueue(event_list *, event *);
 void ev_schedule(event *);
+void ev_schedule_work(event *);
 void ev_postpone(event *);
 int ev_run_list(event_list *);
+int ev_run_list_limited(event_list *, uint);
 
 static inline int
 ev_active(event *e)
diff --git a/nest/proto.c b/nest/proto.c
index 165125d8..b3c6fa47 100644
--- a/nest/proto.c
+++ b/nest/proto.c
@@ -252,7 +252,7 @@ channel_schedule_feed(struct channel *c, int initial)
   c->export_state = ES_FEEDING;
   c->refeeding = !initial;
 
-  ev_schedule(c->feed_event);
+  ev_schedule_work(c->feed_event);
 }
 
 static void
@@ -275,7 +275,7 @@ channel_feed_loop(void *ptr)
   // DBG("Feeding protocol %s continued\n", p->name);
   if (!rt_feed_channel(c))
   {
-    ev_schedule(c->feed_event);
+    ev_schedule_work(c->feed_event);
     return;
   }
 
@@ -291,7 +291,7 @@ channel_feed_loop(void *ptr)
 
     /* Continue in feed - it will process routing table again from beginning */
     c->refeed_count = 0;
-    ev_schedule(c->feed_event);
+    ev_schedule_work(c->feed_event);
     return;
   }
 
@@ -471,7 +471,7 @@ channel_schedule_reload(struct channel *c)
   ASSERT(c->channel_state == CS_UP);
 
   rt_reload_channel_abort(c);
-  ev_schedule(c->reload_event);
+  ev_schedule_work(c->reload_event);
 }
 
 static void
@@ -485,7 +485,7 @@ channel_reload_loop(void *ptr)
 
   if (!rt_reload_channel(c))
   {
-    ev_schedule(c->reload_event);
+    ev_schedule_work(c->reload_event);
     return;
   }
 
diff --git a/sysdep/unix/io.c b/sysdep/unix/io.c
index 9d54a2c3..3d67d0a7 100644
--- a/sysdep/unix/io.c
+++ b/sysdep/unix/io.c
@@ -2161,6 +2161,7 @@ io_init(void)
 {
   init_list(&sock_list);
   init_list(&global_event_list);
+  init_list(&global_work_list);
   krt_io_init();
   // XXX init_times();
   // XXX update_times();
@@ -2172,6 +2173,7 @@ io_init(void)
 
 static int short_loops = 0;
 #define SHORT_LOOP_MAX 10
+#define WORK_EVENTS_MAX 10
 
 void
 io_loop(void)
@@ -2189,6 +2191,7 @@ io_loop(void)
     {
       times_update(&main_timeloop);
       events = ev_run_list(&global_event_list);
+      events = ev_run_list_limited(&global_work_list, WORK_EVENTS_MAX) || events;
       timers_fire(&main_timeloop);
       io_close_event();