Clean up multi-thread barrier-sync hold-down timer

Main thread: don't bother with the barrier sync hold-down timer if
none of the worker threads are busy.

Worker threads: avoid epoll_pwait (10ms timeout) when the
control-plane has been active in the last half-second.

Cherry-pick a recent dangling reference fix: pool_elt_at_index after
e.g. rx callback is required, in case the unix file pool expands.

DO NOT cherry-pick the patch into master!

Change-Id: I130dc72fd95c8a22716fa2f0f98ce35cdccc52d7
Signed-off-by: Dave Barach <dave@barachs.net>

4 files changed, 130 insertions, 16 deletions

diff --git a/src/vlib/main.h b/src/vlib/main.h
index 474756bebd6..8c2719cee52 100644
--- a/src/vlib/main.h
+++ b/src/vlib/main.h
@@ -63,6 +63,9 @@ typedef struct vlib_main_t
   CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
   /* Instruction level timing state. */
   clib_time_t clib_time;
+  /* Offset from main thread time */
+  f64 time_offset;
+  f64 time_last_barrier_release;
 
   /* Time stamp of last node dispatch. */
   u64 cpu_time_last_node_dispatch;
@@ -228,7 +231,7 @@ void vlib_worker_loop (vlib_main_t * vm);
 always_inline f64
 vlib_time_now (vlib_main_t * vm)
 {
-  return clib_time_now (&vm->clib_time);
+  return clib_time_now (&vm->clib_time) + vm->time_offset;
 }
 
 always_inline f64
diff --git a/src/vlib/threads.c b/src/vlib/threads.c
index af60bc6def6..56f9d790ee6 100644
--- a/src/vlib/threads.c
+++ b/src/vlib/threads.c
@@ -1367,7 +1367,9 @@ vlib_worker_thread_barrier_sync_int (vlib_main_t * vm, const char *func_name)
   f64 t_entry;
   f64 t_open;
   f64 t_closed;
+  f64 max_vector_rate, this_vector_rate;
   u32 count;
+  int i;
 
   if (vec_len (vlib_mains) < 2)
     return;
@@ -1388,23 +1390,44 @@ vlib_worker_thread_barrier_sync_int (vlib_main_t * vm, const char *func_name)
       return;
     }
 
+  /*
+   * Need data to decide if we're working hard enough to honor
+   * the barrier hold-down timer.
+   */
+  max_vector_rate = 0.0;
+  for (i = 1; i < vec_len (vlib_mains); i++)
+    {
+      this_vector_rate =
+	vlib_last_vectors_per_main_loop_as_f64 (vlib_mains[i]);
+      if (max_vector_rate < this_vector_rate)
+	max_vector_rate = this_vector_rate;
+    }
+
   vlib_worker_threads[0].barrier_sync_count++;
 
   /* Enforce minimum barrier open time to minimize packet loss */
   ASSERT (vm->barrier_no_close_before <= (now + BARRIER_MINIMUM_OPEN_LIMIT));
 
-  while (1)
+  /*
+   * If any worker thread seems busy, which we define
+   * as a vector rate above 10, we enforce the barrier hold-down timer
+   */
+  if (max_vector_rate > 10.0)
     {
-      now = vlib_time_now (vm);
-      /* Barrier hold-down timer expired? */
-      if (now >= vm->barrier_no_close_before)
-	break;
-      if ((vm->barrier_no_close_before - now)
-	  > (2.0 * BARRIER_MINIMUM_OPEN_LIMIT))
+      while (1)
 	{
-	  clib_warning ("clock change: would have waited for %.4f seconds",
-			(vm->barrier_no_close_before - now));
-	  break;
+	  now = vlib_time_now (vm);
+	  /* Barrier hold-down timer expired? */
+	  if (now >= vm->barrier_no_close_before)
+	    break;
+	  if ((vm->barrier_no_close_before - now)
+	      > (2.0 * BARRIER_MINIMUM_OPEN_LIMIT))
+	    {
+	      clib_warning
+		("clock change: would have waited for %.4f seconds",
+		 (vm->barrier_no_close_before - now));
+	      break;
+	    }
 	}
     }
   /* Record time of closure */
@@ -1491,6 +1514,14 @@ vlib_worker_thread_barrier_release (vlib_main_t * vm)
 
   deadline = now + BARRIER_SYNC_TIMEOUT;
 
+  /*
+   * Note when we let go of the barrier.
+   * Workers can use this to derive a reasonably accurate
+   * time offset. See vlib_time_now(...)
+   */
+  vm->time_last_barrier_release = vlib_time_now (vm);
+  CLIB_MEMORY_STORE_BARRIER ();
+
   *vlib_worker_threads->wait_at_barrier = 0;
 
   while (*vlib_worker_threads->workers_at_barrier > 0)
@@ -1789,6 +1820,45 @@ threads_init (vlib_main_t * vm)
 
 VLIB_INIT_FUNCTION (threads_init);
 
+
+static clib_error_t *
+show_clock_command_fn (vlib_main_t * vm,
+		       unformat_input_t * input, vlib_cli_command_t * cmd)
+{
+  int i;
+  f64 now;
+
+  now = vlib_time_now (vm);
+
+  vlib_cli_output (vm, "Time now %.9f", now);
+
+  if (vec_len (vlib_mains) == 1)
+    return 0;
+
+  vlib_cli_output (vm, "Time last barrier release %.9f",
+		   vm->time_last_barrier_release);
+
+  for (i = 1; i < vec_len (vlib_mains); i++)
+    {
+      if (vlib_mains[i] == 0)
+	continue;
+      vlib_cli_output (vm, "Thread %d offset %.9f error %.9f", i,
+		       vlib_mains[i]->time_offset,
+		       vm->time_last_barrier_release -
+		       vlib_mains[i]->time_last_barrier_release);
+    }
+  return 0;
+}
+
+/* *INDENT-OFF* */
+VLIB_CLI_COMMAND (f_command, static) =
+{
+  .path = "show clock",
+  .short_help = "show clock",
+  .function = show_clock_command_fn,
+};
+/* *INDENT-ON* */
+
 /*
  * fd.io coding-style-patch-verification: ON
  *
diff --git a/src/vlib/threads.h b/src/vlib/threads.h
index 0ceb21e1218..5a295e03319 100644
--- a/src/vlib/threads.h
+++ b/src/vlib/threads.h
@@ -423,6 +423,21 @@ vlib_worker_thread_barrier_check (void)
 	}
       while (*vlib_worker_threads->wait_at_barrier)
 	;
+
+      /*
+       * Recompute the offset from thread-0 time.
+       * Note that vlib_time_now adds vm->time_offset, so
+       * clear it first. Save the resulting idea of "now", to
+       * see how well we're doing. See show_clock_command_fn(...)
+       */
+      {
+	f64 now;
+	vm->time_offset = 0.0;
+	now = vlib_time_now (vm);
+	vm->time_offset = vlib_global_main.time_last_barrier_release - now;
+	vm->time_last_barrier_release = vlib_time_now (vm);
+      }
+
       if (CLIB_DEBUG > 0)
 	vm->parked_at_barrier = 0;
       clib_atomic_fetch_add (vlib_worker_threads->workers_at_barrier, -1);
diff --git a/src/vlib/unix/input.c b/src/vlib/unix/input.c
index 370e5dddaf6..76337f65865 100644
--- a/src/vlib/unix/input.c
+++ b/src/vlib/unix/input.c
@@ -145,9 +145,13 @@ linux_epoll_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
     vlib_node_main_t *nm = &vm->node_main;
     u32 ticks_until_expiration;
     f64 timeout;
+    f64 now;
     int timeout_ms = 0, max_timeout_ms = 10;
     f64 vector_rate = vlib_last_vectors_per_main_loop (vm);
 
+    if (is_main == 0)
+      now = vlib_time_now (vm);
+
     /*
      * If we've been asked for a fixed-sleep between main loop polls,
      * do so right away.
@@ -194,8 +198,9 @@ linux_epoll_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	  }
 	node->input_main_loops_per_call = 0;
       }
-    else if (is_main == 0 && vector_rate < 2 &&
-	     nm->input_node_counts_by_state[VLIB_NODE_STATE_POLLING] == 0)
+    else if (is_main == 0 && vector_rate < 2
+	     && (vlib_global_main.time_last_barrier_release + 0.5 < now)
+	     && nm->input_node_counts_by_state[VLIB_NODE_STATE_POLLING] == 0)
       {
 	timeout = 10e-3;
 	timeout_ms = max_timeout_ms;
@@ -223,12 +228,32 @@ linux_epoll_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 				      em->epoll_events,
 				      vec_len (em->epoll_events), timeout_ms);
 	  }
+
       }
     else
       {
+	/*
+	 * Worker thread, no epoll fd's, sleep for 100us at a time
+	 * and check for a barrier sync request
+	 */
 	if (timeout_ms)
-	  usleep (timeout_ms * 1000);
-	return 0;
+	  {
+	    struct timespec ts, tsrem;
+	    f64 limit = now + (f64) timeout_ms * 1e-3;
+
+	    while (vlib_time_now (vm) < limit)
+	      {
+		/* Sleep for 100us at a time */
+		ts.tv_sec = 0;
+		ts.tv_nsec = 1000 * 100;
+
+		while (nanosleep (&ts, &tsrem) < 0)
+		  ts = tsrem;
+		if (*vlib_worker_threads->wait_at_barrier)
+		  goto done;
+	      }
+	  }
+	goto done;
       }
   }
 
@@ -238,7 +263,7 @@ linux_epoll_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	vlib_panic_with_error (vm, clib_error_return_unix (0, "epoll_wait"));
 
       /* non fatal error (e.g. EINTR). */
-      return 0;
+      goto done;
     }
 
   em->epoll_waits += 1;
@@ -318,6 +343,7 @@ linux_epoll_input_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
 	}
     }
 
+done:
   return 0;
 }