sort worker-thread init functions in advance

Otherwise, all N worker threads try to sort the list at the same time:
a good way to have a bad day.

This approach performs *far* better than maintaing order by adding a
spin-lock. By direct measurement w/ elog + g2: 11 threads execute the
per-thread init function list in 22us, vs. 50ms with a CLIB_PAUSE()
enabled spin-lock.

Change-Id: I1745f2a213c0561260139a60114dcb981e0c64e5
Signed-off-by: Dave Barach <dave@barachs.net>

1 files changed, 11 insertions, 0 deletions

diff --git a/src/vlib/main.c b/src/vlib/main.c
index b6006e8bab0..43400f8594c 100644
--- a/src/vlib/main.c
+++ b/src/vlib/main.c
@@ -1721,6 +1721,14 @@ vlib_main_or_worker_loop (vlib_main_t * vm, int is_main)
   if (is_main)
     {
       uword i;
+
+      /*
+       * Perform an initial barrier sync. Pays no attention to
+       * the barrier sync hold-down timer scheme, which won't work
+       * at this point in time.
+       */
+      vlib_worker_thread_initial_barrier_sync_and_release (vm);
+
       nm->current_process_index = ~0;
       for (i = 0; i < vec_len (nm->processes); i++)
 	cpu_time_now = dispatch_process (vm, nm->processes[i], /* frame */ 0,
@@ -2090,6 +2098,9 @@ vlib_main (vlib_main_t * volatile vm, unformat_input_t * input)
   if ((error = vlib_call_all_config_functions (vm, input, 0 /* is_early */ )))
     goto done;
 
+  /* Sort per-thread init functions before we start threads */
+  vlib_sort_init_exit_functions (&vm->worker_init_function_registrations);
+
   /* Call all main loop enter functions. */
   {
     clib_error_t *sub_error;