Store per-thread node error counters

This fixed performance issue in muti-threaded setup
due to sharing of the same cacheline between multiple threads

Change-Id: I930ee44c17a83d4da350d15b4b97b8bb4633a9b0
Signed-off-by: Damjan Marion <damarion@cisco.com>

2 files changed, 65 insertions, 15 deletions

diff --git a/vlib/vlib/error.c b/vlib/vlib/error.c
index 59b89cefc3a..b04ac916d8e 100644
--- a/vlib/vlib/error.c
+++ b/vlib/vlib/error.c
@@ -151,6 +151,8 @@ void vlib_register_errors (vlib_main_t * vm,
   vlib_node_t * n = vlib_get_node (vm, node_index);
   uword l;
 
+  ASSERT(os_get_cpu_number() == 0);
+
   /* Free up any previous error strings. */
   if (n->n_errors > 0)
     heap_dealloc (em->error_strings_heap, n->error_heap_handle);
@@ -209,26 +211,59 @@ show_errors (vlib_main_t * vm,
   vlib_node_t * n;
   u32 code, i, ni;
   u64 c;
+  int index = 0;
+  int verbose = 0;
+  u64 * sums = 0;
+
+  if (unformat (input, "verbose"))
+    verbose = 1;
+
+  vec_validate(sums, vec_len(em->counters));
 
   vlib_cli_output (vm, "%=16s%=40s%=20s", "Count", "Node", "Reason");
 
+  foreach_vlib_main(({
+    em = &this_vlib_main->error_main;
+
+    if (verbose)
+      vlib_cli_output(vm, "Thread %u (%v):", index, vlib_worker_threads[index].name);
+
+    for (ni = 0; ni < vec_len (this_vlib_main->node_main.nodes); ni++)
+      {
+	n = vlib_get_node (this_vlib_main, ni);
+	for (code = 0; code < n->n_errors; code++)
+	  {
+	    i = n->error_heap_index + code;
+	    c = em->counters[i];
+	    if (i < vec_len (em->counters_last_clear))
+	      c -= em->counters_last_clear[i];
+	    sums[i] += c;
+
+	    if (c == 0 || !verbose)
+	      continue;
+
+	    vlib_cli_output (vm, "%16Ld%=40v%s", c, n->name, em->error_strings_heap[i]);
+	  }
+      }
+    index++;
+  }));
+
+  if (verbose)
+   vlib_cli_output(vm, "Total:");
+
   for (ni = 0; ni < vec_len (vm->node_main.nodes); ni++)
     {
       n = vlib_get_node (vm, ni);
       for (code = 0; code < n->n_errors; code++)
 	{
 	  i = n->error_heap_index + code;
-	  c = em->counters[i];
-	  if (i < vec_len (em->counters_last_clear))
-	    c -= em->counters_last_clear[i];
-
-	  if (c == 0)
-	    continue;
-
-	  vlib_cli_output (vm, "%16Ld%=40v%s", c, n->name, em->error_strings_heap[i]);
+	  if (sums[i])
+	    vlib_cli_output (vm, "%16Ld%=40v%s", sums[i], n->name, em->error_strings_heap[i]);
 	}
     }
 
+  vec_free(sums);
+
   return 0;
 }
 
@@ -249,12 +284,15 @@ clear_error_counters (vlib_main_t * vm,
 		      unformat_input_t * input,
 		      vlib_cli_command_t * cmd)
 {
-  vlib_error_main_t * em = &vm->error_main;
+  vlib_error_main_t * em;
   u32 i;
 
-  vec_validate (em->counters_last_clear, vec_len (em->counters) - 1);
-  for (i = 0; i < vec_len (em->counters); i++)
-    em->counters_last_clear[i] = em->counters[i];
+  foreach_vlib_main(({
+    em = &this_vlib_main->error_main;
+    vec_validate (em->counters_last_clear, vec_len (em->counters) - 1);
+    for (i = 0; i < vec_len (em->counters); i++)
+      em->counters_last_clear[i] = em->counters[i];
+  }));
   return 0;
 }
 
diff --git a/vlib/vlib/threads.c b/vlib/vlib/threads.c
index 4621f843dd5..405a4d644bc 100644
--- a/vlib/vlib/threads.c
+++ b/vlib/vlib/threads.c
@@ -70,7 +70,7 @@ os_get_cpu_number (void)
   /* Get any old stack address. */
   sp = &sp;
 
-  n = ((uword)sp - (uword)vlib_thread_stacks[0]) 
+  n = ((uword)sp - (uword)vlib_thread_stacks[0])
       >> VLIB_LOG2_THREAD_STACK_SIZE;
 
   /* "processes" have their own stacks, and they always run in thread 0 */
@@ -675,6 +675,11 @@ static clib_error_t * start_workers (vlib_main_t * vm)
 
             unix_physmem_init (vm_clone, 0 /* physmem not required */);
 
+	    vm_clone->error_main.counters =
+	      vec_dup(vlib_mains[0]->error_main.counters);
+	    vm_clone->error_main.counters_last_clear =
+	      vec_dup(vlib_mains[0]->error_main.counters_last_clear);
+
             /* Fork the vlib_buffer_main_t free lists, etc. */
             bm_clone = vec_dup (vm_clone->buffer_main);
             vm_clone->buffer_main = bm_clone;
@@ -817,16 +822,23 @@ void vlib_worker_thread_node_runtime_update(void)
       vlib_node_runtime_t * rt;
       w = vlib_worker_threads + i;
       oldheap = clib_mem_set_heap (w->thread_mheap);
-      
+
       vm_clone = vlib_mains[i];
 
       /* Re-clone error heap */
+      u64 * old_counters = vm_clone->error_main.counters;
+      u64 * old_counters_all_clear = vm_clone->error_main.counters_last_clear;
       memcpy (&vm_clone->error_main, &vm->error_main, sizeof (vm->error_main));
+      j = vec_len(vm->error_main.counters) - 1;
+      vec_validate_aligned(old_counters, j, CLIB_CACHE_LINE_BYTES);
+      vec_validate_aligned(old_counters_all_clear, j, CLIB_CACHE_LINE_BYTES);
+      vm_clone->error_main.counters = old_counters;
+      vm_clone->error_main.counters_last_clear = old_counters_all_clear;
 
       nm_clone = &vm_clone->node_main;
       vec_free (nm_clone->next_frames);
       nm_clone->next_frames = vec_dup (nm->next_frames);
-      
+
       for (j = 0; j < vec_len (nm_clone->next_frames); j++)
         {
           vlib_next_frame_t *nf = &nm_clone->next_frames[j];