perfmon: collect data on selected thread(s)

Add missing pre-input node runtime fork and refork code.
unix-epoll-input runs on all threads; each instance needs its own
runtime stats.

Change-Id: I16b02e42d0c95f863161176c4bb9f9917bef809d
Signed-off-by: Dave Barach <dave@barachs.net>

5 files changed, 129 insertions, 27 deletions

diff --git a/src/plugins/perfmon/perfmon.c b/src/plugins/perfmon/perfmon.c
index 359555705aa..7f621a1e671 100644
--- a/src/plugins/perfmon/perfmon.c
+++ b/src/plugins/perfmon/perfmon.c
@@ -301,11 +301,15 @@ set_pmc_command_fn (vlib_main_t * vm,
 		    unformat_input_t * input, vlib_cli_command_t * cmd)
 {
   perfmon_main_t *pm = &perfmon_main;
+  vlib_thread_main_t *vtm = vlib_get_thread_main ();
+  int num_threads = 1 + vtm->n_threads;
   unformat_input_t _line_input, *line_input = &_line_input;
   perfmon_event_config_t ec;
   f64 delay;
   u32 timeout_seconds;
   u32 deadman;
+  int last_set;
+  clib_error_t *error;
 
   vec_reset_length (pm->single_events_to_collect);
   vec_reset_length (pm->paired_events_to_collect);
@@ -315,6 +319,8 @@ set_pmc_command_fn (vlib_main_t * vm,
   if (!unformat_user (input, unformat_line_input, line_input))
     return clib_error_return (0, "counter names required...");
 
+  clib_bitmap_zero (pm->thread_bitmap);
+
   while (unformat_check_input (line_input) != UNFORMAT_END_OF_INPUT)
     {
       if (unformat (line_input, "timeout %u", &timeout_seconds))
@@ -343,6 +349,12 @@ set_pmc_command_fn (vlib_main_t * vm,
 	  ec.pe_config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS;
 	  vec_add1 (pm->paired_events_to_collect, ec);
 	}
+      else if (unformat (line_input, "threads %U",
+			 unformat_bitmap_list, &pm->thread_bitmap))
+	;
+      else if (unformat (line_input, "thread %U",
+			 unformat_bitmap_list, &pm->thread_bitmap))
+	;
       else if (unformat (line_input, "%U", unformat_processor_event, pm, &ec))
 	{
 	  vec_add1 (pm->single_events_to_collect, ec);
@@ -358,10 +370,20 @@ set_pmc_command_fn (vlib_main_t * vm,
       foreach_perfmon_event
 #undef _
 	else
-	return clib_error_return (0, "unknown input '%U'",
-				  format_unformat_error, line_input);
+	{
+	  error = clib_error_return (0, "unknown input '%U'",
+				     format_unformat_error, line_input);
+	  unformat_free (line_input);
+	  return error;
+	}
     }
 
+  unformat_free (line_input);
+
+  last_set = clib_bitmap_last_set (pm->thread_bitmap);
+  if (last_set != ~0 && last_set >= num_threads)
+    return clib_error_return (0, "thread %d does not exist", last_set);
+
   /* Stick paired events at the front of the (unified) list */
   if (vec_len (pm->paired_events_to_collect) > 0)
     {
@@ -410,7 +432,7 @@ set_pmc_command_fn (vlib_main_t * vm,
 VLIB_CLI_COMMAND (set_pmc_command, static) =
 {
   .path = "set pmc",
-  .short_help = "set pmc c1 [..., use \"show pmc events\"]",
+  .short_help = "set pmc [threads n,n1-n2] c1... [see \"show pmc events\"]",
   .function = set_pmc_command_fn,
   .is_mp_safe = 1,
 };
diff --git a/src/plugins/perfmon/perfmon.h b/src/plugins/perfmon/perfmon.h
index 9663dae36d1..9c4c34e36c1 100644
--- a/src/plugins/perfmon/perfmon.h
+++ b/src/plugins/perfmon/perfmon.h
@@ -121,6 +121,9 @@ typedef struct
   /* Current perf_event file descriptors, per thread */
   int **pm_fds;
 
+  /* thread bitmap */
+  uword *thread_bitmap;
+
   /* Logging */
   vlib_log_class_t log_class;
 
diff --git a/src/plugins/perfmon/perfmon_periodic.c b/src/plugins/perfmon/perfmon_periodic.c
index ccf3e9eed81..0811439d7eb 100644
--- a/src/plugins/perfmon/perfmon_periodic.c
+++ b/src/plugins/perfmon/perfmon_periodic.c
@@ -21,6 +21,7 @@
 #include <asm/unistd.h>
 #include <sys/ioctl.h>
 
+/* "not in glibc" */
 static long
 perf_event_open (struct perf_event_attr *hw_event, pid_t pid, int cpu,
 		 int group_fd, unsigned long flags)
@@ -114,6 +115,7 @@ enable_current_events (perfmon_main_t * pm)
   u32 my_thread_index = vm->thread_index;
   u32 index;
   int i, limit = 1;
+  int cpu;
 
   if ((pm->current_event + 1) < vec_len (pm->single_events_to_collect))
     limit = 2;
@@ -140,7 +142,9 @@ enable_current_events (perfmon_main_t * pm)
 	  pe.exclude_hv = 1;
 	}
 
-      fd = perf_event_open (&pe, 0, -1, -1, 0);
+      cpu = vm->cpu_index;
+
+      fd = perf_event_open (&pe, 0, cpu, -1, 0);
       if (fd == -1)
 	{
 	  clib_unix_warning ("event open: type %d config %d", c->pe_type,
@@ -237,25 +241,38 @@ static void
 start_event (perfmon_main_t * pm, f64 now, uword event_data)
 {
   int i;
+  int last_set;
+  int all = 0;
   pm->current_event = 0;
+
   if (vec_len (pm->single_events_to_collect) == 0)
     {
       pm->state = PERFMON_STATE_OFF;
       return;
     }
+
+  last_set = clib_bitmap_last_set (pm->thread_bitmap);
+  all = (last_set == ~0);
+
   pm->state = PERFMON_STATE_RUNNING;
   clear_counters (pm);
 
-  /* Start collection on this thread */
-  enable_current_events (pm);
+  /* Start collection on thread 0? */
+  if (all || clib_bitmap_get (pm->thread_bitmap, 0))
+    {
+      /* Start collection on this thread */
+      enable_current_events (pm);
+    }
 
   /* And also on worker threads */
   for (i = 1; i < vec_len (vlib_mains); i++)
     {
       if (vlib_mains[i] == 0)
 	continue;
-      vlib_mains[i]->worker_thread_main_loop_callback = (void *)
-	worker_thread_start_event;
+
+      if (all || clib_bitmap_get (pm->thread_bitmap, i))
+	vlib_mains[i]->worker_thread_main_loop_callback = (void *)
+	  worker_thread_start_event;
     }
 }
 
@@ -397,23 +414,46 @@ scrape_and_clear_counters (perfmon_main_t * pm)
 }
 
 static void
-handle_timeout (perfmon_main_t * pm, f64 now)
+handle_timeout (vlib_main_t * vm, perfmon_main_t * pm, f64 now)
 {
   int i;
-  disable_events (pm);
+  int last_set, all;
+
+  last_set = clib_bitmap_last_set (pm->thread_bitmap);
+  all = (last_set == ~0);
+
+  if (all || clib_bitmap_get (pm->thread_bitmap, 0))
+    disable_events (pm);
 
   /* And also on worker threads */
   for (i = 1; i < vec_len (vlib_mains); i++)
     {
       if (vlib_mains[i] == 0)
 	continue;
-      vlib_mains[i]->worker_thread_main_loop_callback = (void *)
-	worker_thread_stop_event;
+      if (all || clib_bitmap_get (pm->thread_bitmap, i))
+	vlib_mains[i]->worker_thread_main_loop_callback = (void *)
+	  worker_thread_stop_event;
     }
 
-  /* Short delay to make sure workers have stopped collection */
+  /* Make sure workers have stopped collection */
   if (i > 1)
-    vlib_process_suspend (pm->vlib_main, 1e-3);
+    {
+      f64 deadman = vlib_time_now (vm) + 1.0;
+
+      for (i = 1; i < vec_len (vlib_mains); i++)
+	{
+	  /* Has the worker actually stopped collecting data? */
+	  while (vlib_mains[i]->worker_thread_main_loop_callback)
+	    {
+	      if (vlib_time_now (vm) > deadman)
+		{
+		  clib_warning ("Thread %d deadman timeout!", i);
+		  break;
+		}
+	      vlib_process_suspend (pm->vlib_main, 1e-3);
+	    }
+	}
+    }
   scrape_and_clear_counters (pm);
   pm->current_event += pm->n_active;
   if (pm->current_event >= vec_len (pm->single_events_to_collect))
@@ -422,15 +462,18 @@ handle_timeout (perfmon_main_t * pm, f64 now)
       pm->state = PERFMON_STATE_OFF;
       return;
     }
-  enable_current_events (pm);
+
+  if (all || clib_bitmap_get (pm->thread_bitmap, 0))
+    enable_current_events (pm);
 
   /* And also on worker threads */
   for (i = 1; i < vec_len (vlib_mains); i++)
     {
       if (vlib_mains[i] == 0)
 	continue;
-      vlib_mains[i]->worker_thread_main_loop_callback = (void *)
-	worker_thread_start_event;
+      if (all || clib_bitmap_get (pm->thread_bitmap, i))
+	vlib_mains[i]->worker_thread_main_loop_callback = (void *)
+	  worker_thread_start_event;
     }
 }
 
@@ -464,7 +507,7 @@ perfmon_periodic_process (vlib_main_t * vm,
 
 	  /* Handle timeout */
 	case ~0:
-	  handle_timeout (pm, now);
+	  handle_timeout (vm, pm, now);
 	  break;
 
 	default:
diff --git a/src/vlib/main.c b/src/vlib/main.c
index 0e480fabe2a..3048a0d72de 100644
--- a/src/vlib/main.c
+++ b/src/vlib/main.c
@@ -1105,8 +1105,6 @@ dispatch_pcap_trace (vlib_main_t * vm,
     }
 }
 
-u64 oingo0, oingo1;
-
 static_always_inline u64
 dispatch_node (vlib_main_t * vm,
 	       vlib_node_runtime_t * node,
@@ -1207,12 +1205,6 @@ dispatch_node (vlib_main_t * vm,
       vm->main_loop_vectors_processed += n;
       vm->main_loop_nodes_processed += n > 0;
 
-      if (pmc_delta[0] || pmc_delta[1])
-	{
-	  oingo0 += pmc_delta[0];
-	  oingo1 += pmc_delta[1];
-	}
-
       v = vlib_node_runtime_update_stats (vm, node,
 					  /* n_calls */ 1,
 					  /* n_vectors */ n,
diff --git a/src/vlib/threads.c b/src/vlib/threads.c
index 45e4d89b7ab..e6ac6db543a 100644
--- a/src/vlib/threads.c
+++ b/src/vlib/threads.c
@@ -801,7 +801,7 @@ start_workers (vlib_main_t * vm)
 
 	      /* fork the frame dispatch queue */
 	      nm_clone->pending_frames = 0;
-	      vec_validate (nm_clone->pending_frames, 10);	/* $$$$$?????? */
+	      vec_validate (nm_clone->pending_frames, 10);
 	      _vec_len (nm_clone->pending_frames) = 0;
 
 	      /* fork nodes */
@@ -850,6 +850,21 @@ start_workers (vlib_main_t * vm)
 					 n->runtime_data_bytes));
 	      }
 
+	      nm_clone->nodes_by_type[VLIB_NODE_TYPE_PRE_INPUT] =
+		vec_dup_aligned (nm->nodes_by_type[VLIB_NODE_TYPE_PRE_INPUT],
+				 CLIB_CACHE_LINE_BYTES);
+	      vec_foreach (rt,
+			   nm_clone->nodes_by_type[VLIB_NODE_TYPE_PRE_INPUT])
+	      {
+		vlib_node_t *n = vlib_get_node (vm, rt->node_index);
+		rt->thread_index = vm_clone->thread_index;
+		/* copy initial runtime_data from node */
+		if (n->runtime_data && n->runtime_data_bytes > 0)
+		  clib_memcpy (rt->runtime_data, n->runtime_data,
+			       clib_min (VLIB_NODE_RUNTIME_DATA_SIZE,
+					 n->runtime_data_bytes));
+	      }
+
 	      nm_clone->processes = vec_dup_aligned (nm->processes,
 						     CLIB_CACHE_LINE_BYTES);
 
@@ -1173,6 +1188,33 @@ vlib_worker_thread_node_refork (void)
 
   vec_free (old_rt);
 
+  /* re-clone pre-input nodes */
+  old_rt = nm_clone->nodes_by_type[VLIB_NODE_TYPE_PRE_INPUT];
+  nm_clone->nodes_by_type[VLIB_NODE_TYPE_PRE_INPUT] =
+    vec_dup_aligned (nm->nodes_by_type[VLIB_NODE_TYPE_PRE_INPUT],
+		     CLIB_CACHE_LINE_BYTES);
+
+  vec_foreach (rt, nm_clone->nodes_by_type[VLIB_NODE_TYPE_PRE_INPUT])
+  {
+    vlib_node_t *n = vlib_get_node (vm, rt->node_index);
+    rt->thread_index = vm_clone->thread_index;
+    /* copy runtime_data, will be overwritten later for existing rt */
+    if (n->runtime_data && n->runtime_data_bytes > 0)
+      clib_memcpy_fast (rt->runtime_data, n->runtime_data,
+			clib_min (VLIB_NODE_RUNTIME_DATA_SIZE,
+				  n->runtime_data_bytes));
+  }
+
+  for (j = 0; j < vec_len (old_rt); j++)
+    {
+      rt = vlib_node_get_runtime (vm_clone, old_rt[j].node_index);
+      rt->state = old_rt[j].state;
+      clib_memcpy_fast (rt->runtime_data, old_rt[j].runtime_data,
+			VLIB_NODE_RUNTIME_DATA_SIZE);
+    }
+
+  vec_free (old_rt);
+
   nm_clone->processes = vec_dup_aligned (nm->processes,
 					 CLIB_CACHE_LINE_BYTES);
 }