vppinfra: numa vector placement support

Type: feature

Signed-off-by: Dave Barach <dave@barachs.net>
Change-Id: I7e7d95a089dd849c1f01ecea84529d8dbf239f21

4 files changed, 100 insertions, 16 deletions

diff --git a/src/vlib/cli.c b/src/vlib/cli.c
index bb6c5746537..85049884db6 100644
--- a/src/vlib/cli.c
+++ b/src/vlib/cli.c
@@ -733,9 +733,10 @@ show_memory_usage (vlib_main_t * vm,
 		   unformat_input_t * input, vlib_cli_command_t * cmd)
 {
   int verbose __attribute__ ((unused)) = 0;
-  int api_segment = 0, stats_segment = 0, main_heap = 0;
+  int api_segment = 0, stats_segment = 0, main_heap = 0, numa_heaps = 0;
   clib_error_t *error;
   u32 index = 0;
+  int i;
   uword clib_mem_trace_enable_disable (uword enable);
   uword was_enabled;
 
@@ -750,6 +751,8 @@ show_memory_usage (vlib_main_t * vm,
 	stats_segment = 1;
       else if (unformat (input, "main-heap"))
 	main_heap = 1;
+      else if (unformat (input, "numa-heaps"))
+	numa_heaps = 1;
       else
 	{
 	  error = clib_error_return (0, "unknown input `%U'",
@@ -758,9 +761,9 @@ show_memory_usage (vlib_main_t * vm,
 	}
     }
 
-  if ((api_segment + stats_segment + main_heap) == 0)
+  if ((api_segment + stats_segment + main_heap + numa_heaps) == 0)
     return clib_error_return
-      (0, "Please supply one of api-segment, stats-segment or main-heap");
+      (0, "Need one of api-segment, stats-segment, main-heap or numa-heaps");
 
   if (api_segment)
     {
@@ -801,6 +804,7 @@ show_memory_usage (vlib_main_t * vm,
       vec_free (s);
     }
 
+
 #if USE_DLMALLOC == 0
   /* *INDENT-OFF* */
   foreach_vlib_main (
@@ -849,6 +853,32 @@ show_memory_usage (vlib_main_t * vm,
 	/* Restore the trace flag */
 	clib_mem_trace_enable_disable (was_enabled);
       }
+    if (numa_heaps)
+      {
+	struct dlmallinfo mi;
+	void *mspace;
+
+	for (i = 0; i < ARRAY_LEN (clib_per_numa_mheaps); i++)
+	  {
+	    if (clib_per_numa_mheaps[i] == 0)
+	      continue;
+	    if (clib_per_numa_mheaps[i] == clib_per_cpu_mheaps[i])
+	      {
+		vlib_cli_output (vm, "Numa %d uses the main heap...", i);
+		continue;
+	      }
+	    was_enabled = clib_mem_trace_enable_disable (0);
+	    mspace = clib_per_numa_mheaps[i];
+
+	    mi = mspace_mallinfo (mspace);
+	    vlib_cli_output (vm, "Numa %d:", i);
+	    vlib_cli_output (vm, "  %U\n", format_page_map,
+			     pointer_to_uword (mspace_least_addr (mspace)),
+			     mi.arena);
+	    vlib_cli_output (vm, "  %U\n", format_mheap,
+			     clib_per_numa_mheaps[index], verbose);
+	  }
+      }
   }
 #endif /* USE_DLMALLOC */
   return 0;
@@ -857,7 +887,8 @@ show_memory_usage (vlib_main_t * vm,
 /* *INDENT-OFF* */
 VLIB_CLI_COMMAND (show_memory_usage_command, static) = {
   .path = "show memory",
-  .short_help = "show memory [api-segment][stats-segment][verbose]",
+  .short_help = "show memory [api-segment][stats-segment][verbose]\n"
+  "            [numa-heaps]",
   .function = show_memory_usage,
 };
 /* *INDENT-ON* */
@@ -905,6 +936,7 @@ enable_disable_memory_trace (vlib_main_t * vm,
   int api_segment = 0;
   int stats_segment = 0;
   int main_heap = 0;
+  u32 numa_id = ~0;
   void *oldheap;
 
   if (!unformat_user (input, unformat_line_input, line_input))
@@ -920,6 +952,8 @@ enable_disable_memory_trace (vlib_main_t * vm,
 	stats_segment = 1;
       else if (unformat (line_input, "main-heap"))
 	main_heap = 1;
+      else if (unformat (line_input, "numa-heap %d", &numa_id))
+	;
       else
 	{
 	  unformat_free (line_input);
@@ -928,10 +962,12 @@ enable_disable_memory_trace (vlib_main_t * vm,
     }
   unformat_free (line_input);
 
-  if ((api_segment + stats_segment + main_heap + (enable == 0)) == 0)
+  if ((api_segment + stats_segment + main_heap + (enable == 0)
+       + (numa_id != ~0)) == 0)
     {
       return clib_error_return
-	(0, "Need one of main-heap, stats-segment or api-segment");
+	(0, "Need one of main-heap, stats-segment, api-segment,\n"
+	 "numa-heap <nn> or disable");
     }
 
   /* Turn off current trace, if any */
@@ -975,13 +1011,31 @@ enable_disable_memory_trace (vlib_main_t * vm,
       clib_mem_trace (main_heap);
     }
 
+  if (numa_id != ~0)
+    {
+      if (numa_id >= ARRAY_LEN (clib_per_numa_mheaps))
+	return clib_error_return (0, "Numa %d out of range", numa_id);
+      if (clib_per_numa_mheaps[numa_id] == 0)
+	return clib_error_return (0, "Numa %d heap not active", numa_id);
+
+      if (clib_per_numa_mheaps[numa_id] == clib_mem_get_heap ())
+	return clib_error_return (0, "Numa %d uses the main heap...",
+				  numa_id);
+      current_traced_heap = clib_per_numa_mheaps[numa_id];
+      oldheap = clib_mem_set_heap (current_traced_heap);
+      clib_mem_trace (1);
+      clib_mem_set_heap (oldheap);
+    }
+
+
   return 0;
 }
 
 /* *INDENT-OFF* */
 VLIB_CLI_COMMAND (enable_disable_memory_trace_command, static) = {
   .path = "memory-trace",
-  .short_help = "memory-trace on|off [api-segment][stats-segment][main-heap]\n",
+  .short_help = "memory-trace on|off [api-segment][stats-segment][main-heap]\n"
+  "                   [numa-heap <numa-id>]\n",
   .function = enable_disable_memory_trace,
 };
 /* *INDENT-ON* */
diff --git a/src/vlib/threads.c b/src/vlib/threads.c
index e6733d55b6f..a827e3594e8 100644
--- a/src/vlib/threads.c
+++ b/src/vlib/threads.c
@@ -577,12 +577,12 @@ vlib_worker_thread_bootstrap_fn (void *arg)
   return rv;
 }
 
-static void
-vlib_get_thread_core_socket (vlib_worker_thread_t * w, unsigned cpu_id)
+void
+vlib_get_thread_core_numa (vlib_worker_thread_t * w, unsigned cpu_id)
 {
   const char *sys_cpu_path = "/sys/devices/system/cpu/cpu";
   u8 *p = 0;
-  int core_id = -1, socket_id = -1;
+  int core_id = -1, numa_id = -1;
 
   p = format (p, "%s%u/topology/core_id%c", sys_cpu_path, cpu_id, 0);
   clib_sysfs_read ((char *) p, "%d", &core_id);
@@ -590,11 +590,11 @@ vlib_get_thread_core_socket (vlib_worker_thread_t * w, unsigned cpu_id)
   p =
     format (p, "%s%u/topology/physical_package_id%c", sys_cpu_path, cpu_id,
 	    0);
-  clib_sysfs_read ((char *) p, "%d", &socket_id);
+  clib_sysfs_read ((char *) p, "%d", &numa_id);
   vec_free (p);
 
   w->core_id = core_id;
-  w->socket_id = socket_id;
+  w->numa_id = numa_id;
 }
 
 static clib_error_t *
@@ -602,9 +602,29 @@ vlib_launch_thread_int (void *fp, vlib_worker_thread_t * w, unsigned cpu_id)
 {
   vlib_thread_main_t *tm = &vlib_thread_main;
   void *(*fp_arg) (void *) = fp;
+  void *numa_heap;
 
   w->cpu_id = cpu_id;
-  vlib_get_thread_core_socket (w, cpu_id);
+  vlib_get_thread_core_numa (w, cpu_id);
+  os_set_numa_index (w->numa_id);
+
+  /* Set up NUMA-bound heap if indicated */
+  if (clib_per_numa_mheaps[w->numa_id] == 0)
+    {
+      /* If the user requested a NUMA heap, create it... */
+      if (tm->numa_heap_size)
+	{
+	  numa_heap = clib_mem_init_thread_safe_numa
+	    (0 /* DIY */ , tm->numa_heap_size);
+	  clib_per_numa_mheaps[w->numa_id] = numa_heap;
+	}
+      else
+	{
+	  /* Or, use the main heap */
+	  clib_per_numa_mheaps[w->numa_id] = w->thread_mheap;
+	}
+    }
+
   if (tm->cb.vlib_launch_thread_cb && !w->registration->use_pthreads)
     return tm->cb.vlib_launch_thread_cb (fp, (void *) w, cpu_id);
   else
@@ -1242,6 +1262,9 @@ cpu_config (vlib_main_t * vm, unformat_input_t * input)
 	;
       else if (unformat (input, "skip-cores %u", &tm->skip_cores))
 	;
+      else if (unformat (input, "numa-heap-size %U",
+			 unformat_memory_size, &tm->numa_heap_size))
+	;
       else if (unformat (input, "coremask-%s %U", &name,
 			 unformat_bitmap_mask, &bitmap) ||
 	       unformat (input, "corelist-%s %U", &name,
diff --git a/src/vlib/threads.h b/src/vlib/threads.h
index 312323c096d..c1188cea933 100644
--- a/src/vlib/threads.h
+++ b/src/vlib/threads.h
@@ -110,7 +110,7 @@ typedef struct
   long lwp;
   int cpu_id;
   int core_id;
-  int socket_id;
+  int numa_id;
   pthread_t thread_id;
 } vlib_worker_thread_t;
 
@@ -338,6 +338,10 @@ typedef struct
   /* callbacks */
   vlib_thread_callbacks_t cb;
   int extern_thread_mgmt;
+
+  /* NUMA-bound heap size */
+  uword numa_heap_size;
+
 } vlib_thread_main_t;
 
 extern vlib_thread_main_t vlib_thread_main;
@@ -613,6 +617,9 @@ void
 vlib_process_signal_event_mt_helper (vlib_process_signal_event_mt_args_t *
 				     args);
 void vlib_rpc_call_main_thread (void *function, u8 * args, u32 size);
+void vlib_get_thread_core_numa (vlib_worker_thread_t * w, unsigned cpu_id);
+
+
 #endif /* included_vlib_threads_h */
 
 /*
diff --git a/src/vlib/threads_cli.c b/src/vlib/threads_cli.c
index 65b3e2a5496..bcb85ec69fb 100644
--- a/src/vlib/threads_cli.c
+++ b/src/vlib/threads_cli.c
@@ -67,8 +67,8 @@ show_threads_fn (vlib_main_t * vm,
       if (cpu_id > -1)
 	{
 	  int core_id = w->core_id;
-	  int socket_id = w->socket_id;
-	  line = format (line, "%-7u%-7u%-7u%", cpu_id, core_id, socket_id);
+	  int numa_id = w->numa_id;
+	  line = format (line, "%-7u%-7u%-7u%", cpu_id, core_id, numa_id);
 	}
       else
 	{