buffers: support cases where numa node memory is not available

Change-Id: I550ef893e41d86310d4e5af16d6100b8e557b68a
Signed-off-by: Damjan Marion <damarion@cisco.com>

3 files changed, 82 insertions, 40 deletions

diff --git a/src/vlib/buffer.c b/src/vlib/buffer.c
index 9312e8c0e47..0e972caadc9 100644
--- a/src/vlib/buffer.c
+++ b/src/vlib/buffer.c
@@ -493,9 +493,9 @@ vlib_buffer_chain_append_data_with_alloc (vlib_main_t * vm,
   return copied;
 }
 
-clib_error_t *
-vlib_buffer_pool_create (vlib_main_t * vm, u8 index, char *name,
-			 u32 data_size, u32 physmem_map_index)
+u8
+vlib_buffer_pool_create (vlib_main_t * vm, char *name, u32 data_size,
+			 u32 physmem_map_index)
 {
   vlib_buffer_main_t *bm = vm->buffer_main;
   vlib_buffer_pool_t *bp;
@@ -503,17 +503,12 @@ vlib_buffer_pool_create (vlib_main_t * vm, u8 index, char *name,
   uword start = pointer_to_uword (m->base);
   uword size = (uword) m->n_pages << m->log2_page_size;
   uword i, j;
-  u32 alloc_size, n_alloc_per_page;;
-
-  vec_validate_aligned (bm->buffer_pools, index, CLIB_CACHE_LINE_BYTES);
-  bp = vec_elt_at_index (bm->buffer_pools, index);
+  u32 alloc_size, n_alloc_per_page;
 
-  if (bp->start)
-    return clib_error_return (0, "buffer with index %u already exists",
-			      index);
+  if (vec_len (bm->buffer_pools) >= 255)
+    return ~0;
 
-  if (index >= 255)
-    return clib_error_return (0, "buffer index must be < 255", index);
+  vec_add2_aligned (bm->buffer_pools, bp, 1, CLIB_LOG2_CACHE_LINE_BYTES);
 
   if (bm->buffer_mem_size == 0)
     {
@@ -581,8 +576,7 @@ vlib_buffer_pool_create (vlib_main_t * vm, u8 index, char *name,
       }
 
   bp->n_buffers = vec_len (bp->buffers);
-
-  return 0;
+  return bp->index;
 }
 
 static u8 *
@@ -659,7 +653,8 @@ vlib_buffer_worker_init (vlib_main_t * vm)
 VLIB_WORKER_INIT_FUNCTION (vlib_buffer_worker_init);
 
 static clib_error_t *
-vlib_buffer_main_init_numa_node (struct vlib_main_t *vm, u32 numa_node)
+vlib_buffer_main_init_numa_node (struct vlib_main_t *vm, u32 numa_node,
+				 u8 * index)
 {
   vlib_buffer_main_t *bm = vm->buffer_main;
   clib_error_t *error;
@@ -703,9 +698,14 @@ retry:
   vec_reset_length (name);
   name = format (name, "default-numa-%d%c", numa_node, 0);
 
-  return vlib_buffer_pool_create (vm, numa_node, (char *) name,
-				  vlib_buffer_get_default_data_size (vm),
-				  physmem_map_index);
+  *index = vlib_buffer_pool_create (vm, (char *) name,
+				    vlib_buffer_get_default_data_size (vm),
+				    physmem_map_index);
+
+  if (*index == (u8) ~ 0)
+    return clib_error_return (0, "maximum number of buffer pools reached");
+
+  return 0;
 }
 
 void
@@ -790,10 +790,10 @@ vlib_buffer_main_init (struct vlib_main_t * vm)
 {
   vlib_buffer_main_t *bm;
   clib_error_t *err;
-  clib_bitmap_t *bmp = 0;
+  clib_bitmap_t *bmp = 0, *bmp_has_memory = 0;
   u32 numa_node;
   vlib_buffer_pool_t *bp;
-  u8 *name;
+  u8 *name = 0, first_valid_buffer_pool_index = ~0;
 
   vlib_buffer_main_alloc (vm);
 
@@ -803,42 +803,82 @@ vlib_buffer_main_init (struct vlib_main_t * vm)
 
   clib_spinlock_init (&bm->buffer_known_hash_lockp);
 
-  err = clib_sysfs_read ("/sys/devices/system/node/possible", "%U",
-			 unformat_bitmap_list, &bmp);
-  if (err)
+  if ((err = clib_sysfs_read ("/sys/devices/system/node/online", "%U",
+			      unformat_bitmap_list, &bmp)))
+    clib_error_free (err);
+
+  if ((err = clib_sysfs_read ("/sys/devices/system/node/has_memory", "%U",
+			      unformat_bitmap_list, &bmp_has_memory)))
+    clib_error_free (err);
+
+  if (bmp && bmp_has_memory)
+    bmp = clib_bitmap_and (bmp, bmp_has_memory);
+
+  /* no info from sysfs, assuming that only numa 0 exists */
+  if (bmp == 0)
+    bmp = clib_bitmap_set (bmp, 0, 1);
+
+  if (clib_bitmap_last_set (bmp) >= VLIB_BUFFER_MAX_NUMA_NODES)
+    clib_panic ("system have more than %u NUMA nodes",
+		VLIB_BUFFER_MAX_NUMA_NODES);
+
+  /* *INDENT-OFF* */
+  clib_bitmap_foreach (numa_node, bmp,
+    {
+      u8 *index = bm->default_buffer_pool_index_for_numa + numa_node;
+      index[0] = ~0;
+      if ((err = vlib_buffer_main_init_numa_node (vm, numa_node, index)))
+        {
+	  clib_error_report (err);
+	  clib_error_free (err);
+	  continue;
+	}
+
+      if (first_valid_buffer_pool_index == 0xff)
+        first_valid_buffer_pool_index = index[0];
+    });
+  /* *INDENT-ON* */
+
+  if (first_valid_buffer_pool_index == (u8) ~ 0)
     {
-      /* no info from sysfs, assuming that only numa 0 exists */
-      clib_error_free (err);
-      bmp = clib_bitmap_set (bmp, 0, 1);
+      err = clib_error_return (0, "failed to allocate buffer pool(s)");
+      goto done;
     }
 
   /* *INDENT-OFF* */
-  clib_bitmap_foreach (numa_node, bmp, {
-      if ((err = vlib_buffer_main_init_numa_node(vm, numa_node)))
-	  goto done;
+  clib_bitmap_foreach (numa_node, bmp,
+    {
+      if (bm->default_buffer_pool_index_for_numa[numa_node]  == (u8) ~0)
+	bm->default_buffer_pool_index_for_numa[numa_node] =
+	  first_valid_buffer_pool_index;
     });
   /* *INDENT-ON* */
 
-  bm->n_numa_nodes = clib_bitmap_last_set (bmp) + 1;
-
   vec_foreach (bp, bm->buffer_pools)
   {
-    name = format (0, "/buffer-pools/%s/cached%c", bp->name, 0);
+    if (bp->n_buffers == 0)
+      continue;
+
+    vec_reset_length (name);
+    name = format (name, "/buffer-pools/%s/cached%c", bp->name, 0);
     stat_segment_register_gauge (name, buffer_gauges_update_cached_fn,
 				 bp - bm->buffer_pools);
-    vec_free (name);
-    name = format (0, "/buffer-pools/%s/used%c", bp->name, 0);
+
+    vec_reset_length (name);
+    name = format (name, "/buffer-pools/%s/used%c", bp->name, 0);
     stat_segment_register_gauge (name, buffer_gauges_update_used_fn,
 				 bp - bm->buffer_pools);
-    vec_free (name);
-    name = format (0, "/buffer-pools/%s/available%c", bp->name, 0);
+
+    vec_reset_length (name);
+    name = format (name, "/buffer-pools/%s/available%c", bp->name, 0);
     stat_segment_register_gauge (name, buffer_gauges_update_available_fn,
 				 bp - bm->buffer_pools);
-    vec_free (name);
   }
 
 done:
   vec_free (bmp);
+  vec_free (bmp_has_memory);
+  vec_free (name);
   return err;
 }
 
diff --git a/src/vlib/buffer.h b/src/vlib/buffer.h
index 31baf5fd21e..514e73bd7ca 100644
--- a/src/vlib/buffer.h
+++ b/src/vlib/buffer.h
@@ -394,6 +394,8 @@ typedef struct
   vlib_buffer_t buffer_template;
 } vlib_buffer_pool_t;
 
+#define VLIB_BUFFER_MAX_NUMA_NODES 32
+
 typedef struct
 {
   CLIB_CACHE_LINE_ALIGN_MARK (cacheline0);
@@ -409,7 +411,7 @@ typedef struct
      has never been allocated. */
   uword *buffer_known_hash;
   clib_spinlock_t buffer_known_hash_lockp;
-  u32 n_numa_nodes;
+  u8 default_buffer_pool_index_for_numa[VLIB_BUFFER_MAX_NUMA_NODES];
 
   /* config */
   u32 buffers_per_numa;
diff --git a/src/vlib/buffer_funcs.h b/src/vlib/buffer_funcs.h
index 04f775da1f5..3e0234b6b51 100644
--- a/src/vlib/buffer_funcs.h
+++ b/src/vlib/buffer_funcs.h
@@ -126,8 +126,8 @@ vlib_buffer_copy_template (vlib_buffer_t * b, vlib_buffer_t * bt)
 always_inline u8
 vlib_buffer_pool_get_default_for_numa (vlib_main_t * vm, u32 numa_node)
 {
-  ASSERT (numa_node < vm->buffer_main->n_numa_nodes);
-  return numa_node;
+  ASSERT (numa_node < VLIB_BUFFER_MAX_NUMA_NODES);
+  return vm->buffer_main->default_buffer_pool_index_for_numa[numa_node];
 }
 
 /** \brief Translate array of buffer indices into buffer pointers with offset