diff --git a/heap.c b/heap.c
index 820f4ac..d372d9c 100644
--- a/heap.c
+++ b/heap.c
@@ -11,9 +11,9 @@ static mblock_t* mblock_by_data(void* data)
     return (mblock_t*)((u8*)data - MBLOCK_SIZE);
 }
 
-static u64 mblock_get_size(mblock_t* block)
+static size_t mblock_get_size(mblock_t* block)
 {
-    return MBLOCK_ALIGN_SIZE(block->size);
+    return MBLOCK_ALIGN_SIZE(block->size & ~MBLOCK_ATTR_ALLOC);
 }
 
 void heap_init(mheap_t* heap, void* data, uword size)
@@ -25,7 +25,7 @@ void heap_init(mheap_t* heap, void* data, uword size)
 void heap_join(mblock_t* start, int dir)
 {
     mblock_t* cur;
-    u64 blk_size;
+    size_t blk_size;
     
     blk_size = 0;
     cur = start;
@@ -62,7 +62,7 @@ void heap_split(mblock_t* block, size_t req_size)
     size_t tot_size = mblock_get_size(block);
     mblock_t* split;
     
-    block->size = (req_size)|MBLOCK_ATTR_ALLOC;
+    block->size = MBLOCK_ALIGN_SIZE(req_size)|MBLOCK_ATTR_ALLOC;
     split = mblock_get_next(block);
     
     //Init prev and next
diff --git a/heap.h b/heap.h
index 2d2b859..aa4cc02 100644
--- a/heap.h
+++ b/heap.h
@@ -15,18 +15,18 @@ typedef struct {
     uword size;
 } mheap_t;
 
-#define MBLOCK_SIZE             (CU_WORD_SIZE*3)
-#define MBLOCK_SIZE_MASK        ((CU_WORD_BITS/8) - 1)
-#define MBLOCK_ALIGN_SIZE(size) ((size) & ~MBLOCK_SIZE_MASK)
-#define MBLOCK_ATTR_ALLOC       (1<<0)
-#define IS_MBLOCK_ALLOC(mblock) (mblock->size & MBLOCK_ATTR_ALLOC)
-
 #ifdef CU_64BIT
 #   define  MBLOCK_ALIGN    5
 #else
 #   define  MBLOCK_ALIGN    4
 #endif
 
+#define MBLOCK_SIZE             (CU_WORD_SIZE*3)
+#define MBLOCK_SIZE_MASK        ((CU_WORD_BITS/8) - 1)
+#define MBLOCK_ALIGN_SIZE(size) (cu_round2_up((size), MBLOCK_ALIGN))
+#define MBLOCK_ATTR_ALLOC       (1<<0)
+#define IS_MBLOCK_ALLOC(mblock) (mblock->size & MBLOCK_ATTR_ALLOC)
+
 void heap_init(mheap_t* heap, void* data, uword size);
 void heap_join(mblock_t* start, int dir);
 void heap_split(mblock_t* block, size_t req_size);
diff --git a/test.c b/test.c
index ea37957..763d490 100644
--- a/test.c
+++ b/test.c
@@ -58,9 +58,9 @@ u8 internal_heap[4096];
 int main()
 {
     cutil_init(internal_heap, 4096);
-    cu_malloc = malloc;
+    /*cu_malloc = malloc;
     cu_realloc = realloc;
-    cu_free = free;
+    cu_free = free;*/
 
     printf("[list]\n");
     list_t list;
@@ -116,6 +116,7 @@ int main()
     printf("[bitmap]\n");
     printf("bit_align2\t%u\n", bit_align2(354, 3));
     printf("bit_log2\t%u\n", bit_log2(34));
+    printf("cu_align2_up\t%u\n", cu_round2_up(354, 3));
     num = 5;
     bit_set(num, 1, 1);
     printf("bit_get\t%u\n", bit_get(num, 0));