diff --git a/src/small_allocator.c b/src/small_allocator.c
index c92d1c1..61586c5 100644
--- a/src/small_allocator.c
+++ b/src/small_allocator.c
@@ -19,7 +19,9 @@
 
 
 #define ALIGNMENT 16
-#define ALIGN(n) (((n) + (ALIGNMENT - 1)) & ~(ALIGNMENT - 1))
+#define ALIGN_UP(n) (((n) + (ALIGNMENT - 1)) & ~(ALIGNMENT - 1))
+#define ALIGN_DOWN(n) ((n) & ~(ALIGNMENT - 1))
+
 #define NULL_BLOCK (0xffffffff)
 #define BLOCK_FREE  ((uint32_t)1)
 
@@ -103,7 +105,7 @@ void small_allocator_init(SmallAllocator* a, void* backing_store, uint32_t size)
     a->size = size;
 
     BlockHeader* first = backing_store;
-    uint32_t usable = ALIGN(size - sizeof(BlockHeader));
+    uint32_t usable = ALIGN_DOWN(size - sizeof(BlockHeader));
     block_set_free(first);
     block_set_size(first, usable);
     first->next = NULL_BLOCK;
@@ -115,7 +117,7 @@ void small_allocator_init(SmallAllocator* a, void* backing_store, uint32_t size)
 
 static void split_block(SmallAllocator *a, BlockHeader* block, uint32_t size)
 {
-    size = ALIGN(size);
+    size = ALIGN_UP(size);
     uint32_t bsize = block_size(block);
 
     if (bsize < size + sizeof(BlockHeader) + ALIGNMENT)
@@ -125,7 +127,7 @@ static void split_block(SmallAllocator *a, BlockHeader* block, uint32_t size)
     BlockHeader* new_block =
         (BlockHeader*)(base + sizeof(BlockHeader) + size);
 
-    uint32_t new_size = ALIGN(bsize - size - sizeof(BlockHeader));
+    uint32_t new_size = ALIGN_UP(bsize - size - sizeof(BlockHeader));
 
     block_set_size(block, size);
     block_set_used(block);
@@ -149,7 +151,7 @@ static void merge_with_next(SmallAllocator *a, BlockHeader* block)
     if (!next || !block_is_free(next))
         return;
 
-    uint32_t new_size = ALIGN(block_size(block) +
+    uint32_t new_size = ALIGN_UP(block_size(block) +
                             sizeof(BlockHeader) +
                             block_size(next));
 
@@ -162,7 +164,7 @@ static void merge_with_next(SmallAllocator *a, BlockHeader* block)
 
 void* small_alloc(SmallAllocator* a, uint32_t size)
 {
-    size = ALIGN(size);
+    size = ALIGN_UP(size);
 
     BlockHeader* start = a->last_alloc ? a->last_alloc : a->heap_start;
     BlockHeader* curr = start;
@@ -221,7 +223,7 @@ void debug_alloc(const SmallAllocator *a)
     puts("--debug end--");
 }
 
-char buff[1024*1024];
+char buff[1023*1023];
 int main(void)
 {
     printf("bh size %zu\n", sizeof(BlockHeader));