diff --git a/[refs] b/[refs]
index d9b39358baa9..479dfc64afe5 100644
--- a/[refs]
+++ b/[refs]
@@ -1,2 +1,2 @@
 ---
-refs/heads/master: c36ef4b1762302a493c6cb754073bded084700e2
+refs/heads/master: ea2e7057c0234cfb8b09467d8f137760d371fc72
diff --git a/trunk/arch/arm/mm/dma-mapping.c b/trunk/arch/arm/mm/dma-mapping.c
index ab584569930e..1aa664a1999f 100644
--- a/trunk/arch/arm/mm/dma-mapping.c
+++ b/trunk/arch/arm/mm/dma-mapping.c
@@ -332,6 +332,15 @@ __dma_alloc(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp,
 	struct page *page;
 	void *addr;
 
+	/*
+	 * Following is a work-around (a.k.a. hack) to prevent pages
+	 * with __GFP_COMP being passed to split_page() which cannot
+	 * handle them.  The real problem is that this flag probably
+	 * should be 0 on ARM as it is not supported on this
+	 * platform; see CONFIG_HUGETLBFS.
+	 */
+	gfp &= ~(__GFP_COMP);
+
 	*handle = ~0;
 	size = PAGE_ALIGN(size);