On Thu, May 31, 2012 at 04:35:53PM -0400, KOSAKI Motohiro wrote:

(5/31/12 4:25 PM), Andrew Morton wrote:

quoted

On Thu, 31 May 2012 16:09:15 -0400
KOSAKI Motohiro[off-list ref]  wrote:

quoted

quoted

--- a/mm/shmem.c
+++ b/mm/shmem.c

@@ -929,7 +929,7 @@ static struct page *shmem_alloc_page(gfp_t gfp,
   	/*
   	 * alloc_page_vma() will drop the shared policy reference
   	 */
-	return alloc_page_vma(gfp,&pvma, 0);
+	return alloc_page_vma(gfp,&pvma, info->node_offset<<   PAGE_SHIFT );

3rd argument of alloc_page_vma() is an address. This is type error.

Well, it's an unsigned long...

But yes, it is conceptually wrong and *looks* weird.  I think we can
address that by overcoming our peculair aversion to documenting our
code, sigh.  This?

Sorry, no.

addr agrument of alloc_pages_vma() have two meanings.

1) interleave node seed
2) look-up key of shmem policy

I think this patch break (2). shmem_get_policy(pol, addr) assume caller honor to
pass correct address.

But the pseudo vma we generated in shmem_alloc_page the vm_ops are set to NULL.
So get_vma_policy will return the policy provided by the pseudo vma and not reach
the shmem_get_policy.