mm: fix infinite loop in filemap_fault

filemap_fault will go into an infinite loop if ->readpage() fails
asynchronously.

AFAICS the bug was introduced by this commit, which removed the wait after the
final readpage:

   commit d00806b
   Author: Nick Piggin <npiggin@suse.de>
   Date:   Thu Jul 19 01:46:57 2007 -0700

       mm: fix fault vs invalidate race for linear mappings

Fix by reintroducing the wait_on_page_locked() after ->readpage() to make sure
the page is up-to-date before jumping back to the beginning of the function.

I've noticed this while testing nfs exporting on fuse.  The patch
fixes it.

Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
Cc: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/filemap.c

@@ -1461,6 +1461,11 @@ int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 	 */
 	ClearPageError(page);
 	error = mapping->a_ops->readpage(file, page);
+	if (!error) {
+		wait_on_page_locked(page);
+		if (!PageUptodate(page))
+			error = -EIO;
+	}
 	page_cache_release(page);
 
 	if (!error || error == AOP_TRUNCATED_PAGE)