RE: Hugepage regression

["Chen, Kenneth W" <kenneth.w.chen@intel.com>]

Hugh Dickins wrote on Tuesday, October 10, 2006 1:10 PM
> > can we reverse that order (call unmap_region
> > and then nulls out vma->vmfile and fput)?
> 
> I'm pretty sure we cannot: the ordering is quite intentional, that if
> a driver ->mmap failed, then it'd be wrong to call down to driver in
> the unmap_region (if a driver is nicely behaved, that unmap_region
> shouldn't be unnecessary; but some do rely on us clearing ptes there).


Even not something like the following?  I believe you that nullifying
vma->vm_file can not be done after unmap_region(), I just want to make
sure we are talking the same thing. It looks OK to me to defer the fput
in the do_mmap_pgoff clean up path.


--- ./mm/mmap.c.orig	2006-10-10 15:58:17.000000000 -0700
+++ ./mm/mmap.c	2006-10-10 15:59:02.000000000 -0700
@@ -1159,11 +1159,12 @@ out:	
 unmap_and_free_vma:
 	if (correct_wcount)
 		atomic_inc(&inode->i_writecount);
-	vma->vm_file = NULL;
-	fput(file);
 
 	/* Undo any partial mapping done by a device driver. */
 	unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
+
+	vma->vm_file = NULL;
+	fput(file);
 	charged = 0;
 free_vma:
 	kmem_cache_free(vm_area_cachep, vma);