[PATCH] fix VmSize and VmData after mremap

mremap's move_vma is applying __vm_stat_account to the old vma which may
have already been freed: move it to just before the do_munmap.

mremapping to and fro with CONFIG_DEBUG_SLAB=y showed /proc/<pid>/status
VmSize and VmData wrapping just like in kernel bugzilla #4842, and fixed by
this patch - worth including in 2.6.13, though not yet confirmed that it
fixes that specific report from Frank van Maarseveen.

Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

mm/mremap.c

@@ -229,6 +229,7 @@ static unsigned long move_vma(struct vm_area_struct *vma,
 	 * since do_munmap() will decrement it by old_len == new_len
 	 */
 	mm->total_vm += new_len >> PAGE_SHIFT;
+	__vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
 
 	if (do_munmap(mm, old_addr, old_len) < 0) {
 		/* OOM: unable to split vma, just get accounts right */
@@ -243,7 +244,6 @@ static unsigned long move_vma(struct vm_area_struct *vma,
 			vma->vm_next->vm_flags |= VM_ACCOUNT;
 	}
 
-	__vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
 	if (vm_flags & VM_LOCKED) {
 		mm->locked_vm += new_len >> PAGE_SHIFT;
 		if (new_len > old_len)