[PATCH] Hugetlb page bug fix for i386 in PAE mode

Hit a bug check when unmap a hugetlb vma in PAE mode on i386 (and x86-64).

 Bad page state at free_hot_cold_page (in process 'a.out', page c165cc40)
 flags:0x20000000 mapping:f75e1d00 mapped:0 count:0
 Backtrace:
 Call Trace:
  [<c0133e0d>] bad_page+0x79/0x9e
  [<c0134550>] free_hot_cold_page+0x71/0xfa
  [<c0115d60>] unmap_hugepage_range+0xa3/0xbf
  [<c013d375>] unmap_vmas+0xac/0x252
  [<c0117691>] default_wake_function+0x0/0xc
  [<c0140bea>] unmap_region+0xd8/0x145
  [<c0140f2d>] do_munmap+0xfc/0x14d
  [<c01b8a56>] sys_shmdt+0xa5/0x126
  [<c010a2ad>] sys_ipc+0x23c/0x27f
  [<c014a85e>] sys_write+0x38/0x59
  [<c0103e1b>] syscall_call+0x7/0xb

It turns out there is a bug in hugetlb_prefault(): with 3 level page table,
huge_pte_alloc() might return a pmd that points to a PTE page.  It happens
if the virtual address for hugetlb mmap is recycled from previously used
normal page mmap.  free_pgtables() might not scrub the pmd entry on munmap
and hugetlb_prefault skips on any pmd presence regardless what type it is. 
Patch to fix the bug.
Signed-off-by: Ken Chen <kenneth.w.chen@intel.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

arch/i386/mm/hugetlbpage.c

@@ -244,8 +244,15 @@ int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
 			ret = -ENOMEM;
 			goto out;
 		}
-		if (!pte_none(*pte))
-			continue;
+
+		if (!pte_none(*pte)) {
+			pmd_t *pmd = (pmd_t *) pte;
+
+			page = pmd_page(*pmd);
+			pmd_clear(pmd);
+			dec_page_state(nr_page_table_pages);
+			page_cache_release(page);
+		}
 
 		idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
 			+ (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));