x86: fix pmd_bad and pud_bad to support huge pages

I recently stumbled upon a problem in the support for huge pages. If a
program using huge pages does not explicitly unmap them, they remain
mapped (and therefore, are lost) after the program exits.

I observed that the free huge page count in /proc/meminfo decreased when
running my program, and it did not increase after the program exited.
After running the program a few times, no more huge pages could be
allocated.

The reason for this seems to be that the x86 pmd_bad and pud_bad
consider pmd/pud entries having the PSE bit set invalid. I think there
is nothing wrong with this bit being set, it just indicates that the
lowest level of translation has been reached. This bit has to be (and
is) checked after the basic validity of the entry has been checked, like
in this fragment from follow_page() in mm/memory.c:

  if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
          goto no_page_table;

  if (pmd_huge(*pmd)) {
          BUG_ON(flags & FOLL_GET);
          page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
          goto out;
  }

Note that this code currently doesn't work as intended if the pmd refers
to a huge page, the pmd_huge() check can not be reached if the page is
huge.

Extending pmd_bad() (and, for future 1GB page support, pud_bad()) to
allow for the PSE bit being set fixes this. For similar reasons,
allowing the NX bit being set is necessary, too. I have seen huge pages
having the NX bit set in their pmd entry, which would cause the same
problem.

Signed-Off-By: Hans Rosenfeld <hans.rosenfeld@amd.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

1 files changed, 3 insertions, 1 deletions

diff --git a/include/asm-x86/pgtable_32.h b/include/asm-x86/pgtable_32.h
index a842c7222b1e..b478efa971e0 100644
--- a/include/asm-x86/pgtable_32.h
+++ b/include/asm-x86/pgtable_32.h
@@ -91,7 +91,9 @@ extern unsigned long pg0[];
 /* To avoid harmful races, pmd_none(x) should check only the lower when PAE */
 #define pmd_none(x)     (!(unsigned long)pmd_val(x))
 #define pmd_present(x)  (pmd_val(x) & _PAGE_PRESENT)
-#define pmd_bad(x)      ((pmd_val(x) & (~PAGE_MASK & ~_PAGE_USER)) != _KERNPG_TABLE)
+#define pmd_bad(x)      ((pmd_val(x) \
+                          & ~(PAGE_MASK | _PAGE_USER | _PAGE_PSE | _PAGE_NX)) \
+                         != _KERNPG_TABLE)
 
 
 #define pages_to_mb(x) ((x) >> (20-PAGE_SHIFT))