mm: numa: Do not mark PTEs pte_numa when splitting huge pages
authorMel Gorman <mgorman@suse.de>
Thu, 2 Oct 2014 18:47:42 +0000 (19:47 +0100)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 2 Oct 2014 18:57:18 +0000 (11:57 -0700)
This patch reverts 1ba6e0b50b ("mm: numa: split_huge_page: transfer the
NUMA type from the pmd to the pte"). If a huge page is being split due
a protection change and the tail will be in a PROT_NONE vma then NUMA
hinting PTEs are temporarily created in the protected VMA.

 VM_RW|VM_PROTNONE
|-----------------|
      ^
      split here

In the specific case above, it should get fixed up by change_pte_range()
but there is a window of opportunity for weirdness to happen. Similarly,
if a huge page is shrunk and split during a protection update but before
pmd_numa is cleared then a pte_numa can be left behind.

Instead of adding complexity trying to deal with the case, this patch
will not mark PTEs NUMA when splitting a huge page. NUMA hinting faults
will not be triggered which is marginal in comparison to the complexity
in dealing with the corner cases during THP split.

Cc: stable@vger.kernel.org
Signed-off-by: Mel Gorman <mgorman@suse.de>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/huge_memory.c

index d9a21d06b8623571cabe5f73532d02412a94ae9c..f8ffd9412ec502d10f2c103554ffe333306e2e3f 100644 (file)
@@ -1795,14 +1795,17 @@ static int __split_huge_page_map(struct page *page,
                for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
                        pte_t *pte, entry;
                        BUG_ON(PageCompound(page+i));
+                       /*
+                        * Note that pmd_numa is not transferred deliberately
+                        * to avoid any possibility that pte_numa leaks to
+                        * a PROT_NONE VMA by accident.
+                        */
                        entry = mk_pte(page + i, vma->vm_page_prot);
                        entry = maybe_mkwrite(pte_mkdirty(entry), vma);
                        if (!pmd_write(*pmd))
                                entry = pte_wrprotect(entry);
                        if (!pmd_young(*pmd))
                                entry = pte_mkold(entry);
-                       if (pmd_numa(*pmd))
-                               entry = pte_mknuma(entry);
                        pte = pte_offset_map(&_pmd, haddr);
                        BUG_ON(!pte_none(*pte));
                        set_pte_at(mm, haddr, pte, entry);