On Thu, 2012-07-12 at 13:16 +0200, Michal Hocko wrote:

On Wed 11-07-12 18:48:02, Will Deacon wrote:

quoted

On Tue, Jul 10, 2012 at 11:42:34AM +0100, Will Deacon wrote:

quoted

On Tue, Jul 10, 2012 at 10:45:13AM +0100, Will Deacon wrote:

quoted

On Tue, Jul 10, 2012 at 12:57:14AM +0100, Hugh Dickins wrote:

quoted

If I start to grep the architectures for non-empty flush_dcache_page(),
I soon find things in arch/arm such as v4_mc_copy_user_highpage() doing
if (!test_and_set_bit(PG_dcache_clean,)) __flush_dcache_page() - where
the naming suggests that I'm right, it's the architecture's responsibility
to arrange whatever flushing is needed in its copy and clear page functions.

[...]

quoted

Ok, so this is exactly the problem. The hugetlb allocator uses its own
pool of huge pages, so free_huge_page followed by a later alloc_huge_page
will give you something where the page flags of the compound head do not
guarantee that PG_arch_1 is clear.

Just to confirm, the following quick hack at least results in the correct
flushing for me (on ARM):

diff --git a/mm/hugetlb.c b/mm/hugetlb.c
index e198831..7a7c9d3 100644
--- a/mm/hugetlb.c
+++ b/mm/hugetlb.c

@@ -1141,6 +1141,7 @@ static struct page *alloc_huge_page(struct vm_area_struct *vma,
        }
 
        set_page_private(page, (unsigned long)spool);
+       clear_bit(PG_arch_1, &page->flags);
 
        vma_commit_reservation(h, vma, addr);
 

The question is whether we should tidy that up for the core code or get
architectures to clear the bit in arch_make_huge_pte (which also seems to
work).

This should go into arch specific code IMO. Even the page flag name
suggests this shouldn't be in the base code.

Agree completely.

PG_Arch_1 is mostly used for flushing implementations, but it's meaning
isn't unified.  For instance Arm and Parisc have opposite meanings for
this flag.  Touching it in generic code can therefore *never* be the
right thing to do.  What you're looking for here is the correct flushing
(or rather flush notification) interface.

James