Re: [PATCH] mm/numa/thp: fix assumptions of migrate_misplaced_transhuge_page()

From: Mel Gorman
Date: Thu May 05 2016 - 05:54:05 EST

On Tue, May 03, 2016 at 02:33:51PM +0200, jglisse@xxxxxxxxxx wrote:
> From: Jérôme Glisse <jglisse@xxxxxxxxxx>
> Fix assumptions in migrate_misplaced_transhuge_page() which is only
> call by do_huge_pmd_numa_page() itself only call by __handle_mm_fault()
> for pmd with PROT_NONE. This means that if the pmd stays the same
> then there can be no concurrent get_user_pages / get_user_pages_fast
> (GUP/GUP_fast). More over because migrate_misplaced_transhuge_page()
> abort if page is mapped more than once then there can be no GUP from
> a different process. Finaly, holding the pmd lock assure us that no
> other part of the kernel can take an extra reference on the page.
> In the end this means that the failure code path should never be
> taken unless something is horribly wrong, so convert it to BUG_ON().
> Signed-off-by: Jérôme Glisse <jglisse@xxxxxxxxxx>
> Acked-by: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
> Cc: Mel Gorman <mgorman@xxxxxxx>
> Cc: Hugh Dickins <hughd@xxxxxxxxxx>
> Cc: Andrea Arcangeli <aarcange@xxxxxxxxxx>

Acked-by: Mel Gorman <mgorman@xxxxxxx>

Mel Gorman