On Wed, Jun 09, 2021 at 11:52:37PM -0700, Hugh Dickins wrote:
> Running certain tests with a DEBUG_VM kernel would crash within hours,
> on the total_mapcount BUG() in split_huge_page_to_list(), while trying
> to free up some memory by punching a hole in a shmem huge page: split's
> try_to_unmap() was unable to find all the mappings of the page (which,
> on a !DEBUG_VM kernel, would then keep the huge page pinned in memory).
>
> Crash dumps showed two tail pages of a shmem huge page remained mapped
> by pte: ptes in a non-huge-aligned vma of a gVisor process, at the end
> of a long unmapped range; and no page table had yet been allocated for
> the head of the huge page to be mapped into.
>
> Although designed to handle these odd misaligned huge-page-mapped-by-pte
> cases, page_vma_mapped_walk() falls short by returning false prematurely
> when !pmd_present or !pud_present or !p4d_present or !pgd_present: there
> are cases when a huge page may span the boundary, with ptes present in
> the next.
>
> Restructure page_vma_mapped_walk() as a loop to continue in these cases,
> while keeping its layout much as before. Add a step_forward() helper to
> advance pvmw->address across those boundaries: originally I tried to use
> mm's standard p?d_addr_end() macros, but hit the same crash 512 times
> less often: because of the way redundant levels are folded together,
> but folded differently in different configurations, it was just too
> difficult to use them correctly; and step_forward() is simpler anyway.
>
> Fixes: ace71a19cec5 ("mm: introduce page_vma_mapped_walk()")
> Signed-off-by: Hugh Dickins <hughd@xxxxxxxxxx>
> Cc: <stable@xxxxxxxxxxxxxxx>
Acked-by: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
--
Kirill A. Shutemov
