On Thu, Jul 19, 2018 at 07:13:39AM -0700, Dave Hansen wrote:
> On 07/19/2018 01:59 AM, Kirill A. Shutemov wrote:
> > On Wed, Jul 18, 2018 at 04:11:57PM -0700, Dave Hansen wrote:
> >> On 07/17/2018 04:20 AM, Kirill A. Shutemov wrote:
> >>> khugepaged allocates page in advance, before we found a VMA for
> >>> collapse. We don't yet know which KeyID to use for the allocation.
> >>
> >> That's not really true. We have the VMA and the address in the caller
> >> (khugepaged_scan_pmd()), but we drop the lock and have to revalidate the
> >> VMA.
> >
> > For !NUMA we allocate the page in khugepaged_do_scan(), well before we
> > know VMA.
>
> Ahh, thanks for clarifying. That's some more very good information
> about the design and progression of your patch that belongs in the
> changelog.
Okay.
> >>> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> >>> index 5ae34097aed1..d116f4ebb622 100644
> >>> --- a/mm/khugepaged.c
> >>> +++ b/mm/khugepaged.c
> >>> @@ -1056,6 +1056,16 @@ static void collapse_huge_page(struct mm_struct *mm,
> >>> */
> >>> anon_vma_unlock_write(vma->anon_vma);
> >>>
> >>> + /*
> >>> + * At this point new_page is allocated as non-encrypted.
> >>> + * If VMA's KeyID is non-zero, we need to prepare it to be encrypted
> >>> + * before coping data.
> >>> + */
> >>> + if (vma_keyid(vma)) {
> >>> + prep_encrypted_page(new_page, HPAGE_PMD_ORDER,
> >>> + vma_keyid(vma), false);
> >>> + }
> >>
> >> I guess this isn't horribly problematic now, but if we ever keep pools
> >> of preassigned-keyids, this won't work any more.
> >
> > I don't get this. What pools of preassigned-keyids are you talking about?
>
> My point was that if we ever teach the allocator or something _near_ the
> allocator to keep pools of pre-zeroed and/or pre-cache-cleared pages,
> this approach will need to get changed otherwise we will double-prep pages.
It shouldn't be a problem here. It's pretty slow path. We often wait
memory to be compacted before page for khugepaged gets allocated.
Double-prep shouldn't have visible impact.
> My overall concern with prep_encrypted_page() in this patch set is that
> it's inserted pretty ad-hoc. It seems easy to miss spots where it
> should be. I'm also unsure of the failure mode and anything we've done
> to ensure that if we get this wrong, we scream clearly and loudly about
> what happened. Do we do something like that?
I have debugging patch that puts BUG_ONs around set_pte_at() to check if
the page's keyid matches VMA's keyid. But that's not very systematic.
We would need something better than this.
--
Kirill A. Shutemov
