On Tue, Jul 25, 2023 at 8:49 PM Yin Fengwei <fengwei.yin@xxxxxxxxx> wrote:
>
>
> On 7/25/23 13:55, Yu Zhao wrote:
> > On Fri, Jul 21, 2023 at 3:41 AM Yin Fengwei <fengwei.yin@xxxxxxxxx> wrote:
> >>
> >> Currently, in function madvise_cold_or_pageout_pte_range(), the
> >> young bit of pte/pmd is cleared notify subscripter.
> >>
> >> Using notify-able API to make sure the subscripter is signaled about
> >> the young bit clearing.
> >>
> >> Signed-off-by: Yin Fengwei <fengwei.yin@xxxxxxxxx>
> >> ---
> >> mm/madvise.c | 18 ++----------------
> >> 1 file changed, 2 insertions(+), 16 deletions(-)
> >>
> >> diff --git a/mm/madvise.c b/mm/madvise.c
> >> index f12933ebcc24..b236e201a738 100644
> >> --- a/mm/madvise.c
> >> +++ b/mm/madvise.c
> >> @@ -403,14 +403,7 @@ static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
> >> return 0;
> >> }
> >>
> >> - if (pmd_young(orig_pmd)) {
> >> - pmdp_invalidate(vma, addr, pmd);
> >> - orig_pmd = pmd_mkold(orig_pmd);
> >> -
> >> - set_pmd_at(mm, addr, pmd, orig_pmd);
> >> - tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
> >> - }
> >> -
> >> + pmdp_clear_flush_young_notify(vma, addr, pmd);
> >> folio_clear_referenced(folio);
> >> folio_test_clear_young(folio);
> >> if (folio_test_active(folio))
> >> @@ -496,14 +489,7 @@ static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
> >>
> >> VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
> >>
> >> - if (pte_young(ptent)) {
> >> - ptent = ptep_get_and_clear_full(mm, addr, pte,
> >> - tlb->fullmm);
> >> - ptent = pte_mkold(ptent);
> >> - set_pte_at(mm, addr, pte, ptent);
> >> - tlb_remove_tlb_entry(tlb, pte, addr);
> >> - }
> >> -
> >> + ptep_clear_flush_young_notify(vma, addr, pte);
> >
> > These two places are tricky.
> >
> > I agree there is a problem here, i.e., we are not consulting the mmu
> > notifier. In fact, we do pageout on VMs on ChromeOS, and it's been a
> > known problem to me for a while (not a high priority one).
> >
> > tlb_remove_tlb_entry() is batched flush, ptep_clear_flush_young() is
> > not. But, on x86, we might see a performance improvement since
> > ptep_clear_flush_young() doesn't flush TLB at all. On ARM, there might
> > be regressions though.
> >
> > I'd go with ptep_clear_young_notify(), but IIRC, Minchan mentioned he
> > prefers flush. So I'll let him chime in.
> I am OK with either way even no flush way here is more efficient for
> arm64. Let's wait for Minchan's comment.
Yes, and I don't think there would be any "negative" consequences
without tlb flushes when clearing the A-bit.
> > If we do end up with ptep_clear_young_notify(), please remove
> > mmu_gather -- it should have been done in this patch.
>
> I suppose "remove mmu_gather" means to trigger flush tlb operation in
> batched way to make sure no stale data in TLB for long time on arm64
> platform.
In madvise_cold_or_pageout_pte_range(), we only need struct
mmu_gather *tlb because of tlb_remove_pmd_tlb_entry(), i.e., flushing
tlb after clearing the A-bit. There is no correction, e.g., potential
data corruption, involved there.
