Re: [RFC PATCH v2 2/4] madvise: Use notify-able API to clear and flush page table entries

[Yu Zhao]

On Wed, Jul 26, 2023 at 12:21 AM Yin Fengwei <fengwei.yin@xxxxxxxxx> wrote:
>
>
> On 7/26/23 13:40, Yu Zhao wrote:
> > On Tue, Jul 25, 2023 at 10:44 PM Yin Fengwei <fengwei.yin@xxxxxxxxx> wrote:
> >>
> >>
> >> On 7/26/23 11:26, Yu Zhao wrote:
> >>> 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.
> >>
> >> From https://lore.kernel.org/lkml/20181029105515.GD14127@xxxxxxx/,
> >> the reason that arm64 didn't drop whole flush tlb in ptep_clear_flush_young()
> >> is to prevent the stale data in TLB. I suppose there is no correction issue
> >> there also.
> >>
> >> So why keep stale data in TLB in madvise_cold_or_pageout_pte_range() is fine?
> >
> > Sorry, I'm not sure I understand your question here.
> Oh. Sorry for the confusion. I will explain my understanding and
> question in detail.
>
> >
> > In this patch, you removed tlb_remove_tlb_entry(), so we don't need
> > struct mmu_gather *tlb any more.
> Yes. You are right.
>
> >
> > If you are asking why I prefer ptep_clear_young_notify() (no flush),
> > which also doesn't need tlb_remove_tlb_entry(), then the answer is
> > that the TLB size doesn't scale like DRAM does: the gap has been
> > growing exponentially. So there is no way TLB can hold stale entries
> > long enough to cause a measurable effect on the A-bit. This isn't a
> > conjecture -- it's been proven conversely: we encountered bugs (almost
> > every year) caused by missing TLB flushes and resulting in data
> > corruption. They were never easy to reproduce, meaning stale entries
> > never stayed long in TLB.
>
> when I read https://lore.kernel.org/lkml/20181029105515.GD14127@xxxxxxx/,
>
> my understanding is that arm64 still keep something in ptep_clear_flush_young.
> The reason is finishing TLB flush by next context-switch to make sure no
> stale entries in TLB cross next context switch.
>
> Should we still keep same behavior (no stable entries in TLB cross next
> context switch) for madvise_cold_or_pageout_pte_range()?
>
> So two versions work (I assume we should keep same behavior) for me:
>   1. using xxx_flush_xxx() version. but with possible regression on arm64.
>   2. using none flush version. But do batched TLB flush.

I see. I don't think we need to flush at all, i.e., the no flush
version *without* batched TLB flush. So far nobody can actually prove
that flushing TLB while clearing the A-bit is beneficial, not even in
theory :)