On Wed, Aug 31, 2022 at 12:43 PM Yang Shi <shy828301@xxxxxxxxx> wrote:
>
> On Wed, Aug 31, 2022 at 12:36 PM David Hildenbrand <david@xxxxxxxxxx> wrote:
> >
> > On 31.08.22 21:34, Yang Shi wrote:
> > > On Wed, Aug 31, 2022 at 12:15 PM David Hildenbrand <david@xxxxxxxxxx> wrote:
> > >>
> > >> On 31.08.22 21:08, Yang Shi wrote:
> > >>> On Wed, Aug 31, 2022 at 11:29 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
> > >>>>
> > >>>> On 31.08.22 19:55, Yang Shi wrote:
> > >>>>> On Wed, Aug 31, 2022 at 1:30 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
> > >>>>>>
> > >>>>>> The comment is stale, because a TLB flush is no longer sufficient and
> > >>>>>> required to synchronize against concurrent GUP-fast. This used to be true
> > >>>>>> in the past, whereby a TLB flush would have implied an IPI on architectures
> > >>>>>> that support GUP-fast, resulting in GUP-fast that disables local interrupts
> > >>>>>> from completing before completing the flush.
> > >>>>>
> > >>>>> Hmm... it seems there might be problem for THP collapse IIUC. THP
> > >>>>> collapse clears and flushes pmd before doing anything on pte and
> > >>>>> relies on interrupt disable of fast GUP to serialize against fast GUP.
> > >>>>> But if TLB flush is no longer sufficient, then we may run into the
> > >>>>> below race IIUC:
> > >>>>>
> > >>>>> CPU A CPU B
> > >>>>> THP collapse fast GUP
> > >>>>>
> > >>>>> gup_pmd_range() <-- see valid pmd
> > >>>>>
> > >>>>> gup_pte_range() <-- work on pte
> > >>>>> clear pmd and flush TLB
> > >>>>> __collapse_huge_page_isolate()
> > >>>>> isolate page <-- before GUP bump refcount
> > >>>>>
> > >>>>> pin the page
> > >>>>> __collapse_huge_page_copy()
> > >>>>> copy data to huge page
> > >>>>> clear pte (don't flush TLB)
> > >>>>> Install huge pmd for huge page
> > >>>>>
> > >>>>> return the obsolete page
> > >>>>
> > >>>> Hm, the is_refcount_suitable() check runs while the PTE hasn't been
> > >>>> cleared yet. And we don't check if the PMD changed once we're in
> > >>>> gup_pte_range().
> > >>>
> > >>> Yes
> > >>>
> > >>>>
> > >>>> The comment most certainly should be stale as well -- unless there is
> > >>>> some kind of an implicit IPI broadcast being done.
> > >>>>
> > >>>> 2667f50e8b81 mentions: "The RCU page table free logic coupled with an
> > >>>> IPI broadcast on THP split (which is a rare event), allows one to
> > >>>> protect a page table walker by merely disabling the interrupts during
> > >>>> the walk."
> > >>>>
> > >>>> I'm not able to quickly locate that IPI broadcast -- maybe there is one
> > >>>> being done here (in collapse) as well?
> > >>>
> > >>> The TLB flush may call IPI. I'm supposed it is arch dependent, right?
> > >>> Some do use IPI, some may not.
> > >>
> > >> Right, and the whole idea of the RCU GUP-fast was to support
> > >> architectures that don't do it. x86-64 does it. IIRC, powerpc doesn't do
> > >> it -- but maybe it does so for PMDs?
> > >
> > > It looks powerpc does issue IPI for pmd flush. But arm64 doesn't IIRC.
> > >
> > > So maybe we should implement pmdp_collapse_flush() for those arches to
> > > issue IPI.
> >
> > ... or find another way to detect and handle this in GUP-fast?
> >
> > Not sure if, for handling PMDs, it could be sufficient to propagate the
> > pmdp pointer + value and double check that the values didn't change.
>
> Should work too, right before pinning the page.
I actually mean the same place for checking pte. So, something like:
diff --git a/mm/gup.c b/mm/gup.c
index 5abdaf487460..2b0703403902 100644
--- a/mm/gup.c
+++ b/mm/gup.c
@@ -2392,7 +2392,8 @@ static int gup_pte_range(pmd_t pmd, unsigned
long addr, unsigned long end,
goto pte_unmap;
}
- if (unlikely(pte_val(pte) != pte_val(*ptep))) {
+ if (unlikely(pmd_val(pmd) != pmd_val(*pmdp)) ||
+ unlikely(pte_val(pte) != pte_val(*ptep))) {
gup_put_folio(folio, 1, flags);
goto pte_unmap;
}
It doesn't build, just shows the idea.
>
> pmdp_collapse_flush() is actually just called by khugepaged, so arch
> specific implementation should not be a problem and we avoid making
> gup fast more complicated.
>
> >
> > --
> > Thanks,
> >
> > David / dhildenb
> >
