Re: [PATCH 0/2] arm64: hugetlb: Fix page fault loop for sw-dirty/hw-clean contiguous PTEs

[James Houghton]

On Wed, Dec 6, 2023 at 2:24 AM Ryan Roberts <ryan.roberts@xxxxxxx> wrote:
>
> On 05/12/2023 17:54, James Houghton wrote:
> > On Tue, Dec 5, 2023 at 6:43 AM Ryan Roberts <ryan.roberts@xxxxxxx> wrote:
> > Thanks for pointing this out. So (1) is definitely a bug. The second
> > patch in this series makes it impossible to create such a PTE via
> > pte_modify (by forcing sw-dirty PTEs to be hw-dirty as well).
>
> Yes; I think the second patch should be sufficient; I took a quick look at the
> other helpers and I don't see anything else that could get the PTE to the
> invalid state.
>
> I have a series that starts using the contpte bit for (multi-size) THP
> opportunistically. This bug will affect that too I think. Your patch #2 will fix
> for both hugetlb and my series. I'd rather not apply an equivalent to your patch
> #1 because its not quite as straightforward in my code path. But I'm pretty
> confident that patch # is all that's needed here.

There is no need to apply a patch #1-equivalent for multi-size THPs.
:) If multi-size THP has the same problem as HugeTLB, patch #2 will
fix it too. I don't think multi-size THP has the equivalent problem --
in fact, I'm not sure how multi-size THP keeps the PTE_DIRTY,
PTE_WRITE (DBM), and the PTE_RDONLY bits in sync (they do need to be
in-sync with each other when the contiguous bit is being used,
right?).

I included patch #1 (with cc:stable) because it's a more direct fix
for HugeTLB that might be slightly easier to backport. If you think
that patch #1 should be dropped and patch #2 should be backported,
please let me know.

Thanks for the review!