[PATCH] arm64: mm: use ptep_clear() instead of pte_clear() in clear_flush()

[Andrew Morton]

On Fri, 11 Aug 2023 19:28:41 +0800 Qi Zheng <zhengqi.arch at bytedance.com> wrote:

> Will Deacon <will at kernel.org>于2023年8月11日 周五19:21写道：
> 
> > On Fri, Aug 11, 2023 at 07:16:20PM +0800, Qi Zheng wrote:
> > >    Will Deacon <[1]will at kernel.org>于2023年8月11日 周五19:03写道：
> > >
> > >      On Thu, Aug 10, 2023 at 09:32:41AM +0000, Qi Zheng wrote:
> > >      > From: Qi Zheng <[2]zhengqi.arch at bytedance.com>
> > >      >
> > >      > In clear_flush(), the original pte may be a present entry, so we
> > >      should
> > >      > use ptep_clear() to let page_table_check track the pte clearing
> > >      operation,
> > >      > otherwise it may cause false positive in subsequent set_pte_at().
> > >
> > >      Isn't this true for most users of pte_clear()? There are some in the
> > >      core
> > >      code, so could they trigger the false positive as well?
> > >
> > >    No, the PTE entry in other places where pte_clear() is used is
> > non-present
> > >    PTE.
> > >    The page_table_check does not does track the pte operation in this
> > case,
> > >    so it will not cause false positives.
> >
> > Are you sure? For example, the call from flush_all_zero_pkmaps() in
> > highmem.c really looks like it's clearing a valid entry. Not that arm64
> > cares about highmem, but still.
> 
> 
> Ah, this is init_mm, not user mm, page_table_check does not care about this
> case.

It's unclear where we stand with this patch.  An ack or a nack, please?