[PATCH V2 1/4] arm64/mm: Use pmdp_get() for PMD accesses

[Anshuman Khandual]

On Wed, Sep 23, 2026 at 01:50:34PM +0100, Ryan Roberts wrote:
> On 23/09/2026 13:22, Anshuman Khandual wrote:
> > On Wed, Sep 23, 2026 at 12:31:42PM +0100, Ryan Roberts wrote:
> >> On 23/09/2026 12:01, Anshuman Khandual wrote:
> >>> On Wed, Sep 23, 2026 at 09:32:17AM +0100, Ryan Roberts wrote:
> >>>> On 22/09/2026 07:16, Anshuman Khandual wrote:
> >>>>> Replace READ_ONCE() with pmdp_get() for PMD accesses in preparation for
> >>>>> supporting both D64 and D128 translation table formats.
> >>>>>
> >>>>> READ_ONCE() cannot currently be used for 128-bit page table entries on
> >>>>> arm64 because it does not provide the required 128-bit single-copy
> >>>>> atomicity, causing builds to fail for accesses wider than 64 bits.
> >>>>>
> >>>>> Although LDP/STP provide the required atomicity when FEAT_LSE is
> >>>>> available (as required by FEAT_D128), extending READ_ONCE() to support
> >>>>> 128-bit accesses is undesirable. READ_ONCE() is a general-purpose API,
> >>>>> so doing so could encourage other 128-bit users that would either fail
> >>>>> to build in configurations without D128 support or, if D128 becomes a
> >>>>> runtime option, silently permit tearing on systems without the required
> >>>>> hardware support.
> >>>>>
> >>>>> Instead, standardize PMD accesses on the existing page-table helpers.
> >>>>> These can be overridden on arm64 to provide 128-bit single-copy
> >>>>> atomicity when required. No functional change intended.
> >>>>
> >>>> I notice you have an unconverted READ_ONCE(*pmdp) in
> >>>> pmdp_test_and_clear_young(). Is that intentional?
> >>>
> >>> Converting the above READ_ONCE() instances as pmdp_get() cuases a build
> >>> failure as pmpd_get() which is defined in generic pgtable header is not
> >>> visible inside platform specific pgtable header, even with a forward
> >>> declaration.
> 
> I was wondering why the forward declaration doesn't work, but looks like some c
> files include asm/pgtable.h directly, which explains it. Let's assume there are
> good reasons and unpicking it so that only linux/pgtable.h includes
> asm/pgtable.h is not practical.

Agreed.

> 
> >>>
> >>> In D128 V2 series these conversions were performed later in the series,
> >>> when platform specific pxdp_get() overrides were added via ptval_get(). 
> >>>
> >>> But to avoid this, we could just move pmdp_test_and_clear_young() inside
> >>> arch/arm64/mm/mmu.c file instead. Folding in the following change builds.
> >>
> >> I'm not sure if/how this might affect performance? Probably not significant
> >> since they are for higher levels, so not called as frequently as pte helpers.
> > 
> > Right, should not have much performance impact.
> > 
> >>
> >> Another option would be to implement arm64-specific pxdp_get() helpers in
> >> asm/pgtable.h to override the generic versions?
> > 
> > For now those will still be based on READ_ONCE() just like the default ones.
> > Although trying to drop similar redundancies from other platforms :)
> > 
> > https://lore.kernel.org/linux-mm/20260923043226.331880-1-anshuman.khandual@arm.com/
> > 
> > But given D128 is going to define pxdp_get() anyway it might make sense to
> > define them now and withdraw the above series :) Seems like s390 platform
> > also has similar situation.
> 
> Given the desired end state is that arm64 defines it's own pxdp_get() helpers, I
> think the best approach is just to define them now (as READ_ONCE()).

Sure, will do that.

> 
> Thanks,
> Ryan
> 
>