[PATCH V3 1/4] arm64/mm: Use pmdp_get() for PMD accesses

[Anshuman Khandual]

On Mon, Oct 12, 2026 at 12:04:11PM +0100, Catalin Marinas wrote:
> On Wed, Oct 07, 2026 at 08:11:52AM +0530, Anshuman Khandual wrote:
> > On Mon, Oct 05, 2026 at 04:31:28PM +0100, Catalin Marinas wrote:
> > > On Thu, Sep 24, 2026 at 11:52:11AM +0530, Anshuman Khandual wrote:
> > > > diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
> > > > index 763c5a411d64..9ad2d03e86ad 100644
> > > > --- a/arch/arm64/include/asm/pgtable.h
> > > > +++ b/arch/arm64/include/asm/pgtable.h
> > > > @@ -84,6 +84,12 @@ static inline void arch_leave_lazy_mmu_mode(void)
> > > >  	arch_flush_lazy_mmu_mode();
> > > >  }
> > > >  
> > > > +#define pmdp_get pmdp_get
> > > > +static inline pmd_t pmdp_get(pmd_t *pmdp)
> > > > +{
> > > > +	return READ_ONCE(*pmdp);
> > > > +}
> > > 
> > > This is the same as the generic version. Does it need redefining?
> > 
> > Without this pmdp_get() override, replacing READ_ONCE() in pmdp_test_and_clear_young()
> > with pmdp_get() causes a build failure as pmpd_get() which is defined in the generic
> > pgtable header is not visible from inside platform specific pgtable header, even with
> > a forward declaration.
> > 
> > Earlier had moved pmdp_test_and_clear_young() inside arch/arm64/mm/mmu.c to avoid the
> > above described dependency problem. But then realised that pmdp_get() gets eventually
> > overridden on arm64 platform for D128, why not just redefine here it now as standard
> > READ_ONCE() to avoid the current build problem and the code movement.
> > 
> > Please find some context here from earlier discussions during V2.
> > 
> > https://lore.kernel.org/all/212067fe-1b16-4b87-87a7-4ce50847eb24@arm.com/
> 
> Thanks for link summary of the past discussion. I agree, for D128 we'll
> have to define our p*d_get() but I haven't decided D128 is worth merging
> yet (it depends on future hardware availability, I need to get more
> visibility into this).
> 
> For the time being, I'd keep the p*d_get() accessors generic and move
> pmdp_test_and_clear_young() to mmu.c.

OK.

Although there are similar cases involving [p4d|pud]_offset[_phys]() helpers
which would require code movement inside arch/arm64/mm/mmu.c file. Something
like the following. Will do all the necessary changes.

diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index b9427e9d54dab..e03bb7017a57e 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c
@@ -2465,4 +2465,17 @@ pud_t *pud_offset(p4d_t *p4dp, unsigned long addr)
 }
 #endif /* CONFIG_PGTABLE_LEVELS > 3 */

+#if CONFIG_PGTABLE_LEVELS > 4
+phys_addr_t p4d_offset_phys(pgd_t *pgdp, unsigned long addr)
+{
+       VM_WARN_ON_ONCE(!pgtable_l5_enabled());
+
+       return pgd_page_paddr(pgdp_get(pgdp)) + p4d_index(addr) * sizeof(p4d_t);
+}
+
+p4d_t *p4d_offset(pgd_t *pgdp, unsigned long addr)
+{
+       return p4d_offset_lockless(pgdp, pgdp_get(pgdp), addr);
+}
+#endif /* CONFIG_PGTABLE_LEVELS > 4 */
 #endif

> 
> > > >  #ifdef CONFIG_TRANSPARENT_HUGEPAGE
> > > >  #define __HAVE_ARCH_FLUSH_PMD_TLB_RANGE
> > > >  
> > > > @@ -811,7 +817,8 @@ static inline unsigned long pmd_page_vaddr(pmd_t pmd)
> > > >  }
> > > >  
> > > >  /* Find an entry in the third-level page table. */
> > > > -#define pte_offset_phys(dir,addr)	(pmd_page_paddr(READ_ONCE(*(dir))) + pte_index(addr) * sizeof(pte_t))
> > > > +#define pte_offset_phys(dir, addr)	(pmd_page_paddr(pmdp_get(dir)) + \
> > > > +					 pte_index(addr) * sizeof(pte_t))
> > > 
> > > Why not make this change in the generic version?
> > 
> > Could you please elaborate on this ? pte_offset_phys() is arm64 specific helper which
> > is also local to this header file. AFAICT there is no generic version for this helper.
> 
> You are right, I got confused.
> 
> Anyway, I'd push this series post -rc1. It depends on the mm tree and
> it's getting pretty close to the merging window.

Sure, no problem.

Will rebase this series post v7.4-rc1 and repost as V4 with all necessary changes
discussed earlier.

> 
> -- 
> Catalin