Re: [RFC V1 10/16] arm64/mm: Route all pgtable writes via ptdesc_set()

[Ryan Roberts]

On 26/02/2026 12:54, Anshuman Khandual wrote:
> 
> 
> On 26/02/26 6:07 PM, Usama Arif wrote:
>> On Tue, 24 Feb 2026 10:41:47 +0530 Anshuman Khandual <anshuman.khandual@xxxxxxx> wrote:
>>
>>> Currently ptdesc_set() is defined as WRITE_ONCE() but this will change for
>>> D128 pgtable builds, for which WRITE_ONCE() is not sufficient for single
>>> copy atomicity.
>>>
>>> In future this infrastructure can be used for D128 to maintain single copy
>>> atomicity semantics with inline asm blocks.
>>>
>>> Cc: Catalin Marinas <catalin.marinas@xxxxxxx>
>>> Cc: Will Deacon <will@xxxxxxxxxx>
>>> Cc: Ryan Roberts <ryan.roberts@xxxxxxx>
>>> Cc: Mark Rutland <mark.rutland@xxxxxxx>
>>> Cc: linux-arm-kernel@xxxxxxxxxxxxxxxxxxx
>>> Cc: linux-kernel@xxxxxxxxxxxxxxx
>>> Signed-off-by: Anshuman Khandual <anshuman.khandual@xxxxxxx>
>>> ---
>>>  arch/arm64/include/asm/pgtable.h | 11 ++++++-----
>>>  arch/arm64/mm/mmu.c              |  4 ++--
>>>  mm/debug_vm_pgtable.c            |  4 ++--
>>>  3 files changed, 10 insertions(+), 9 deletions(-)
>>>
>>> diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
>>> index 804ef49aea88..42124d2f323d 100644
>>> --- a/arch/arm64/include/asm/pgtable.h
>>> +++ b/arch/arm64/include/asm/pgtable.h
>>> @@ -85,6 +85,7 @@ static inline void arch_leave_lazy_mmu_mode(void)
>>>  }
>>>  
>>>  #define ptdesc_get(x)		READ_ONCE(x)
>>> +#define ptdesc_set(x, val)	WRITE_ONCE(x, val)
>>>  
>>>  #define pmdp_get pmdp_get
>>>  static inline pmd_t pmdp_get(pmd_t *pmdp)
>>> @@ -389,7 +390,7 @@ static inline pte_t pte_clear_uffd_wp(pte_t pte)
>>>  
>>>  static inline void __set_pte_nosync(pte_t *ptep, pte_t pte)
>>>  {
>>> -	WRITE_ONCE(*ptep, pte);
>>> +	ptdesc_set(*ptep, pte);
>>>  }
>>>  
>>>  static inline void __set_pte_complete(pte_t pte)
>>> @@ -856,7 +857,7 @@ static inline void set_pmd(pmd_t *pmdp, pmd_t pmd)
>>>  	}
>>>  #endif /* __PAGETABLE_PMD_FOLDED */
>>>  
>>> -	WRITE_ONCE(*pmdp, pmd);
>>> +	ptdesc_set(*pmdp, pmd);
>>>  
>>>  	if (pmd_valid(pmd))
>>>  		queue_pte_barriers();
>>> @@ -917,7 +918,7 @@ static inline void set_pud(pud_t *pudp, pud_t pud)
>>>  		return;
>>>  	}
>>>  
>>> -	WRITE_ONCE(*pudp, pud);
>>> +	ptdesc_set(*pudp, pud);
>>>  
>>>  	if (pud_valid(pud))
>>>  		queue_pte_barriers();
>>> @@ -999,7 +1000,7 @@ static inline void set_p4d(p4d_t *p4dp, p4d_t p4d)
>>>  		return;
>>>  	}
>>>  
>>> -	WRITE_ONCE(*p4dp, p4d);
>>> +	ptdesc_set(*p4dp, p4d);
>>>  	queue_pte_barriers();
>>>  }
>>>  
>>> @@ -1120,7 +1121,7 @@ static inline void set_pgd(pgd_t *pgdp, pgd_t pgd)
>>>  		return;
>>>  	}
>>>  
>>> -	WRITE_ONCE(*pgdp, pgd);
>>> +	ptdesc_set(*pgdp, pgd);
>>>  	queue_pte_barriers();
>>>  }
>>>  
>>> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
>>> index bcf32d1a92de..ffd307c546f5 100644
>>> --- a/arch/arm64/mm/mmu.c
>>> +++ b/arch/arm64/mm/mmu.c
>>> @@ -83,7 +83,7 @@ void noinstr set_swapper_pgd(pgd_t *pgdp, pgd_t pgd)
>>>  	 * writable in the kernel mapping.
>>>  	 */
>>>  	if (rodata_is_rw) {
>>> -		WRITE_ONCE(*pgdp, pgd);
>>> +		ptdesc_set(*pgdp, pgd);
>>>  		dsb(ishst);
>>>  		isb();
>>>  		return;
>>> @@ -91,7 +91,7 @@ void noinstr set_swapper_pgd(pgd_t *pgdp, pgd_t pgd)
>>>  
>>>  	spin_lock(&swapper_pgdir_lock);
>>>  	fixmap_pgdp = pgd_set_fixmap(__pa_symbol(pgdp));
>>> -	WRITE_ONCE(*fixmap_pgdp, pgd);
>>> +	ptdesc_set(*fixmap_pgdp, pgd);
>>>  	/*
>>>  	 * We need dsb(ishst) here to ensure the page-table-walker sees
>>>  	 * our new entry before set_p?d() returns. The fixmap's
>>> diff --git a/mm/debug_vm_pgtable.c b/mm/debug_vm_pgtable.c
>>> index 83cf07269f13..faf6a19a89a1 100644
>>> --- a/mm/debug_vm_pgtable.c
>>> +++ b/mm/debug_vm_pgtable.c
>>> @@ -445,7 +445,7 @@ static void __init pmd_huge_tests(struct pgtable_debug_args *args)
>>>  	 * X86 defined pmd_set_huge() verifies that the given
>>>  	 * PMD is not a populated non-leaf entry.
>>>  	 */
>>> -	WRITE_ONCE(*args->pmdp, __pmd(0));
>>> +	ptdesc_set(*args->pmdp, __pmd(0));
>>
>> The ptdesc_set() and ptdesc_get() macros are defined in
>> arch/arm64/include/asm/pgtable.h and are arm64-specific. This change is
>> in mm/debug_vm_pgtable.c which is generic code compiled for all
>> architectures. Other architectures do not define ptdesc_set(), so this
>> will would cause a build failure on other architectures..
> 
> Agreed. Probably will add a local fallback for ptdesc_set() which
> defaults to WRITE_ONCE() for platforms not overriding. OR does it
> require making ptdesc_get/set() much more generic construct ?
> 
> Probably missed this problem while building on some other platforms
> as DEBUG_VM_PGTABLE might not have been enabled in their defconfig.

I don't think there is any need for WRITE_ONCE() in this file? (and I feel like
this has come up before in some other context). Could you just change it to a
plain C assignment? Or even better, pmd_clear()?

> 
>>
>>>  	WARN_ON(!pmd_set_huge(args->pmdp, __pfn_to_phys(args->fixed_pmd_pfn), args->page_prot));
>>>  	WARN_ON(!pmd_clear_huge(args->pmdp));
>>>  	pmd = pmdp_get(args->pmdp);
>>> @@ -465,7 +465,7 @@ static void __init pud_huge_tests(struct pgtable_debug_args *args)
>>>  	 * X86 defined pud_set_huge() verifies that the given
>>>  	 * PUD is not a populated non-leaf entry.
>>>  	 */
>>> -	WRITE_ONCE(*args->pudp, __pud(0));
>>> +	ptdesc_set(*args->pudp, __pud(0));
>>>  	WARN_ON(!pud_set_huge(args->pudp, __pfn_to_phys(args->fixed_pud_pfn), args->page_prot));
>>>  	WARN_ON(!pud_clear_huge(args->pudp));
>>>  	pud = pudp_get(args->pudp);
>>> -- 
>>> 2.43.0
>>>
>>>
>>
>