Re: [PATCH 3/6] arm64: convert between HW PTEs and SW PTE values

[Muhammad Usama Anjum <usama.anjum@arm.com>]

On 21/09/2026 9:26 am, Ryan Roberts wrote:
> On 14/09/2026 14:51, Muhammad Usama Anjum wrote:
>> __ptep_get() returns a SW PTE value, and __set_pte_nosync() accepts one.
>> When HW PTEs use a distinct hw_pte_t, directly reading or writing *ptep
>> as a SW PTE value no longer satisfies those interfaces.
>>
>> Use __pte_from_hw after READ_ONCE() to obtain a SW PTE value and add
>> __hw_pte before WRITE_ONCE() to form an HW PTE.
>>
>> Define __hw_pte for both the wrapper and alias configurations so the
>> same accessor code works in either case.
>>
>> Signed-off-by: Muhammad Usama Anjum <usama.anjum@arm.com>
>> ---
>>  arch/arm64/include/asm/pgtable.h | 4 ++--
>>  include/linux/pgtable_types.h    | 2 ++
>>  2 files changed, 4 insertions(+), 2 deletions(-)
>>
>> diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
>> index 4768ec59de555..67c4a6179e154 100644
>> --- a/arch/arm64/include/asm/pgtable.h
>> +++ b/arch/arm64/include/asm/pgtable.h
>> @@ -360,7 +360,7 @@ static inline pte_t pte_clear_uffd(pte_t pte)
>>  
>>  static inline void __set_pte_nosync(hw_pte_t *ptep, pte_t pte)
>>  {
>> -	WRITE_ONCE(*ptep, pte);
>> +	WRITE_ONCE(*ptep, __hw_pte(pte));
> 
> Would this work as an alternative? It removes the need for the new __hw_pte()
> helper. In future, converting a pte_t to a hw_pte_t won't be trivial so having a
> generic helper for it doesn't feel like a good fit.
> 
> 	WRITE_ONCE(hw_pte_val(*ptep), pte_val(pte));

Yeah, this would work. It removes the need for the generic conversion
helpers which is even better.

> 
> In general, I wonder if we should avoid defining both the __hw_pte() and
> __pte_from_hw() generic helpers. Conversion between sw and hw representations
> will become arch-specific in future and for arm64 that conversion will need
> extra information so it's can't just specialize these interfaces as they are.

Agreed.

> 
>>  }
>>  
>>  static inline void __set_pte_complete(pte_t pte)
>> @@ -381,7 +381,7 @@ static inline void __set_pte(hw_pte_t *ptep, pte_t pte)
>>  
>>  static inline pte_t __ptep_get(hw_pte_t *ptep)
>>  {
>> -	return READ_ONCE(*ptep);
>> +	return __pte_from_hw(READ_ONCE(*ptep));
> 
> 
> ...Corresponds with:
> 
> 	return __pte(READ_ONCE(hw_pte_val(*ptep)));

I'll update. 

> 
> Thanks,
> Ryan
> 
>>  }
>>  
>>  extern void __sync_icache_dcache(pte_t pteval);
>> diff --git a/include/linux/pgtable_types.h b/include/linux/pgtable_types.h
>> index ee4eace5c3e1c..b5862a16aa497 100644
>> --- a/include/linux/pgtable_types.h
>> +++ b/include/linux/pgtable_types.h
>> @@ -9,11 +9,13 @@
>>  #ifdef CONFIG_ARCH_HAS_HW_PTE_T
>>  typedef struct __hw_pte_t { pte_t __pte; } hw_pte_t;
>>  #define __pte_from_hw(pte)	((pte).__pte)
>> +#define __hw_pte(pte)  ((hw_pte_t) { (pte) })
>>  
>>  #define hw_pte_val(x)  pte_val((x).__pte)
>>  #else
>>  #define hw_pte_t pte_t
>>  #define __pte_from_hw(pte)	(pte)
>> +#define __hw_pte(pte)  (pte)
>>  
>>  #define hw_pte_val(x)  pte_val(x)
>>  #endif
>>
> 


-- 
Thanks,
Usama