Re: [PATCH 6/9] mm: convert PTE table entry to pte

[Muhammad Usama Anjum <usama.anjum@xxxxxxx>]

On 10/08/2026 7:44 am, Alexander Gordeev wrote:
> On Fri, Aug 07, 2026 at 05:26:04PM +0100, Muhammad Usama Anjum wrote:
>> On 07/08/2026 7:58 am, Alexander Gordeev wrote:
>>> On Thu, Aug 06, 2026 at 09:38:44AM +0100, Muhammad Usama Anjum wrote:
>>>> The non-MMU stub receives hw_pte_t but returns a logical pte_t
>>>> value. Convert the stored entry through __pte_from_hw() before
>>>> returning.
>>>>
>>>> Signed-off-by: Muhammad Usama Anjum <usama.anjum@xxxxxxx>
>>>> ---
>>>>  include/linux/hugetlb.h | 2 +-
>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>>
>>>> diff --git a/include/linux/hugetlb.h b/include/linux/hugetlb.h
>>>> index bc0b9c65aa1d0..9e8b391aa4bc9 100644
>>>> --- a/include/linux/hugetlb.h
>>>> +++ b/include/linux/hugetlb.h
>>>> @@ -1283,7 +1283,7 @@ static inline pte_t huge_ptep_clear_flush(struct 
>>>> vm_area_struct *vma,
>>>>  #ifdef CONFIG_MMU
>>>>    return ptep_get(ptep);
>>>>  #else
>>>> -  return *ptep;
>>>> +  return __pte_from_hw(*ptep);
>>>
>>> But this is a direct dereferencing, which breaks the whole point, isn't it?
>> Yes, this is particular line is for non MMU. In this case, 
>> CONIFG_ARCH_HAS_HW_PTE
>> would never be defined. Hence hw_pte_t is just pte_t and direct dereference 
>> is
>> allowed. I'd thought a lot about it; is better to leave direct dereference 
>> here
>> or use some helper. Then used __pte_from_hw() was already being used in 
>> generic
>> ptep_get().
> 
> But in case CONIFG_ARCH_HAS_HW_PTE=n __pte_from_hw() is still gets called.
> That looks inconsistent to me. Why not just call ptep_deref() (see below)?

Agreed. Calling __pte_from_hw() directly exposes the representation
conversion at the call site. I will introduce ptep_deref() and use it
here.

> 
>> There are only two users of __pte_from_hw() at this time. 
>>
>>>
>>> What about introducing something like pte_t ptep_get_sw(hw_pte_t *ptep)
>>> to be used in exactly situations like this? With that the semantics of
>>> hw_pte_t pointers becomes straightforward and closes the still ongoing
>>> "storage vs lifetime" discussion:
>>>
>>> hw_pte_t*     points to HW-formatted page table entries
>>>
>>> ptep_get()    is used to obtain HW-linked/attached entries, and may wire
>>>               extra code like [1] or [2]
>>>
>>> ptep_get_sw() is used to obtain HW-unlinked/unattached entries and in
>>>               most cases is just a direct dereference
>> ptep_get_sw() or ptep_get_deref() is better name here?
> 
> ptep_deref() would be it.
> 
> Do you agree to the suggested API requirements?

Yes. hw_pte_t * identifies storage containing hardware-formatted PTEs,
regardless of whether it is attached. ptep_get() is used for attached
entries and may provide additional architecture-specific handling.
ptep_deref() is used for unattached entries and performs only the raw
storage-to-value conversion.

For review, this patch would become:

diff --git a/include/linux/hugetlb.h b/include/linux/hugetlb.h
index bc0b9c65aa1d0..ce900d2652d91 100644
--- a/include/linux/hugetlb.h
+++ b/include/linux/hugetlb.h
@@ -1283,7 +1283,7 @@ static inline pte_t huge_ptep_clear_flush(struct 
vm_area_struct *vma,
 #ifdef CONFIG_MMU
        return ptep_get(ptep);
 #else
-       return *ptep;
+       return ptep_deref(ptep);
 #endif
 }
 
diff --git a/include/linux/pgtable.h b/include/linux/pgtable.h
index 1768421755a9c..08613593f3320 100644
--- a/include/linux/pgtable.h
+++ b/include/linux/pgtable.h
@@ -490,6 +490,13 @@ static inline int pudp_set_access_flags(struct 
vm_area_struct *vma,
 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 #endif
 
+#ifndef ptep_deref
+static inline pte_t ptep_deref(hw_pte_t *ptep)
+{
+       return __pte_from_hw(*ptep);
+}
+#endif
+
 #ifndef ptep_get
 static inline pte_t ptep_get(hw_pte_t *ptep)
 {

-- 
Thanks,
Usama