Re: [PATCH] mm/ptdump: Replace READ_ONCE() with standard page table accessors

[Dev Jain <dev.jain@arm.com>]

On 30/09/25 8:13 pm, David Hildenbrand wrote:
> On 30.09.25 09:00, Anshuman Khandual wrote:
>>
>>
>> On 30/09/25 12:11 PM, David Hildenbrand wrote:
>>> On 30.09.25 06:37, Dev Jain wrote:
>>>>
>>>> On 30/09/25 8:22 am, Anshuman Khandual wrote:
>>>>> Replace READ_ONCE() with standard page table accessors i.e 
>>>>> pxdp_get() which
>>>>> anyways default into READ_ONCE() in cases where platform does not 
>>>>> override.
>>>>>
>>>>> Cc: Andrew Morton <akpm@linux-foundation.org>
>>>>> Cc: David Hildenbrand <david@redhat.com>
>>>>> Cc: linux-mm@kvack.org
>>>>> Cc: linux-kernel@vger.kernel.org
>>>>> Signed-off-by: Anshuman Khandual <anshuman.khandual@arm.com>
>>>>> ---
>>>>>     mm/ptdump.c | 8 ++++----
>>>>>     1 file changed, 4 insertions(+), 4 deletions(-)
>>>>>
>>>>> diff --git a/mm/ptdump.c b/mm/ptdump.c
>>>>> index b600c7f864b8..18861501b533 100644
>>>>> --- a/mm/ptdump.c
>>>>> +++ b/mm/ptdump.c
>>>>> @@ -31,7 +31,7 @@ static int ptdump_pgd_entry(pgd_t *pgd, unsigned 
>>>>> long addr,
>>>>>                     unsigned long next, struct mm_walk *walk)
>>>>>     {
>>>>>         struct ptdump_state *st = walk->private;
>>>>> -    pgd_t val = READ_ONCE(*pgd);
>>>>> +    pgd_t val = pgdp_get(pgd);
>>>>>        #if CONFIG_PGTABLE_LEVELS > 4 && \
>>>>>             (defined(CONFIG_KASAN_GENERIC) || 
>>>>> defined(CONFIG_KASAN_SW_TAGS))
>>>>> @@ -54,7 +54,7 @@ static int ptdump_p4d_entry(p4d_t *p4d, unsigned 
>>>>> long addr,
>>>>>                     unsigned long next, struct mm_walk *walk)
>>>>>     {
>>>>>         struct ptdump_state *st = walk->private;
>>>>> -    p4d_t val = READ_ONCE(*p4d);
>>>>> +    p4d_t val = p4dp_get(p4d);
>>>>>        #if CONFIG_PGTABLE_LEVELS > 3 && \
>>>>>             (defined(CONFIG_KASAN_GENERIC) || 
>>>>> defined(CONFIG_KASAN_SW_TAGS))
>>>>> @@ -77,7 +77,7 @@ static int ptdump_pud_entry(pud_t *pud, unsigned 
>>>>> long addr,
>>>>>                     unsigned long next, struct mm_walk *walk)
>>>>>     {
>>>>>         struct ptdump_state *st = walk->private;
>>>>> -    pud_t val = READ_ONCE(*pud);
>>>>> +    pud_t val = pudp_get(pud);
>>>>>        #if CONFIG_PGTABLE_LEVELS > 2 && \
>>>>>             (defined(CONFIG_KASAN_GENERIC) || 
>>>>> defined(CONFIG_KASAN_SW_TAGS))
>>>>> @@ -100,7 +100,7 @@ static int ptdump_pmd_entry(pmd_t *pmd, 
>>>>> unsigned long addr,
>>>>>                     unsigned long next, struct mm_walk *walk)
>>>>>     {
>>>>>         struct ptdump_state *st = walk->private;
>>>>> -    pmd_t val = READ_ONCE(*pmd);
>>>>> +    pmd_t val = pmdp_get(pmd);
>>>>
>>>> I believe this should go through pmdp_get_lockless(). I can see in 
>>>> pgtable.h that
>>>> some magic is required on some arches to decode the pmd correctly 
>>>> in case walking
>>>> without locks.
>>>
>>> pmdp_get_lockless() is a nasty thingy to handle selected 32bit 
>>> architectures.
>>>
>>> But given that we're using ptep_get_lockless() in ptdump_pmd_entry() 
>>> it probably wouldn't hurt to use pmdp_get_lockless() here.
>>>
>>> Staring at ARCH_HAS_PTDUMP, I don't think any 32bit arch would 
>>> actually end up compiling ptdump.c.
>>>
>>> E.g., on x86 only X86_64 ends up selecting ARCH_HAS_PTDUMP.
>>>
>>
>> pxdp_get_lockless() not really required here, let's stick with 
>> pxdp_get() instead.
>
> I'd suggest that we keep it consistent. That is, also removing the 
> ptep_get_lockless() if not really required.

So in theory, the _lockless variant is not required, but surely 
consistency would mean that since we are

dereferencing the pte without the PTL, we should use the _lockless variant?