Re: [PATCH v3 07/12] mm/debug_vm_pgtable: Use struct pgtable_debug_args in PTE modifying tests

[Gavin Shan <gshan@redhat.com>]

Hi Anshuman,

On 7/23/21 12:39 PM, Anshuman Khandual wrote:
> On 7/22/21 12:07 PM, Gavin Shan wrote:
>>
>>>>          pgdp = pgd_offset(mm, vaddr);
>>>>        p4dp = p4d_alloc(mm, pgdp, vaddr);
>>>> @@ -1272,11 +1275,11 @@ static int __init debug_vm_pgtable(void)
>>>>         * Page table modifying tests. They need to hold
>>>>         * proper page table lock.
>>>>         */
>>>> -
>>>> -    ptep = pte_offset_map_lock(mm, pmdp, vaddr, &ptl);
>>>> -    pte_clear_tests(mm, ptep, pte_aligned, vaddr, prot);
>>>> -    pte_advanced_tests(mm, vma, ptep, pte_aligned, vaddr, prot);
>>>> -    pte_unmap_unlock(ptep, ptl);
>>>> +    ptl = pte_lockptr(args.mm, args.pmdp);
>>>> +    spin_lock(ptl);
>>>> +    pte_clear_tests(&args);
>>>> +    pte_advanced_tests(&args);
>>>> +    spin_unlock(ptl);
>>>
>>> Why pte_offset_map_lock()/pte_unmap_unlock() has been dropped and
>>> spin_lock()/spin_unlock() sequence has been added ? Please dont
>>> change the tests in these patches.
>>>
>>
>> The semantics of pte_offset_map_lock() is to grab and take the lock
>> and return the PTE entry, which is mapped if needed. We already had
>> the PTE entry tracked by args->ptep in init_args(). So some of the
>> operations covered by pte_offset_map_lock() isn't needed any more
> 
> To keep the patch on purpose, please avoid this change here. But if
> required, you could send a follow up patch later.
> 

In order to use pte_offset_map_lock() and pte_unmap_unlock(), we need
a temporary variable @ptep to store the return value from pte_offset_map_lock().
The temporary variable @ptep is passed to pte_unmap_unlock(). It means
we just need the temporary variable to keep the original implementation
in this regard. I will keep pte_offset_map_lock() and pte_unmap_unlock()
in v4.

If we really want to remove the temporary variable (@ptep), we can do it
after this series gets merged.

Thanks,
Gavin