Re: [PATCH 2/5] mm,hugetlb: Document the reason to lock the folio in the faulting path

[David Hildenbrand <david@redhat.com>]

On 17.06.25 14:04, Oscar Salvador wrote:
> On Tue, Jun 17, 2025 at 01:27:18PM +0200, David Hildenbrand wrote:
>>>    @@ -6198,6 +6198,8 @@ static vm_fault_t hugetlb_wp(struct vm_fault *vmf)
>>>     	 * in scenarios that used to work. As a side effect, there can still
>>>     	 * be leaks between processes, for example, with FOLL_GET users.
>>>     	 */
>>>    +	if (folio_test_anon(old_folio))
>>>    +		folio_lock(old_folio);
>>
>> If holding the PTL, this would not work. You'd have to unlock PTL, lock
>> folio, retake PTL, check pte_same.
> 
> Why so?
> 
> hugetlb_no_page() has already checked pte_same under PTL, then mapped the page
> and called hugetlb_wp().
> 
>   hugetlb_no_page
>    vmf->ptl = huge_pte_lock()
>    pte_same
>    set_huge_pte_at
>    hugetlb_wp
> 
> and in hugetlb_wp() we're still holding the PTL.
> Why do we have to release PTL in order to lock the folio?
> This folio can't have been unmapped because we're holding PTL, right?
> And it can't have been truncaed for the same reason.
> 
> It's because some lock-order issue?

folio lock is a sleeping lock, PTL is a spinlock. :)

-- 
Cheers,

David / dhildenb