Re: [PATCH v2 3/3] mm/huge_memory: don't mark refcounted folios special in vmf_insert_folio_pud()

[David Hildenbrand <david@redhat.com>]

On 12.06.25 19:08, Lorenzo Stoakes wrote:
> On Thu, Jun 12, 2025 at 07:00:01PM +0200, David Hildenbrand wrote:
>> On 12.06.25 18:49, Lorenzo Stoakes wrote:
>>> On Wed, Jun 11, 2025 at 02:06:54PM +0200, David Hildenbrand wrote:
>>>> Marking PUDs that map a "normal" refcounted folios as special is
>>>> against our rules documented for vm_normal_page().
>>>
>>> Might be worth referring to specifically which rule. I'm guessing it's the
>>> general one of special == don't touch (from vm_normal_page() comment):
>>>
>>> /*
>>>    * vm_normal_page -- This function gets the "struct page" associated with a pte.
>>>    *
>>>    * "Special" mappings do not wish to be associated with a "struct page" (either
>>>    * it doesn't exist, or it exists but they don't want to touch it). In this
>>>    * case, NULL is returned here. "Normal" mappings do have a struct page.
>>>    *
>>>    * ...
>>>    *
>>>    */
>>
>> Well, yes, the one vm_normal_page() is all about ... ? :)
> 
> Lol yes to be fair that is pretty obvious...
> 
>>
>>>
>>> But don't we already violate this E.g.:
>>>
>>> 		if (vma->vm_ops && vma->vm_ops->find_special_page)
>>> 			return vma->vm_ops->find_special_page(vma, addr);
>>>> I mean this in itself perhaps means we should update this comment to say
>> 'except
>>> when file-backed and there is a find_special_page() hook'.
>>
>> I rather hope we severely break this case such that we can remove that hack.
>>
>> Read as in: I couldn't care less about this XEN hack, in particular, not
>> documenting it.
>>
>> I was already wondering about hiding it behind a XEN config so not each and
>> every sane user of this function has to perform this crappy-hack check.
> 
> Yeah, I'm not a fan of generalised hooks if they can be avoided, especially ones
> where you pass critical data structures like VMAs.
> 
> It means you can, in theory, make no assumptions about what the caller does and
> yeah.
> 
> To do this for such a stupid edge case is ridiculous.

Also, I am not sure if this works at all as intended. I want to look 
into cleaning that up ...

When we inserted the page, we sure must have taken a reference, but when 
we inserted it we set pte_special() and ... didn't take a reference? Hmmmm

> 
>>
>> [...]
>>
>>>>    	}
>>>>
>>>> -	entry = pud_mkhuge(pfn_t_pud(pfn, prot));
>>>> -	if (pfn_t_devmap(pfn))
>>>> -		entry = pud_mkdevmap(entry);
>>>> -	else
>>>> -		entry = pud_mkspecial(entry);
>>>> +	if (fop.is_folio) {
>>>> +		entry = folio_mk_pud(fop.folio, vma->vm_page_prot);
>>>> +
>>>> +		folio_get(fop.folio);
>>>> +		folio_add_file_rmap_pud(fop.folio, &fop.folio->page, vma);
>>>> +		add_mm_counter(mm, mm_counter_file(fop.folio), HPAGE_PUD_NR);
>>>
>>> Nit, but might be nice to abstract for PMD/PUD.
>>
>> Which part exactly? Likely a follow-up if it should be abstracted.
> 
> Ah on second thoughts it doesn't matter, because you're using pud variants of
> everything such that it wouldn't be worth it.
> 
> Disregard this ;)

Ah, I was already suspecting that you might have missed the sneaky _pud :)

-- 
Cheers,

David / dhildenb