Re: [PATCH 07/12] mm: Remove redundant pXd_devmap calls

[David Hildenbrand <david@redhat.com>]

On 02.06.25 11:33, David Hildenbrand wrote:
>> --- a/mm/huge_memory.c
>> +++ b/mm/huge_memory.c
>> @@ -1398,10 +1398,7 @@ static int insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
>>    	}
>>    
>>    	entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
>> -	if (pfn_t_devmap(pfn))
>> -		entry = pmd_mkdevmap(entry);
>> -	else
>> -		entry = pmd_mkspecial(entry);
>> +	entry = pmd_mkspecial(entry);
>>    	if (write) {
> 
> 
> I just stumbled over this, and I think there is something off here in
> the PMD/PUD case.
> 
> vmf_insert_folio_pmd() does a folio_get() + folio_add_file_rmap_pmd().
> 
> But then, we go ahead and turn this into a special mapping by setting it
> pmd_mkdevmap()/pmd_mkspecial().
> 
> Consequently, vm_normal_page_pmd() would ignore them, not following the
> rules documented for vm_normal_page() and behaving differently than
> vmf_insert_page_mkwrite()->insert_page().
> 
> 
> folio_add_file_rmap_pmd() should never set these things special/devmap
> in the first place :/
> 
> What am I missing?
> 
> Note that fs/dax.c calls vmf_insert_folio_pmd() for PMDs and
> vmf_insert_page_mkwrite() for PTEs.
> 
> Consequently, PTEs will never be marked special (corner case, shared
> zeropage), but PMDs would always.
> 
> Hm?

What an ugly piece of crap this pmd handling code is.

Just noting because I stumbled over that myself:

pmd_mkdevmap() does *not* imply pmd_special().

... in contrast to pte_mkdevmap(), which will imply pte_special().


-- 
Cheers,

David / dhildenb