Re: [PATCH v7 3/3] mm/madvise: optimize lazyfreeing with mTHP in madvise_free

[David Hildenbrand <david@redhat.com>]

>> +			nr = madvise_folio_pte_batch(addr, end, folio, pte,
>> +						     ptent, &any_young, &any_dirty);
>> +
>> +			if (nr < folio_nr_pages(folio)) {
>> +				if (folio_likely_mapped_shared(folio))
>> +					continue;
>> +
>> +				arch_leave_lazy_mmu_mode();
>> +				if (madvise_pte_split_folio(mm, pmd, addr,
>> +							    folio, &start_pte, &ptl))
>> +					nr = 0;
>> +				if (!start_pte)
>> +					break;
>> +				pte = start_pte;
>> +				arch_enter_lazy_mmu_mode();
>> +				continue;
>> +			}
>> +
>> +			if (any_young)
>> +				ptent = pte_mkyoung(ptent);
>> +			if (any_dirty)
>> +				ptent = pte_mkdirty(ptent);
>>   		}
>>   
>> +		if (folio_mapcount(folio) != folio_nr_pages(folio))
>> +			continue;
> 
> Why is this here? I thought we had previously concluded to only do this test
> inside the below if statement (where you have it duplicated).

I stumbled over these same while reviewing. It's not exactly duplicate, 
because it's unreliable without the folio lock. It looks more like an 
best-effort early check.

But then, we also add it to cases where we previously wouldn't check the 
mapcount at all: when the folio was added to the swapcache or is already 
dirty.

In that case, we would even see a change for order-0 folios with that 
new check.

-- 
Cheers,

David / dhildenb