Re: [PATCH 2/2] mm/memory-failure.c: not necessary to recalculate hpage

[David Hildenbrand]

On 02.12.19 23:28, Wei Yang wrote:
> On Wed, Nov 20, 2019 at 04:07:38PM +0100, David Hildenbrand wrote:
>> On 18.11.19 09:20, Wei Yang wrote:
>>> hpage is not changed.
>>>
>>> Signed-off-by: Wei Yang <richardw.yang@xxxxxxxxxxxxxxx>
>>> ---
>>>   mm/memory-failure.c | 1 -
>>>   1 file changed, 1 deletion(-)
>>>
>>> diff --git a/mm/memory-failure.c b/mm/memory-failure.c
>>> index 392ac277b17d..9784f4339ae7 100644
>>> --- a/mm/memory-failure.c
>>> +++ b/mm/memory-failure.c
>>> @@ -1319,7 +1319,6 @@ int memory_failure(unsigned long pfn, int flags)
>>>   		}
>>>   		unlock_page(p);
>>>   		VM_BUG_ON_PAGE(!page_count(p), p);
>>> -		hpage = compound_head(p);
>>>   	}
>>>   	/*
>>>
>>
>> I am *absolutely* no transparent huge page expert (sorry :) ), but won't the
>> split_huge_page(p) eventually split the compound page, such that
>> compound_head(p) will return something else after that call?
>>
> 
> Hi, David
> 
> Took sometime to look into the code and re-think about it. Found maybe we can
> simplify this in another way.
> 
> First, code touches here means split_huge_page() succeeds and "p" is now a PTE
> page. So compound_head(p) == p.

While this would also be my intuition, I can't state that this is
guaranteed to be the case (IOW, I did not check the code/documentation) :)

> 
> Then let's look at who will use hpage in the following function. There are two
> uses in current upstream:
> 
>     * page_flags calculation
>     * hwpoison_user_mappings()
> 
> The first one would be removed in next patch since PageHuge is handled at the
> beginning.
> 
> And in the second place, comment says if split succeeds, hpage points to page
> "p".
> 
> After all, we don't need to re-calculate hpage after split, and just replace
> hpage in hwpoison_user_mappings() with p is enough.

That assumption would only be true in case all compound pages at this
point are transparent huge pages, no? AFAIK that is not necessarily
true. Or am I missing something?


-- 
Thanks,

David / dhildenb