Re: [PATCH 1/4] mm/memory-failure.c: fix race with changing page compound again

[Mike Kravetz <mike.kravetz@oracle.com>]

On 3/6/22 23:01, HORIGUCHI NAOYA(堀口 直也) wrote:
> On Mon, Mar 07, 2022 at 11:44:20AM +0800, Miaohe Lin wrote:
>> On 2022/3/5 3:32, Mike Kravetz wrote:
>>> On 3/4/22 00:26, HORIGUCHI NAOYA(堀口 直也) wrote:
>>>> On Mon, Feb 28, 2022 at 10:02:42PM +0800, Miaohe Lin wrote:
>>>>> There is a race window where we got the compound_head, the hugetlb page
>>>>> could be freed to buddy, or even changed to another compound page just
>>>>> before we try to get hwpoison page. If this happens, just bail out.
>>>>
>>>> I think that when some hugetlb page is about to change into other type/size
>>>> of compound page, it has to go through buddy allocator because hugetlb pages
>>>> are maintained in separate memory allocator and they never change into other
>>>> normal state directly.  memory_failure_hugetlb() takes refcount before
>>>> lock_page(), so the hugetlb page seems not change between get_hwpoison_page()
>>>> and lock_page(). So it this new check really necessary?
>>>
>>> A hugetlb page could change size without going through buddy via the new
>>> demote functionality [1].  Only hugetlb pages on the hugetlb free list can
>>> be demoted.  
>>>
>>> We should not demote a page if poison is set.  However, there is no check in
>>> the demote code.  IIUC, poison is set early in the memory error handling
>>> process, even before taking ref on page.  Demote code needs to be fixed so
>>> that poisoned pages are not demoted.  I can do that.
>>>
>>> With this change in place, then I think Naoya's statement that hugetlb pages
>>> can not change state is correct and this patch is not necessary.
>>>
>>
>> Sorry for my confusing commit words. What I mean to tell is indeed the below race:
>>   CPU 1							CPU 2
>>   memory_failure_hugetlb
>>   struct page *head = compound_head(p);
>> 							hugetlb page is freed to buddy, or
>> 							even changed to another compound page
>> 							as we haven't held the page refcnt now
>>   get_hwpoison_page -- page is not what we want now...
>>
>> Does this make sense for both of you? Many thanks for comment and reply! :)
> 
> Thanks for elaboration, I agree with you (I simply overlooked this race, sorry).

Yes, thank you.

> And please add this in the commit log.
> 
>> +
>> +       /**
>> +        * The page could have changed compound pages due to race window.
>> +        * If this happens just bail out.
>> +        */
>> +       if (!PageHuge(p) || compound_head(p) != head) {
>> +               action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED);
>> +               res = -EBUSY;
>> +               goto out;
>> +       }
> 
> Let me have one comment on the diff. The result code MF_MSG_DIFFERENT_COMPOUND
> might not fit when PageHuge is false in the check (because it's no longer a
> compound page).  Maybe you may invent another result code, or changes
> MF_MSG_DIFFERENT_COMPOUND (for example) to MF_MSG_DIFFERENT_PAGE_SIZE?
> 

Suppose we do encounter this race.  Also, suppose p != head.
At the beginning of memory_failure_hugetlb, we do:

struct page *head = compound_head(p);
...
if (TestSetPageHWPoison(head))

So, it could be that we set Poison in the 'head' page but the error was really
in another page.  Is that correct?

Now with the race, head is not a huge page and the pages could even be on
buddy.  Does this mean we could have poison set on the wrong page in buddy?

-- 
Mike Kravetz