On 4/18/24 9:01 PM, Matthew Wilcox wrote:
> On Thu, Apr 18, 2024 at 11:45:52AM -0700, Andrew Morton wrote:
>> It indeed appears that I can move
>>
>> mm-create-folio_flag_false-and-folio_type_ops-macros.patch
>> mm-support-page_mapcount-on-page_has_type-pages.patch
>> mm-turn-folio_test_hugetlb-into-a-pagetype.patch
>> mm-turn-folio_test_hugetlb-into-a-pagetype-fix.patch
>>
>> without merge or build issues. I added
>>
>> Fixes: 9c5ccf2db04b ("mm: remove HUGETLB_PAGE_DTOR")
>> Cc: <stable@xxxxxxxxxxxxxxx>
>>
>> to all patches.
>>
>> But a question: 9c5ccf2db04b is from August 2023. Why are we seeing
>> this issue now?
>
> We saw it earlier, we just didn't know how to fix it.
> eg December 2023:
> https://lore.kernel.org/all/ZXNhGsX32y19a2Xv@xxxxxxxxxxxxxxxxxxxx/
>
> I think there were earlier reports, but I'm not finding them now.
It's a race and needs CONFIG_DEBUG_VM to be visible, which probably makes it
rare. What seems to be new are the stress tests that have more "luck" at
hitting it? Or maybe also the changes towards more high-order folio usage
make it more likely.
