Re: WARNING in memory_failure() at include/linux/huge_mm.h:635 triggered

[Zi Yan <ziy@nvidia.com>]

On 4 Feb 2026, at 14:55, David Hildenbrand (arm) wrote:

> On 2/4/26 20:48, Zi Yan wrote:
>> On 4 Feb 2026, at 14:18, David Hildenbrand (arm) wrote:
>>
>>> On 2/4/26 18:41, Zi Yan wrote:
>>>>
>>>>
>>>> More details:
>>>> later at sg_vma_fault(), the driver just handles a page fault by supplying
>>>> a subpage from a pre-allocated compound page[3]. We then get a large folio
>>>> without !CONFIG_TRANSPARENT_HUGEPAGE.
>>>
>>> We can identify such non-folio (but compound) things by looking at PG_large_rmappable IIRC.
>>
>> What do you mean? Changing memory failure code to only handle large_rmappable?
>> large_rmappable is a folio flag, memory failure code should see such
>
> Did you mean "should not" ? :)

Yes.

>
>> non-folio but compound things to begin with, IMHO.
>
> I would say that we could right now reject in memory failure code any compound pages that do not have PG_large_rmappable set.
>
> I have the faint recollection that we don't set PG_large_rmappable on hugetlb folios yet, so they have to identified as well.

Right. My patchset[1] is trying to add it, since hugetlb is used as a folio
in most places and large_rmappable is a folio flag.


[1] https://lore.kernel.org/all/20260130034818.472804-1-ziy@nvidia.com/
>>
>> I think we need to be able to tell between raw page (compound or not),
>> mappable page (compound or not, especially for those used with vm_insert_*),
>> and folio.
>
> We can't identify (small) folios just yet. We'd need another page flag for that (just like PG_large_rmappable), and we all know how that ends ;)

Yes, I am thinking about removing mapcount in struct page to achieve that.
And only pages used for vm_insert_*() and folios need mapcount. Code
uses vm_insert_*() on pages would probably have a struct mappable_page
with mapcount.

>
> With Willy's work we'll be able to identify folios reliably.
>
> How to deal with that vm_insert_* crap, especially for non-folio pages, is also future work based on that.

I think it might the other way around. memdesc does not have mapcount,
if we do not have a separate struct for these mappable pages now,
what do we use at memdesc time? folio?


Best Regards,
Yan, Zi