Re: [PATCH] mm/huge_memory: fix swap entry values of tail pages of THP

[Zi Yan <ziy@nvidia.com>]

[-- Attachment #1: Type: text/plain, Size: 5527 bytes --]

On 27 Feb 2024, at 10:01, David Hildenbrand wrote:

> On 27.02.24 15:52, Zi Yan wrote:
>> On 27 Feb 2024, at 9:11, David Hildenbrand wrote:
>>
>>> On 14.02.24 15:34, David Hildenbrand wrote:
>>>> On 14.02.24 15:18, Matthew Wilcox wrote:
>>>>> On Wed, Feb 14, 2024 at 12:04:10PM +0530, Charan Teja Kalla wrote:
>>>>>>> 1) Is it broken in 5.15? Did you actually try to reproduce or is this
>>>>>>>       just a guess?
>>>>>>>
>>>>>>
>>>>>> We didn't run the tests with THP enabled on 5.15, __so we didn't
>>>>>> encounter this issue__ on older to 6.1 kernels.
>>>>>>
>>>>>> I mentioned that issue exists is based on my understanding after code
>>>>>> walk through. To be specific, I just looked to the
>>>>>> migrate_pages()->..->migrate_page_move_mapping() &
>>>>>> __split_huge_page_tail() where the ->private field of thp sub-pages is
>>>>>> not filled with swap entry. If it could have set, I think these are the
>>>>>> only places where it would have done, per my understanding. CMIW.
>>>>>
>>>>> I think you have a misunderstanding.  David's patch cfeed8ffe55b (part
>>>>> of 6.6) _stopped_ us using the tail ->private entries.  So in 6.1, these
>>>>> tail pages should already have page->private set, and I don't understand
>>>>> what you're fixing.
>>>>
>>>> I think the issue is, that migrate_page_move_mapping() /
>>>> folio_migrate_mapping() would update ->private for a folio in the
>>>> swapcache (head page)
>>>>
>>>> 	newfolio->private = folio_get_private(folio);
>>>>
>>>> but not the ->private of the tail pages.
>>>>
>>>> So once you migrate a THP that is in the swapcache, ->private of the
>>>> tail pages would not be migrated and, therefore, be stale/wrong.
>>>>
>>>> Even before your patch that was the case.
>>>>
>>>> Looking at migrate_page_move_mapping(), we had:
>>>>
>>>> 	if (PageSwapBacked(page)) {
>>>> 		__SetPageSwapBacked(newpage);
>>>> 		if (PageSwapCache(page)) {
>>>> 			SetPageSwapCache(newpage);
>>>> 			set_page_private(newpage, page_private(page));
>>>> 		}
>>>> 	} else {
>>>> 		VM_BUG_ON_PAGE(PageSwapCache(page), page);
>>>> 	}
>>>>
>>>>
>>>> I don't immediately see where the tail pages would similarly get updated
>>>> (via set_page_private).
>>>>
>>>> With my patch the problem is gone, because the tail page entries don't
>>>> have to be migrated, because they are unused.
>>>>
>>>>
>>>> Maybe this was an oversight from THP_SWAP -- 38d8b4e6bdc8 ("mm, THP,
>>>> swap: delay splitting THP during swap out").
>>>>
>>>> It did update __add_to_swap_cache():
>>>>
>>>> for (i = 0; i < nr; i++) {
>>>>            set_page_private(page + i, entry.val + i);
>>>>            error = radix_tree_insert(&address_space->page_tree,
>>>>                                      idx + i, page + i);
>>>>            if (unlikely(error))
>>>>                    break;
>>>> }
>>>>
>>>> and similarly __delete_from_swap_cache().
>>>>
>>>> But I don't see any updates to migration code.
>>>>
>>>> Now, it could be that THP migration was added later (post 2017), in that
>>>> case the introducing commit would not have been 38d8b4e6bdc8.
>>>>
>>>
>>> Let's continue:
>>>
>>> The introducing commit is likely either
>>>
>>> (1) 38d8b4e6bdc87 ("mm, THP, swap: delay splitting THP during swap out")
>>>
>>> That one added THP_SWAP, but THP migration wasn't supported yet AFAIKS.
>>>
>>> -> v4.13
>>>
>>> (2) 616b8371539a6 ("mm: thp: enable thp migration in generic path")
>>
>> I think this is the one, since it makes THP entering migrate_page_move_mapping()
>> possible.
>>
>>>
>>> Or likely any of the following that actually allocate THP for migration:
>>>
>>> 8135d8926c08e mm: memory_hotplug: memory hotremove supports thp migration
>>> e8db67eb0ded3 mm: migrate: move_pages() supports thp migration
>>> c8633798497ce mm: mempolicy: mbind and migrate_pages support thp migration
>>>
>>> That actually enable THP migration.
>>>
>>> -> v4.14
>>>
>>>
>>> So likely we'd have to fix the stable kernels:
>>>
>>> 4.19
>>> 5.4
>>> 5.10
>>> 5.15
>>> 6.1
>>>
>>> That's a lot of pre-folio code. A backport of my series likely won't really make any sense.
>>>
>>> Staring at 4.19.307 code base, we likely have to perform a stable-only fix that properly handles the swapcache of compoud pages in migrate_page_move_mapping().
>>
>> Something like (applies to v4.19.307):
>>
>> diff --git a/mm/migrate.c b/mm/migrate.c
>> index 171573613c39..59878459c28c 100644
>> --- a/mm/migrate.c
>> +++ b/mm/migrate.c
>> @@ -514,8 +514,13 @@ int migrate_page_move_mapping(struct address_space *mapping,
>>          if (PageSwapBacked(page)) {
>>                  __SetPageSwapBacked(newpage);
>>                  if (PageSwapCache(page)) {
>> +                       int i;
>> +
>>                          SetPageSwapCache(newpage);
>> -                       set_page_private(newpage, page_private(page));
>> +                       for (i = 0; i < (1 << compound_order(page)); i++) {
>> +                               set_page_private(newpage + i,
>> +                                                page_private(page + i));
>> +                       }
>>                  }
>>          } else {
>>                  VM_BUG_ON_PAGE(PageSwapCache(page), page);
>
> I'm wondering if there is a swapcache update missing as well.

It seems that e71769ae5260 ("mm: enable thp migration for shmem thp") fixed
swapcache entry part, starting from v4.19.

--
Best Regards,
Yan, Zi

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 854 bytes --]