Re: [PATCH v3 1/1] mm: gup: avoid CMA page pinning by retrying migration if no migratable page

[David Hildenbrand <david@redhat.com>]

On 05.06.25 10:39, Hyesoo Yu wrote:
> On Thu, Jun 05, 2025 at 05:04:31PM +0900, Hyesoo Yu wrote:
>> Commit 1aaf8c122918 ("mm: gup: fix infinite loop within __get_longterm_locked")
>> introduced an issue where CMA pages could be pinned by longterm GUP requests.
>> This occurs when unpinnable pages are detected but the movable_page_list is empty;
>> the commit would return success without retrying, allowing unpinnable
>> pages (such as CMA) to become pinned.
>>
>> CMA pages may be temporarily off the LRU due to concurrent isolation,
>> for example when multiple longterm GUP requests are racing and therefore
>> not appear in movable_page_list. Before commit 1aaf8c, the kernel would
>> retry migration in such cases, which helped avoid accidental CMA pinning.
>>
>> The original intent of the commit was to support longterm GUP on non-LRU
>> CMA pages in out-of-tree use cases such as pKVM. However, allowing this
>> can lead to broader CMA pinning issues.
>>
>> To avoid this, the logic is restored to return -EAGAIN instead of success
>> when no folios could be collected but unpinnable pages were found.
>> This ensures that migration is retried until success, and avoids
>> inadvertently pinning unpinnable pages.
>>
>> Fixes: 1aaf8c122918 ("mm: gup: fix infinite loop within __get_longterm_locked")
>> Acked-by: David Hildenbrand <david@redhat.com>
>> Signed-off-by: Hyesoo Yu <hyesoo.yu@samsung.com>
>>
>> ---
>> We have confirmed that this regression causes CMA pages to be pinned
>> in our kernel 6.12-based environment.
>>
>> In addition to CMA allocation failures, we also observed longterm GUP failures
>> when repeatedly accessing the same VMA. Specifically, the first longterm GUP
>> call would pin a CMA page, and a second call on the same region
>> would fail the migration because the cma page was already pinned.
>>
>> After reverting commit 1aaf8c122918, the issue no longer reproduced.
>>
>> Therefore, this fix is important to ensure reliable behavior of longterm GUP
>> and CMA-backed memory, and should be backported to stable.
>> ---
>>   mm/gup.c | 28 ++++++++++++++++++++++------
>>   1 file changed, 22 insertions(+), 6 deletions(-)
>>
>> diff --git a/mm/gup.c b/mm/gup.c
>> index e065a49842a8..66193421c1d4 100644
>> --- a/mm/gup.c
>> +++ b/mm/gup.c
>> @@ -2300,14 +2300,12 @@ static void pofs_unpin(struct pages_or_folios *pofs)
>>   		unpin_user_pages(pofs->pages, pofs->nr_entries);
>>   }
>>   
>> -/*
>> - * Returns the number of collected folios. Return value is always >= 0.
>> - */
>> -static void collect_longterm_unpinnable_folios(
>> +static bool collect_longterm_unpinnable_folios(
>>   		struct list_head *movable_folio_list,
>>   		struct pages_or_folios *pofs)
>>   {
>>   	struct folio *prev_folio = NULL;
>> +	bool any_unpinnable = false;
>>   	bool drain_allow = true;
>>   	unsigned long i;
>>   
>> @@ -2321,6 +2319,8 @@ static void collect_longterm_unpinnable_folios(
>>   		if (folio_is_longterm_pinnable(folio))
>>   			continue;
>>   
>> +		any_unpinnable = true;
>> +
>>   		if (folio_is_device_coherent(folio))
>>   			continue;
>>
>> @@ -2342,6 +2342,8 @@ static void collect_longterm_unpinnable_folios(
>>   				    NR_ISOLATED_ANON + folio_is_file_lru(folio),
>>   				    folio_nr_pages(folio));
>>   	}
>> +
>> +	return any_unpinnable;
>>   }
>>   
>>   /*
>> @@ -2417,11 +2419,25 @@ migrate_longterm_unpinnable_folios(struct list_head *movable_folio_list,
>>   static long
>>   check_and_migrate_movable_pages_or_folios(struct pages_or_folios *pofs)
>>   {
>> +	bool any_unpinnable;
>> +
>>   	LIST_HEAD(movable_folio_list);
>>   
>> -	collect_longterm_unpinnable_folios(&movable_folio_list, pofs);
>> -	if (list_empty(&movable_folio_list))
>> +	any_unpinnable = collect_longterm_unpinnable_folios(&movable_folio_list, pofs);
>> +
> 
> Hi David,
> 
> While re-reviewing the v3 patch, I realized that migrate_longterm_unpinnable_folios()
> should always be called whenever unpinnable folios are present, regardless of whether
> the movable_folio_list is empty.
 > > In collect_longterm_unpinnable_folios(), if 
folio_is_device_coherent() is true,
> the folio is not added to movable_folio_list. However, such device-coherent folios
> still need to be migrated later in migrate_longterm_unpinnable_folios().

Ohh, because we cannot isolate them ... and they are always 
longterm-unpinnable.

> 
> I think the condition `list_empty(&movable_folio_list)` should be removed,
> and it might be better to revert commit 1aaf8c122918 rather than adding a new patch.
> 
> What do you think?

Yeah, with that in mind, a revert might indeed be the better option.

-- 
Cheers,

David / dhildenb