Re: [PATCH V2] mm/gup: Clear the LRU flag of a page before adding to LRU batch

[Ge Yang <yangge1116@126.com>]

在 2024/7/30 17:41, David Hildenbrand 写道:
> On 30.07.24 11:36, Ge Yang wrote:
>>
>>
>> 在 2024/7/30 15:45, David Hildenbrand 写道:
>>>>> Looking at this in more detail, I wonder if we can turn that to
>>>>>
>>>>> if (!folio_test_clear_lru(folio))
>>>>>        return;
>>>>> folio_get(folio);
>>>>>
>>>>> In all cases? The caller must hold a reference, so this should be 
>>>>> fine.
>>>>>
>>>>
>>>> Seems the caller madvise_free_pte_range(...), calling
>>>> folio_mark_lazyfree(...), doesn't hold a reference on folio.
>>>>
>>>
>>> If that would be the case and the folio could get freed concurrently,
>>> the folio_get(folio) would be completely broken.
>>>
>>> In madvise_free_pte_range() we hold the PTL, so the folio cannot get
>>> freed concurrently.
>>>
>>
>> Right.
>>
>>> folio_get() is only allowed when we are sure the folio cannot get freed
>>> concurrently, because we know there is a reference that cannot go away.
>>>
>>>
>>
>> When cpu0 runs folio_activate(), and cpu1 runs folio_put() concurrently,
>> a possible bad scenario would like:
>>
>> cpu0                                           cpu1
>>
>>                                              folio_put_testzero(folio)
>> if (!folio_test_clear_lru(folio))// Seems folio shouldn't be accessed
>>
>>          return;
>> folio_get(folio);
>>                                               __folio_put(folio)
>>                                               __folio_clear_lru(folio)
>>
>>
>> Seems we should use folio_try_get(folio) instead of folio_get(folio).
> 
> In which case is folio_activate() called without the PTL on a mapped 
> page or without a raised refcount?
> 

No such case has been found. But, folio_put() can be run at anytime, so 
folio_activate() may access a folio with a reference count of 0.