Re: [PATCH] mm: migrate: fix data-race in migrate_folio_unmap()

[Jeongjun Park <aha310510@gmail.com>]

> Matthew Wilcox <willy@infradead.org> wrote:
> 
> On Tue, Sep 24, 2024 at 09:28:44AM +0900, Jeongjun Park wrote:
>>> Matthew Wilcox <willy@infradead.org> wrote:
>>> 
>>> On Mon, Sep 23, 2024 at 05:56:40PM +0200, David Hildenbrand wrote:
>>>>>> On 22.09.24 17:17, Jeongjun Park wrote:
>>>>>> I found a report from syzbot [1]
>>>>>> 
>>>>>> When __folio_test_movable() is called in migrate_folio_unmap() to read
>>>>>> folio->mapping, a data race occurs because the folio is read without
>>>>>> protecting it with folio_lock.
>>>>>> 
>>>>>> This can cause unintended behavior because folio->mapping is initialized
>>>>>> to a NULL value. Therefore, I think it is appropriate to call
>>>>>> __folio_test_movable() under the protection of folio_lock to prevent
>>>>>> data-race.
>>>>> 
>>>>> We hold a folio reference, would we really see PAGE_MAPPING_MOVABLE flip?
>>>>> Hmm
>>> 
>>> No; this shows a page cache folio getting truncated.  It's fine; really
>>> a false alarm from the tool.  I don't think the proposed patch
>>> introduces any problems, but it's all a bit meh.
>>> 
>> 
>> Well, I still don't understand why it's okay to read folio->mapping
>> without folio_lock .
> 
> Because it can't be changed in a way which changes the value of
> __folio_test_movable().  We have a refcount on the folio at this point,
> so it can't be freed.  And __folio_set_movable() happens at allocation.
> 

Thanks for the explanation. Then it seems appropriate to annotate 
data-race in __folio_test_movable() so that KCSAN ignores it.

I will apply the change and send you a new patch.

Regards,
Jeongjun Park