Re: [PATCH] mm: Do not reclaim private data from pinned page

[David Hildenbrand <david@redhat.com>]

On 02.05.23 17:26, Peter Xu wrote:
> On Fri, Apr 28, 2023 at 02:41:40PM +0200, Jan Kara wrote:
>> If the page is pinned, there's no point in trying to reclaim it.
>> Furthermore if the page is from the page cache we don't want to reclaim
>> fs-private data from the page because the pinning process may be writing
>> to the page at any time and reclaiming fs private info on a dirty page
>> can upset the filesystem (see link below).
>>
>> Link: https://lore.kernel.org/linux-mm/20180103100430.GE4911@quack2.suse.cz
>> Signed-off-by: Jan Kara <jack@suse.cz>
>> ---
>>   mm/vmscan.c | 10 ++++++++++
>>   1 file changed, 10 insertions(+)
>>
>> This was the non-controversial part of my series [1] dealing with pinned pages
>> in filesystems. It is already a win as it avoids crashes in the filesystem and
>> we can drop workarounds for this in ext4. Can we merge it please?
>>
>> [1] https://lore.kernel.org/all/20230209121046.25360-1-jack@suse.cz/
>>
>> diff --git a/mm/vmscan.c b/mm/vmscan.c
>> index bf3eedf0209c..401a379ea99a 100644
>> --- a/mm/vmscan.c
>> +++ b/mm/vmscan.c
>> @@ -1901,6 +1901,16 @@ static unsigned int shrink_folio_list(struct list_head *folio_list,
>>   			}
>>   		}
>>   
>> +		/*
>> +		 * Folio is unmapped now so it cannot be newly pinned anymore.
>> +		 * No point in trying to reclaim folio if it is pinned.
>> +		 * Furthermore we don't want to reclaim underlying fs metadata
>> +		 * if the folio is pinned and thus potentially modified by the
>> +		 * pinning process as that may upset the filesystem.
>> +		 */
>> +		if (folio_maybe_dma_pinned(folio))
>> +			goto activate_locked;
>> +
>>   		mapping = folio_mapping(folio);
>>   		if (folio_test_dirty(folio)) {
>>   			/*
>> -- 
>> 2.35.3
>>
>>
> 
> IIUC we have similar handling for anon (feb889fb40fafc).  Should we merge
> the two sites and just move the check earlier?  Thanks,
> 

feb889fb40fafc introduced a best-effort check that is racy, as the page 
is still mapped (can still get pinned). Further, we get false positives 
most only if a page is shared very often (1024 times), which happens 
rarely with anon pages. Now that we handle COW+pinning correctly using 
PageAnonExclusive, that check only optimizes for the "already pinned" 
case. But it's not required for correctness anymore (so it can be racy).

Here, however, we want more precision, and not false positives simply 
because a page is mapped many times (which can happen easily) or can 
still get pinned while mapped.
-- 
Thanks,

David / dhildenb