Re: [PATCH 10/16] mm/migration: remove PG_writeback handle in folio_migrate_flags

[Miaohe Lin]

On 2022/3/7 9:21, Andrew Morton wrote:
> On Fri, 4 Mar 2022 17:34:03 +0800 Miaohe Lin <linmiaohe@xxxxxxxxxx> wrote:
> 
>> When newfolio reaches here, it's guaranteed that PG_writeback is not set
>> because caller ensures writeback must have been completed. Remove this
>> unneeded check and cleanup the relevant comment.
> 
> What guarantees that writeback cannot start after the caller has checked?

Drivers should set writeback flag under page lock. But I'am not sure this is always obeyed.

> 
> I see no such check in iomap_migrate_page()?

iomap_migrate_page acts as a migratepage callback. The check is actually done in __unmap_and_move.

> 
>> --- a/mm/migrate.c
>> +++ b/mm/migrate.c
>> @@ -549,18 +549,6 @@ void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
>>  	if (!folio_test_hugetlb(folio))
>>  		folio->private = NULL;
>>  
>> -	/*
>> -	 * If any waiters have accumulated on the new page then
>> -	 * wake them up.
>> -	 */
>> -	if (folio_test_writeback(newfolio))
>> -		folio_end_writeback(newfolio);

IIUC, this should also be broken too. If a page can be flagged writeback under page lock,
we can't simply do folio_end_writeback here as page might still be under the use of drivers.
More works should be done to handle the case correctly.

>> -
>> -	/*
>> -	 * PG_readahead shares the same bit with PG_reclaim.  The above
>> -	 * end_page_writeback() may clear PG_readahead mistakenly, so set the
>> -	 * bit after that.
>> -	 */
>>  	if (folio_test_readahead(folio))
>>  		folio_set_readahead(newfolio);
>>  
> 
> folio_migrate_flags() and folio_migrate_copy() are global,
> export-to-modules functions but have no interface documentation.  That
> was bad of us.  
> 
> I wonder why those two functions are exported to modules anyway.

That's really a pity.

> .
> 

Many thanks for your comment and reply!