Re: [PATCH v3 12/16] mm: remember exclusively mapped anonymous pages with PG_anon_exclusive

[Vlastimil Babka <vbabka@suse.cz>]

On 4/13/22 18:39, David Hildenbrand wrote:
>>> @@ -3035,10 +3083,19 @@ void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
>>>  
>>>  	flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
>>>  	pmdval = pmdp_invalidate(vma, address, pvmw->pmd);
>>> +
>>> +	anon_exclusive = PageAnon(page) && PageAnonExclusive(page);
>>> +	if (anon_exclusive && page_try_share_anon_rmap(page)) {
>>> +		set_pmd_at(mm, address, pvmw->pmd, pmdval);
>>> +		return;
>> 
>> I am admittedly not too familiar with this code, but looks like this means
>> we fail to migrate the THP, right? But we don't seem to be telling the
>> caller, which is try_to_migrate_one(), so it will continue and not terminate
>> the walk and return false?
> 
> Right, we're not returning "false". Returning "false" would be an
> optimization to make rmap_walk_anon() fail faster.

Ah right, that's what I missed, it's an optimization and we will realize
elsewhere afterwards that the page has still mappings and we can't migrate...

> But, after all, the THP is exclusive (-> single mapping), so
> anon_vma_interval_tree_foreach() would most probably not have a lot work
> to do either way I'd assume?
> 
> In  any case, once we return from try_to_migrate(), the page will still
> be mapped.
>