Re: [PATCH] mm/thp: expose deferred split folio memory usage in meminfo

[Ye Liu]

在 2026/7/17 18:29, David Hildenbrand (Arm) 写道:
> On 7/17/26 11:54, Ye Liu wrote:
>>
>>
>> 在 2026/7/17 17:48, Lorenzo Stoakes (ARM) 写道:
>>> On Fri, Jul 17, 2026 at 11:33:05AM +0200, David Hildenbrand (Arm) wrote:
>>>>
>>>> Don't expose it to user space.
>>>
>>> Yeah please stop.
>>>
>>> We don't want this change, it exposes internal implementation details. It's a
>>> no.
>>>
>>> Next time send an RFC or a discussion thread.
>>>
>> Yes, sending an RFC would be more appropriate.                                                  
>>                                                                                                 
>> I'd also like to ask if it's possible to implement this in                                      
>> `/sys/kernel/debug/deferred_split_pages`, or are you completely against querying this statistic?
> 
> It's not pages. It's folios.
> 
> You don't know how much memory is there to reclaim. You don't know if it can
> even be reclaimed.
> 
> You don't even know if there can be any memory reclaimed *at all*, because there
> are scenarios where we add folios that are actually *not* partially mapped today.
> 
> You don't even know the relationship to memcgs.
> 
>>                                                                                                 
>> How can we quickly identify this situation?                                                     
>> I believe some people will still be confused: where did the memory go?                          
> 
> Just like other caches that are shrunk on memory pressure.
> 
> We do have a "split_deferred" and a "nr_anon_partially_mapped" counter in
> "/sys/kernel/mm/transparent_hugepage/hugepages-2048kB/", maybe that helps.
> 
Thank you, it really helped.

-- 
Thanks,
Ye Liu