Re: [PATCH] mm/mglru: Update MG-LRU proactive reclaim statistics only to memcg

[Hao Jia <jiahao.kernel@gmail.com>]

On 2025/7/18 04:18, Yuanchu Xie wrote:

Hi Yuanchu,

> Hi Hao,
> 
> On Thu, Jul 17, 2025 at 1:29 AM Hao Jia <jiahao.kernel@gmail.com> wrote:
>>
>> From: Hao Jia <jiahao1@lixiang.com>
>>
>> Users can use /sys/kernel/debug/lru_gen to trigger proactive memory reclaim
>> of a specified memcg. Currently, statistics such as pgrefill, pgscan and
>> pgsteal will be updated to the /proc/vmstat system memory statistics.
> 
> This is a debugfs interface and it's not meant for use in production
> or provide a stable ABI. Does memory.reclaim not work for your needs?
> 

No, I am comparing the two interfaces.

Thanks for your reminder, but I want to use this interface run_aging() 
to age folios, and separate proactive memory reclamation from multiple 
walk_mm() by combining it with BIT(LRU_GEN_MM_WALK).

For example, user-space agent, enable LRU_GEN_MM_WALK, and then trigger 
run_aging(). Then turn off LRU_GEN_MM_WALK and use cgroup.reclaim to 
trigger proactive reclamation. Avoid the long latency caused by walk_mm().


Maybe it would be more reasonable to put walk_mm() in workqueues?

I don't know if my idea is reasonable, any suggestions are welcome.

Thanks,
Hao

> I'm not against the change; I just hope you don't depend on it
> continuing to exist/behave a certain way.
> 
> Shakeel's comment is accurate. The lru_gen interface uses the internal
> memcg id which is not usually used to interface with the userspace.
> Reading this file does show the cgroup path and memcg id association.
> 
>>
>> This will confuse some system memory pressure monitoring tools, making
>> it difficult to determine whether pgscan and pgsteal are caused by
>> system-level pressure or by proactive memory reclaim of some specific
>> memory cgroup.
>>
>> Therefore, make this interface behave similarly to memory.reclaim.
>> Update proactive memory reclaim statistics only to its memory cgroup.
>>
>> Signed-off-by: Hao Jia <jiahao1@lixiang.com>
> 
> The patch looks okay to me too.
> 
> Thanks,
> Yuanchu