Re: [RFC -next] memcg: Optimize creation performance when LRU_GEN is enabled

[Chen Ridong <chenridong@huaweicloud.com>]

On 2025/11/27 1:15, Johannes Weiner wrote:
> On Wed, Nov 19, 2025 at 08:37:22AM +0000, Chen Ridong wrote:
>> From: Chen Ridong <chenridong@huawei.com>
>>
>> With LRU_GEN=y and LRU_GEN_ENABLED=n, a performance regression occurs
>> when creating a large number of memory cgroups (memcgs):
>>
>> 	# time mkdir testcg_{1..10000}
>>
>> 	real	0m7.167s
>> 	user	0m0.037s
>> 	sys	0m6.773s
>>
>> 	# time mkdir testcg_{1..20000}
>>
>> 	real	0m27.158s
>> 	user	0m0.079s
>> 	sys	0m26.270s
>>
>> In contrast, with LRU_GEN=n, creation of the same number of memcgs
>> performs better:
>>
>> 	# time mkdir testcg_{1..10000}
>>
>> 	real	0m3.386s
>> 	user	0m0.044s
>> 	sys	0m3.009s
>>
>> 	# time mkdir testcg_{1..20000}
>>
>> 	real	0m6.876s
>> 	user	0m0.075s
>> 	sys	0m6.121s
>>
>> The root cause is that lru_gen node onlining uses hlist_nulls_add_tail_rcu,
>> which traverses the entire list to find the tail. This traversal scales
>> with the number of memcgs, even when LRU_GEN is runtime-disabled.
> 
> Can you please look into removing the memcg LRU instead?
> 
> Use mem_cgroup_iter() with a reclaim cookie in shrink_many(), like we
> do in shrink_node_memcgs().
> 
> The memcg LRU is complicated, and it only works for global reclaim; if
> you have a subtree with a memory.max at the top, it'll go through
> shrink_node_memcgs() already anyway.

Hi, all,

I previously attempted to remove the memcg LRU [1], but this change introduced a regression that
increased kswapd overhead significantly.

Now circling back to this issue: does anyone have suggestions on how to address this problem
effectively?

[1] https://lore.kernel.org/cgroups/0b8ea26f-71f7-4f6d-b0d6-7d42e087a7ed@huaweicloud.com/T/#t

-- 
Best regards,
Ridong