Re: [RFC 2/5] memcontrol: add boot option to enable memsw account on dfl

[Roman Gushchin <roman.gushchin@linux.dev>]

Johannes Weiner <hannes@cmpxchg.org> writes:

> On Wed, Mar 19, 2025 at 10:30:20PM +0000, Roman Gushchin wrote:
>> Shakeel Butt <shakeel.butt@linux.dev> writes:
>> 
>> > On Wed, Mar 19, 2025 at 02:41:45PM +0800, Jingxiang Zeng wrote:
>> >> From: Zeng Jingxiang <linuszeng@tencent.com>
>> >> 
>> >> Added cgroup.memsw_account_on_dfl startup parameter, which
>> >> is off by default. When enabled in cgroupv2 mode, the memory
>> >> accounting mode of swap will be reverted to cgroupv1 mode.
>> >> 
>> >> Signed-off-by: Zeng Jingxiang <linuszeng@tencent.com>
>> >> ---
>> >>  include/linux/memcontrol.h |  4 +++-
>> >>  mm/memcontrol.c            | 11 +++++++++++
>> >>  2 files changed, 14 insertions(+), 1 deletion(-)
>> >> 
>> >> diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
>> >> index dcb087ee6e8d..96f2fad1c351 100644
>> >> --- a/include/linux/memcontrol.h
>> >> +++ b/include/linux/memcontrol.h
>> >> @@ -62,10 +62,12 @@ struct mem_cgroup_reclaim_cookie {
>> >>  
>> >>  #ifdef CONFIG_MEMCG
>> >>  
>> >> +DECLARE_STATIC_KEY_FALSE(memsw_account_on_dfl);
>> >>  /* Whether enable memory+swap account in cgroupv2 */
>> >>  static inline bool do_memsw_account_on_dfl(void)
>> >>  {
>> >> -	return IS_ENABLED(CONFIG_MEMSW_ACCOUNT_ON_DFL);
>> >> +	return IS_ENABLED(CONFIG_MEMSW_ACCOUNT_ON_DFL)
>> >> +				|| static_branch_unlikely(&memsw_account_on_dfl);
>> >
>> > Why || in above condition? Shouldn't it be && ?
>> >
>> >>  }
>> >>  
>> >>  #define MEM_CGROUP_ID_SHIFT	16
>> >> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
>> >> index 768d6b15dbfa..c1171fb2bfd6 100644
>> >> --- a/mm/memcontrol.c
>> >> +++ b/mm/memcontrol.c
>> >> @@ -5478,3 +5478,14 @@ static int __init mem_cgroup_swap_init(void)
>> >>  subsys_initcall(mem_cgroup_swap_init);
>> >>  
>> >>  #endif /* CONFIG_SWAP */
>> >> +
>> >> +DEFINE_STATIC_KEY_FALSE(memsw_account_on_dfl);
>> >> +static int __init memsw_account_on_dfl_setup(char *s)
>> >> +{
>> >> +	if (!strcmp(s, "1"))
>> >> +		static_branch_enable(&memsw_account_on_dfl);
>> >> +	else if (!strcmp(s, "0"))
>> >> +		static_branch_disable(&memsw_account_on_dfl);
>> >> +	return 1;
>> >> +}
>> >> +__setup("cgroup.memsw_account_on_dfl=", memsw_account_on_dfl_setup);
>> >
>> > Please keep the above in memcontrol-v1.c
>> 
>> Hm, I'm not sure about this. This feature might be actually useful with
>> cgroup v2, as some companies are dependent on the old cgroup v1
>> semantics here but otherwise would prefer to move to v2.
>> In other words, I see it as a cgroup v2 feature, not as a cgroup v1.
>> So there is no reason to move it into the cgroup v1 code.
>
> Agreed. Let's think of this proposal as making memsw tracking and
> control a full-fledged v2 feature.
>
>> I think it deserves a separate config option (if we're really concerned
>> about the memory overhead in struct mem_cgroup) or IMO better a
>> boot/mount time option.
>
> Yeah, a config option forces distros to enable it :/
>
> I'm hesitant to agree with making it optional in any manner. If you
> consider the functionality that is implemented, the overhead should be
> fairly minimal. It isn't right now, because page_counter contains a
> ton of stuff that isn't applicable to this new user. That overhead is
> still paid for unnecessarily by users who _do_ need to enable it.

Agree. Memcg is already huge, so another page_counter won't add a lot
percentage-wise.

>
> It seems like a good opportunity to refactor struct page_counter.

I don't think it's a hard dependency here, but otherwise fully agree.

Thanks!