Received-SPF: pass (google.com: domain of linux-kernel+bounces-155700-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:45e3:2400::1 as permitted sender) client-ip=2604:1380:45e3:2400::1;
Date: Tue, 23 Apr 2024 10:44:07 -0700
From: Shakeel Butt <shakeel.butt@linux.dev>
To: Johannes Weiner <hannes@cmpxchg.org>
Cc: Andrew Morton <akpm@linux-foundation.org>, 
	Michal Hocko <mhocko@kernel.org>, Roman Gushchin <roman.gushchin@linux.dev>, 
	Muchun Song <muchun.song@linux.dev>, linux-mm@kvack.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH 1/4] mm: rearrange node_stat_item to put memcg stats at
 start
Message-ID: <cupqywok4kl3cxotmpnfrlcsxhkaj7lbo6viehvxlltv3qkt7g@nxxplmgbcfyi>
References: <20240423051826.791934-1-shakeel.butt@linux.dev>
 <20240423051826.791934-2-shakeel.butt@linux.dev>
 <20240423135844.GA21141@cmpxchg.org>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20240423135844.GA21141@cmpxchg.org>

On Tue, Apr 23, 2024 at 09:58:44AM -0400, Johannes Weiner wrote:
> On Mon, Apr 22, 2024 at 10:18:23PM -0700, Shakeel Butt wrote:
> > At the moment the memcg stats are sized based on the size of enum
> > node_stat_item but not all fields in node_stat_item corresponds to memcg
> > stats. So, rearrage the contents of node_stat_item such that all the
> > memcg specific stats are at the top and then the later patches will make
> > sure that the memcg code will not waste space for non-memcg stats.
> > 
> > Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
> 
> This series is a great idea and the savings speak for themselves.
> 
> But rearranging and splitting vmstats along the memcg-nomemcg line
> seems like an undue burden on the non-memcg codebase and interface.
> 
> - It messes with user-visible /proc/vmstat ordering, and sets things
>   up to do so on an ongoing basis as stats are added to memcg.
> 
> - It also separates related stats (like the workingset ones) in
>   /proc/vmstat when memcg only accounts a subset.
> 
> Would it make more sense to have a translation table inside memcg?
> Like we have with memcg1_events.

Thanks for taking a look. I will look into the translation table
approach. The reason I went with this approach was that I am in parallel
looking into rearranging fields of important MM structs and also enums
to improve cache locality. For example, the field NR_SWAPCACHE is always
accessed together with NR_FILE_PAGES, so it makes sense to have them on
same cacheline. So, is the rearrangement of vmstats a big NO or a little
bit here and there is fine unlike what I did with this patch?

thanks,
Shakeel