Re: [patch 1/3] mm: embed the memcg pointer directly into struct page

From: Michal Hocko <mhocko@suse.cz>
To: Johannes Weiner <hannes@cmpxchg.org>
Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com>,
	Andrew Morton <akpm@linux-foundation.org>,
	Vladimir Davydov <vdavydov@parallels.com>,
	Tejun Heo <tj@kernel.org>, David Miller <davem@davemloft.net>,
	linux-mm@kvack.org, cgroups@vger.kernel.org,
	linux-kernel@vger.kernel.org
Subject: Re: [patch 1/3] mm: embed the memcg pointer directly into struct page
Date: Mon, 3 Nov 2014 18:02:28 +0100	[thread overview]
Message-ID: <20141103170228.GH10156@dhcp22.suse.cz> (raw)
In-Reply-To: <20141103150942.GA32052@phnom.home.cmpxchg.org>

On Mon 03-11-14 10:09:42, Johannes Weiner wrote:
> Hi Joonsoo,
> 
> On Mon, Nov 03, 2014 at 05:02:08PM +0900, Joonsoo Kim wrote:
> > On Sat, Nov 01, 2014 at 11:15:54PM -0400, Johannes Weiner wrote:
> > > Memory cgroups used to have 5 per-page pointers.  To allow users to
> > > disable that amount of overhead during runtime, those pointers were
> > > allocated in a separate array, with a translation layer between them
> > > and struct page.
> > 
> > Hello, Johannes.
> > 
> > I'd like to leave this translation layer.
> > Could you just disable that code with #ifdef until next user comes?
> > 
> > In our company, we uses PAGE_OWNER on mm tree which is the feature
> > saying who allocates the page. To use PAGE_OWNER needs modifying
> > struct page and then needs re-compile. This re-compile makes us difficult
> > to use this feature. So, we decide to implement run-time configurable
> > PAGE_OWNER through page_cgroup's translation layer code. Moreover, with
> > this infrastructure, I plan to implement some other debugging feature.
> > 
> > Because of my laziness, it didn't submitted to LKML. But, I will
> > submit it as soon as possible. If the code is removed, I would
> > copy-and-paste the code, but, it would cause lose of the history on
> > that code. So if possible, I'd like to leave that code now.
> 
> Please re-introduce this code when your new usecase is ready to be
> upstreamed.  There is little reason to burden an unrelated feature
> with a sizable chunk of dead code for a vague future user.

Completely agreed! I would hate to have some random feature piggy back
on memcg internal structures. For that to be acceptable the page_cgroup
would have to be abstracted into a more generic structure which wouldn't
be under CONFIG_MEMCG anyway.
-- 
Michal Hocko
SUSE Labs