Re: [PATCH] mm: net: memcg accounting for TCP rx zerocopy

From: Roman Gushchin <guro@fb.com>
To: Arjun Roy <arjunroy@google.com>
Cc: Shakeel Butt <shakeelb@google.com>,
	Yang Shi <shy828301@gmail.com>,
	Johannes Weiner <hannes@cmpxchg.org>,
	Michal Hocko <mhocko@kernel.org>,
	Eric Dumazet <edumazet@google.com>,
	Andrew Morton <akpm@linux-foundation.org>,
	"David S . Miller" <davem@davemloft.net>,
	Jakub Kicinski <kuba@kernel.org>, Linux MM <linux-mm@kvack.org>,
	Cgroups <cgroups@vger.kernel.org>,
	netdev <netdev@vger.kernel.org>,
	Linux Kernel Mailing List <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH] mm: net: memcg accounting for TCP rx zerocopy
Date: Tue, 19 Jan 2021 19:49:49 -0800	[thread overview]
Message-ID: <20210120034949.GA1218369@carbon.dhcp.thefacebook.com> (raw)
In-Reply-To: <CAOFY-A2C4=fWQB69rmP1Ff1Sh=NLCPKT1kD-Lpq29342YJvaWA@mail.gmail.com>

On Tue, Jan 19, 2021 at 07:31:51PM -0800, Arjun Roy wrote:
> On Wed, Jan 13, 2021 at 11:55 AM Shakeel Butt <shakeelb@google.com> wrote:
> >
> > On Wed, Jan 13, 2021 at 11:49 AM Yang Shi <shy828301@gmail.com> wrote:
> > >
> > > On Wed, Jan 13, 2021 at 11:13 AM Shakeel Butt <shakeelb@google.com> wrote:
> > > >
> > > > On Wed, Jan 13, 2021 at 10:43 AM Roman Gushchin <guro@fb.com> wrote:
> > > > >
> > > > > On Tue, Jan 12, 2021 at 04:18:44PM -0800, Shakeel Butt wrote:
> > > > > > On Tue, Jan 12, 2021 at 4:12 PM Arjun Roy <arjunroy@google.com> wrote:
> > > > > > >
> > > > > > > On Tue, Jan 12, 2021 at 3:48 PM Roman Gushchin <guro@fb.com> wrote:
> > > > > > > >
> > > > > > [snip]
> > > > > > > > Historically we have a corresponding vmstat counter to each charged page.
> > > > > > > > It helps with finding accounting/stastistics issues: we can check that
> > > > > > > > memory.current ~= anon + file + sock + slab + percpu + stack.
> > > > > > > > It would be nice to preserve such ability.
> > > > > > > >
> > > > > > >
> > > > > > > Perhaps one option would be to have it count as a file page, or have a
> > > > > > > new category.
> > > > > > >
> > > > > >
> > > > > > Oh these are actually already accounted for in NR_FILE_MAPPED.
> > > > >
> > > > > Well, it's confusing. Can't we fix this by looking at the new page memcg flag?
> > > >
> > > > Yes we can. I am inclined more towards just using NR_FILE_PAGES (as
> > > > Arjun suggested) instead of adding a new metric.
> > >
> > > IMHO I tend to agree with Roman, it sounds confusing. I'm not sure how
> > > people relies on the counter to have ballpark estimation about the
> > > amount of reclaimable memory for specific memcg, but they are
> > > unreclaimable. And, I don't think they are accounted to
> > > NR_ACTIVE_FILE/NR_INACTIVE_FILE, right? So, the disparity between
> > > NR_FILE_PAGES and NR_{IN}ACTIVE_FILE may be confusing either.
> > >
> >
> > Please note that due to shmem/tmpfs there is already disparity between
> > NR_FILE_PAGES and NR_{IN}ACTIVE_FILE.
> >
> > BTW I don't have a strong opinion against adding a new metric. If
> > there is consensus we can add one.
> 
> Just wanted to see if there were any thoughts/consensus on what the
> best way to proceed is - should there be a v2 patch with specific
> changes? Or is NR_FILE_PAGES alright?

I struggle to see why these pages should be considered file pages.
(NR_FILE_MAPPED is a different story).
I'm ok with slab/kmem, sock and a new metric, we can discuss what's
the best option out of three.

> And similar query, for pre-charging vs. post charging.

IMO double accounting is bad. If it means post charging, I vote for
post charging.

Thanks!