Received-SPF: pass (google.com: domain of linux-kernel+bounces-149432-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.80.249 as permitted sender) client-ip=147.75.80.249;
Precedence: bulk
MIME-Version: 1.0
References: <171328983017.3930751.9484082608778623495.stgit@firesoul> <Zh7vuBRbA9rT5OCO@slm.duckdns.org>
In-Reply-To: <Zh7vuBRbA9rT5OCO@slm.duckdns.org>
From: Yosry Ahmed <yosryahmed@google.com>
Date: Wed, 17 Apr 2024 19:13:40 -0700
Message-ID: <CAJD7tkZhjYZQqsnTvUnv9EB1KUNyKijxLbYLOMsEcsRcZw=j3Q@mail.gmail.com>
Subject: Re: [PATCH v1 0/3] cgroup/rstat: global cgroup_rstat_lock changes
To: Tejun Heo <tj@kernel.org>
Cc: Jesper Dangaard Brouer <hawk@kernel.org>, hannes@cmpxchg.org, lizefan.x@bytedance.com, 
	cgroups@vger.kernel.org, longman@redhat.com, netdev@vger.kernel.org, 
	linux-mm@kvack.org, linux-kernel@vger.kernel.org, shakeel.butt@linux.dev, 
	kernel-team@cloudflare.com, Arnaldo Carvalho de Melo <acme@kernel.org>, 
	Sebastian Andrzej Siewior <bigeasy@linutronix.de>, mhocko@kernel.org, Wei Xu <weixugc@google.com>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Tue, Apr 16, 2024 at 2:38=E2=80=AFPM Tejun Heo <tj@kernel.org> wrote:
>
> On Tue, Apr 16, 2024 at 07:51:19PM +0200, Jesper Dangaard Brouer wrote:
> > This patchset is focused on the global cgroup_rstat_lock.
> >
> >  Patch-1: Adds tracepoints to improve measuring lock behavior.
> >  Patch-2: Converts the global lock into a mutex.
> >  Patch-3: Limits userspace triggered pressure on the lock.
>
> Imma wait for people's inputs on patch 2 and 3. ISTR switching the lock t=
o
> mutex made some tail latencies really bad for some workloads at google?
> Yosry, was that you?

I spent some time going through the history of my previous patchsets
to find context.

There were two separate instances where concerns were raised about
using a mutex.

(a) Converting the global rstat spinlock to a mutex:

Shakeel had concerns about priority inversion with a global sleepable
lock. So I never actually tested replacing the spinlock with a mutex
based on Shakeel's concerns as priority inversions would be difficult
to reproduce with synthetic tests.

Generally speaking, other than priority inversions, I was depending on
Wei's synthetic test to measure performance for userspace reads, and a
script I wrote with parallel reclaimers to measure performance for
in-kernel flushers.

(b) Adding a mutex on top of the global rstat spinlock for userspace
reads (to limit contention from userspace on the in-kernel lock):

Wei reported that this significantly affects userspace read latency
[2]. I then proceeded to add per-memcg thresholds for flushing, which
resulted in the regressions from that mutex going away. However, at
that point the mutex didn't really provide much value, so I removed it
[3].

[1]https://lore.kernel.org/lkml/CALvZod441xBoXzhqLWTZ+xnqDOFkHmvrzspr9NAr+n=
ybqXgS-A@mail.gmail.com/
[2]https://lore.kernel.org/lkml/CAAPL-u9D2b=3DiF5Lf_cRnKxUfkiEe0AMDTu6yhrUA=
zX0b6a6rDg@mail.gmail.com/
[3]https://lore.kernel.org/lkml/CAJD7tkZgP3m-VVPn+fF_YuvXeQYK=3DtZZjJHj=3Dd=
zD=3DCcSSpp2qg@mail.gmail.com/