Re: query: [PATCH 2/2] cgroup: Remove call to synchronize_rcu in cgroup_attach_task

From: Paul Menage <menage@google.com>
To: Mike Galbraith <efault@gmx.de>
Cc: LKML <linux-kernel@vger.kernel.org>,
	Colin Cross <ccross@android.com>,
	Peter Zijlstra <a.p.zijlstra@chello.nl>,
	Ingo Molnar <mingo@elte.hu>, Li Zefan <lizf@cn.fujitsu.com>
Subject: Re: query: [PATCH 2/2] cgroup: Remove call to synchronize_rcu in cgroup_attach_task
Date: Wed, 13 Apr 2011 15:10:59 +0200	[thread overview]
Message-ID: <BANLkTikvHtfeC1MqqXL7yTaXhZzZ3JRi8g@mail.gmail.com> (raw)
In-Reply-To: <1302170153.12304.31.camel@marge.simson.net>

On Thu, Apr 7, 2011 at 11:55 AM, Mike Galbraith <efault@gmx.de> wrote:
> Greetings,
>
> Wrt these patches:
>
> https://lkml.org/lkml/2010/11/24/14 [PATCH 1/2] cgroup: Set CGRP_RELEASABLE when adding to a cgroup
> https://lkml.org/lkml/2010/11/24/15 [PATCH 2/2] cgroup: Remove call to synchronize_rcu in cgroup_attach_task
>
> I received a query regarding 2/2 because a large database company is
> apparently moving tasks between cgroups frequently enough that their
> database initialization time dropped from ~11 hours to ~4 hours when
> they applied this patch.

That sounds like a problem in their user-space code too, although I
agree that making cgroup moves faster is a good thing.

>
> Curious why these got no traction.
>

Apart from just my chronic lack of time to work on cgroups, there were
a couple of issues:

1) we had trouble getting the semantics right for the release_agent
notifications. Not that this is something that I suspect many people
care about, but it has been part of the API since the cpuset days. I
spent a while trying to juggle the way that release notifications were
done (via an event counter rather than a simple flag) but never got
them finished.

2) I have this nagging feeling that the synchronize_rcu() call in
cgroup_attach_task() was protecting more than is obvious. Certainly
when cgroups first went in, that synchronize_rcu() call meant that
cgroup_rmdir() could guarantee that if the cgroup was empty, there
were no threads in RCU-read sections accessing their old cgroup via
their RCU-proected current->cgroups pointers, so objects could just be
deleted at that point. A year or two ago we RCU-ified most/all of the
cgroup deletion path, so this shouldn't be an issue now, but I'm still
a bit worried that we missed something. I'm probably being
over-paranoid though.

We're looking at testing these patches at Google, which will give a
little more confidence.

There's a conflicting patchset (allowing moving entire processes by
writing to cgroup.procs) that Ben Blum has been trying to get in for
ages, and which has just gone in to -mm - the RCU change patches will
likely need a bit of merge love.

Paul