Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <d323bd95-476b-0901-855e-14c8796d1b23@bytedance.com>
Date:   Wed, 7 Sep 2022 20:06:30 +0800
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:91.0) Gecko/20100101
 Thunderbird/91.13.0
Subject: Re: [Phishing Risk] Re: [Phishing Risk] [External] Re: [PATCH]
 cgroup/cpuset: Add a new isolated mems.policy type.
To:     Tejun Heo <tj@kernel.org>
Cc:     lizefan.x@bytedance.com, hannes@cmpxchg.org,
        cgroups@vger.kernel.org, linux-kernel@vger.kernel.org
References: <20220902063303.1057-1-hezhongkun.hzk@bytedance.com>
 <YxT/liaotbiOod51@slm.duckdns.org>
 <c05bdeac-b354-0ac7-3233-27f8e5cbb38a@bytedance.com>
 <YxeBGeOaQxvlPLzo@slm.duckdns.org>
From:   Zhongkun He <hezhongkun.hzk@bytedance.com>
In-Reply-To: <YxeBGeOaQxvlPLzo@slm.duckdns.org>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk

> Hello,
> 
> On Mon, Sep 05, 2022 at 06:30:38PM +0800, Zhongkun He wrote:
>> We usually use numactl to set the memory policy, but it cannot be changed
>> dynamically. In addition, the mempolicy of cpuset can provide a more
>> convenient interface for management and control panel.
> 
> But you can write a better tool easily in userspace to do whatever you wanna
> do, right? If you're worried about racing against forks, you can freeze the
> cgroup, iterate all pids applying whatever new policy and then unfreeze. We
> can probably improve the freezer interface so that multiple users don't
> conflict with each other but that shouldn't be too difficult to do and is
> gonna be useful generically.
> 
> I don't see much point in adding something which can be almost trivially
> implemented in userspace as a built-in kernel feature.
> 
>> Sorry,I don't quite understand the meaning of "don't enforce anything
>> resource related". Does it mean mempolicy, such as "prefer:2" must specify
>> node? Or "cpuset.mems.policy" need to specify a default value?
>> (cpuset.mems.policy does not require a default value.)
> 
> In that there's no real resource being distributed hierarchically like cpu
> cycles or memory capacities. All it's doing is changing attributes for a
> group of processes, which can be done from userspace all the same.
> 
> Thanks.
> 
Hi Tejun, thanks for your reply.

It would be better if one process had a way to dynamically modify the
mempolicy of another process. But unfortunately there is no interface or
system call to do that in userspace.

In our use case, we hope to combine memory policy with cgroup for
better use of resources. The current implementation may not be suitable, 
I'll keep trying other approaches.

Thanks again.