Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Thu, 20 Feb 2020 10:56:51 -0500
From:   Tejun Heo <tj@kernel.org>
To:     Daniel Jordan <daniel.m.jordan@oracle.com>
Cc:     Johannes Weiner <hannes@cmpxchg.org>,
        Michal Hocko <mhocko@kernel.org>,
        Andrew Morton <akpm@linux-foundation.org>,
        Roman Gushchin <guro@fb.com>, linux-mm@kvack.org,
        cgroups@vger.kernel.org, linux-kernel@vger.kernel.org,
        kernel-team@fb.com, Peter Zijlstra <peterz@infradead.org>
Subject: Re: [PATCH] mm: memcontrol: asynchronous reclaim for memory.high
Message-ID: <20200220155651.GG698990@mtj.thefacebook.com>
References: <20200219181219.54356-1-hannes@cmpxchg.org>
 <20200219183731.GC11847@dhcp22.suse.cz>
 <20200219191618.GB54486@cmpxchg.org>
 <20200219195332.GE11847@dhcp22.suse.cz>
 <20200219214112.4kt573kyzbvmbvn3@ca-dmjordan1.us.oracle.com>
 <20200219220859.GF54486@cmpxchg.org>
 <20200220154524.dql3i5brnjjwecft@ca-dmjordan1.us.oracle.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20200220154524.dql3i5brnjjwecft@ca-dmjordan1.us.oracle.com>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Thu, Feb 20, 2020 at 10:45:24AM -0500, Daniel Jordan wrote:
> Ok, consistency with io and memory is one advantage to doing it that way.
> Creating kthreads in cgroups also seems viable so far, and it's unclear whether
> either approach is significantly simpler or more maintainable than the other,
> at least to me.

The problem with separate kthread approach is that many of these work
units are tiny, and cgroup membership might not be known or doesn't
agree with the processing context from the beginning

For example, the ownership of network packets can't be determined till
processing has progressed quite a bit in shared contexts and each item
too small to bounce around. The only viable way I can think of
splitting aggregate overhead according to the number of packets (or
some other trivially measureable quntity) processed.

Anything sitting in reclaim layer is the same. Reclaim should be
charged to the cgroup whose memory is reclaimed *but* shouldn't block
other cgroups which are waiting for that memory. It has to happen in
the context of the highest priority entity waiting for memory but the
costs incurred must be charged to the memory owners.

So, one way or the other, I think we'll need back charging and once
back charging is needed for big ticket items like network and reclaim,
it's kinda silly to use separate mechanisms for other stuff.

> Is someone on your side working on remote charging right now?  I was planning
> to post an RFD comparing these soon and it would make sense to include them.

It's been on the to do list but nobody is working on it yet.

Thanks.

-- 
tejun