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;
Sender: Tejun Heo <htejun@gmail.com>
Date:   Fri, 21 Oct 2022 08:04:13 -1000
From:   Tejun Heo <tj@kernel.org>
To:     Peter Zijlstra <peterz@infradead.org>
Cc:     Chuyi Zhou <zhouchuyi@bytedance.com>, mingo@redhat.com,
        juri.lelli@redhat.com, vincent.guittot@linaro.org,
        linux-kernel@vger.kernel.org, htejun@gmail.com,
        lizefan.x@bytedance.com, vschneid@redhat.com, bsegall@google.com,
        Abel Wu <wuyun.abel@bytedance.com>
Subject: Re: [RESEND] sched/fair: Add min_ratio for cfs bandwidth_control
Message-ID: <Y1LfHVENWB8ZDFyr@slm.duckdns.org>
References: <20221019031551.24312-1-zhouchuyi@bytedance.com>
 <Y1BqT59ovFitahJa@slm.duckdns.org>
 <Y1GAffm4aHCpvoBB@hirez.programming.kicks-ass.net>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <Y1GAffm4aHCpvoBB@hirez.programming.kicks-ass.net>
Precedence: bulk

Hello,

On Thu, Oct 20, 2022 at 07:08:13PM +0200, Peter Zijlstra wrote:
> > This is a bit of a bandaid. I think what we really need to do is only
> > throttling when running in userspace. In kernel space, it should just keep
> > accumulating used cycles as debt which should be paid back before userspace
> > code can run again so that we don't throttle at random places in the kernel.
> 
> That's just moving the problem. But yeah; perhaps. Starving random
> userspace is less of a problem I suppose.

Given that our primary mean of guaranteeing forward progress is the fact
that the system runs out of other things to do when there are severe
priority inversions, I don't think we can safely give control of throttling
something running in the kernel to userspace.

IO control takes a similar approach with shared IOs which can have
system-wide impacts and it's been working out pretty well. While some may go
over the limit briefly, it's not that difficult to remain true to the
intended configuration over time.

The only problem is the cases where userspace can cause a large amount of
forced consumptions (e.g. for IOs, creating a lot of metadata updates
without doing anything else), but even in the unlikely case similar problem
exists for CPU, it's pretty easy to add specific control mechanisms around
those (e.g. sth along the style of might_resched()).

So, yeah, I think this is the actual solution.

Thanks.

-- 
tejun