Date: Mon, 21 Nov 2016 12:48:02 +0100
From: Peter Zijlstra <peterz@infradead.org>
To: Viresh Kumar <viresh.kumar@linaro.org>
Cc: Rafael Wysocki <rjw@rjwysocki.net>, Ingo Molnar <mingo@redhat.com>,
        linaro-kernel@lists.linaro.org, linux-pm@vger.kernel.org,
        linux-kernel@vger.kernel.org,
        Vincent Guittot <vincent.guittot@linaro.org>,
        Juri Lelli <Juri.Lelli@arm.com>,
        Robin Randhawa <robin.randhawa@arm.com>,
        Steve Muckle <smuckle.linux@gmail.com>, tkjos@google.com,
        Morten Rasmussen <morten.rasmussen@arm.com>
Subject: Re: [PATCH] cpufreq: schedutil: add up/down frequency transition
 rate limits
Message-ID: <20161121114802.GB3092@twins.programming.kicks-ass.net>
References: <c6248ec9475117a1d6c9ff9aafa8894f6574a82f.1479359903.git.viresh.kumar@linaro.org>
 <20161121100805.GB10014@vireshk-i7>
 <20161121101946.GI3102@twins.programming.kicks-ass.net>
 <20161121104800.GC10014@vireshk-i7>
 <20161121111243.GK3102@twins.programming.kicks-ass.net>
 <20161121113016.GD10014@vireshk-i7>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20161121113016.GD10014@vireshk-i7>
User-Agent: Mutt/1.5.23.1 (2014-03-12)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1237
Lines: 24

On Mon, Nov 21, 2016 at 05:00:16PM +0530, Viresh Kumar wrote:
> On 21-11-16, 12:12, Peter Zijlstra wrote:
> > I think it should be replaced by a value provided by the driver. It
> > makes sense to have a rate-limit in so far as that it doesn't make sense
> > to try and program the hardware faster than it can actually change
> > frequencies and/or have a programming cost amortization. And this very
> > clearly is a driver specific thing.
> 
> We already have something called as transition_latency for that (though it isn't
> used much currently).
> 
> > It however doesn't make sense to me to fudge with this in order to
> > achieve ramp up/down differences.
> 
> So if a platform, for example, can do DVFS in say 100-500 us, then the scheduler
> should try to re-evaluate frequency (and update it) after that short of a
> period? Wouldn't that scheme waste lots of time doing just freq updates? And
> that's the primary reason why cpufreq governors have some sort of sampling-rate
> or rate-limit until now.

Dunno.. there's of course the cost amortization, but by the time we've
reached sugov_should_update_freq() most of the 'expensive' parts have
already been done from the scheduler's POV and its once again down to
the driver.