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, 25 Jul 2019 10:47:42 +0530
From:   Viresh Kumar <viresh.kumar@linaro.org>
To:     Saravana Kannan <saravanak@google.com>
Cc:     MyungJoo Ham <myungjoo.ham@samsung.com>,
        Kyungmin Park <kyungmin.park@samsung.com>,
        Chanwoo Choi <cw00.choi@samsung.com>,
        Viresh Kumar <vireshk@kernel.org>, Nishanth Menon <nm@ti.com>,
        Stephen Boyd <sboyd@kernel.org>,
        "Rafael J. Wysocki" <rjw@rjwysocki.net>,
        Sibi Sankar <sibis@codeaurora.org>,
        Android Kernel Team <kernel-team@android.com>,
        Linux PM <linux-pm@vger.kernel.org>,
        LKML <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH v3 3/5] OPP: Improve require-opps linking
Message-ID: <20190725051742.mn54pi722txkpddg@vireshk-i7>
References: <20190717222340.137578-1-saravanak@google.com>
 <20190717222340.137578-4-saravanak@google.com>
 <20190723102842.t2s45zzylsjuccm4@vireshk-i7>
 <CAGETcx-6M9Ts8tfMf6aA8GjMyzK5sOLr069ZCxTG7RHMFPLzHw@mail.gmail.com>
 <20190725030712.lx3cjogo5r7kc262@vireshk-i7>
 <CAGETcx8QTs2Dqqppb_gwiUa2fte92K_q+B+j_CreRgqU52L7EA@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <CAGETcx8QTs2Dqqppb_gwiUa2fte92K_q+B+j_CreRgqU52L7EA@mail.gmail.com>
User-Agent: NeoMutt/20180716-391-311a52
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On 24-07-19, 21:09, Saravana Kannan wrote:
> On Wed, Jul 24, 2019 at 8:07 PM Viresh Kumar <viresh.kumar@linaro.org> wrote:
> > We should be doing this whenever a new OPP table is created, and see
> > if someone else was waiting for this OPP table to come alive.
> 
> Searching the global OPP table list seems a ton more wasteful than
> doing the lazy linking. I'd rather not do this.

We can see how best to optimize that, but it will be done only once
while a new OPP table is created and putting stress there is the right
thing to do IMO. And doing anything like that in a place like
opp-set-rate is the worst one. It will be a bad choice by design if
you ask me and so I am very much against that.

> > Also we
> > must make sure that we do this linking only if the new OPP table has
> > its own required-opps links fixed, otherwise delay further.
> 
> This can be done. Although even without doing that, this patch is
> making things better by not failing silently like it does today? Can I
> do this later as a separate patch set series?

I would like this to get fixed now in a proper way, there is no hurry
for a quick fix currently. No band-aids please.

> > Even then I don't want to add these checks to those places. For the
> > opp-set-rate routine, add another flag to the OPP table which
> > indicates if we are ready to do dvfs or not and mark it true only
> > after the required-opps are all set.
> 
> Honestly, this seems like extra memory and micro optimization without
> any data to back it.

Again, opp-set-rate isn't supposed to do something like this. It
shouldn't handle initializations of things, that is broken design.

> Show me data that checking all these table
> pointers is noticeably slower than what I'm doing. What's the max
> "required tables count" you've seen in upstream so far?

Running anything extra (specially some initialization stuff) in
opp-set-rate is wrong as per me and as a Maintainer of the OPP core it
is my responsibility to not allow such things to happen.

> I'd even argue that doing it the way I do might actually reduce the
> cache misses/warm the cache because those pointers are going to be
> searched/used right after anyway.

So you want to make the cache hot with data, by running some code at a
place where it is not required to be run really, and the fact that
most of the data cached may not get used anyway ? And that is an
improvement somehow ?

-- 
viresh