linux-kernel - Re: [PATCH] PM / EM: Inefficient OPPs detection

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Thu, 15 Apr 2021 14:59:54 +0000
From:   Quentin Perret <qperret@...gle.com>
To:     Vincent Donnefort <vincent.donnefort@....com>
Cc:     peterz@...radead.org, rjw@...ysocki.net, viresh.kumar@...aro.org,
        vincent.guittot@...aro.org, linux-kernel@...r.kernel.org,
        ionela.voinescu@....com, lukasz.luba@....com,
        dietmar.eggemann@....com
Subject: Re: [PATCH] PM / EM: Inefficient OPPs detection

On Thursday 15 Apr 2021 at 15:34:53 (+0100), Vincent Donnefort wrote:
> On Thu, Apr 15, 2021 at 01:16:35PM +0000, Quentin Perret wrote:
> > On Thursday 08 Apr 2021 at 18:10:29 (+0100), Vincent Donnefort wrote:
> > > --- a/kernel/sched/cpufreq_schedutil.c
> > > +++ b/kernel/sched/cpufreq_schedutil.c
> > > @@ -10,6 +10,7 @@
> > >  
> > >  #include "sched.h"
> > >  
> > > +#include <linux/energy_model.h>
> > >  #include <linux/sched/cpufreq.h>
> > >  #include <trace/events/power.h>
> > >  
> > > @@ -164,6 +165,9 @@ static unsigned int get_next_freq(struct sugov_policy *sg_policy,
> > >  
> > >  	freq = map_util_freq(util, freq, max);
> > >  
> > > +	/* Avoid inefficient performance states */
> > > +	freq = em_pd_get_efficient_freq(em_cpu_get(policy->cpu), freq);
> > 
> > I remember this was discussed when Douglas sent his patches some time
> > ago, but I still find it sad we index the EM table here but still
> > re-index the cpufreq frequency table later :/
> > 
> > Yes in your case this lookup is very inexpensive, but still. EAS relies
> > on the EM's table matching cpufreq's accurately, so this second lookup
> > still feels rather unnecessary ...
> 
> To get only a single lookup, we would need to bring the inefficiency knowledge
> directly to the cpufreq framework. But it has its own limitations: 
> 
>   The cpufreq driver can have its own resolve_freq() callback, which means that
>   not all the drivers would benefit from that feature.
> 
>   The cpufreq_table can be ordered and accessed in several ways which brings
>   many combinations that would need to be supported, ending-up with something
>   much more intrusive. (We can though decide to limit the feature to the low to
>   high access that schedutil needs).
> 
> As the EM needs schedutil to exist anyway, it seemed to be the right place for
> this code. It allows any cpufreq driver to benefit from the feature, simplify a
> potential extension for a usage by devfreq devices and as a bonus it speeds-up
> energy computing, allowing a more complex Energy Model.

I was thinking of something a bit simpler. cpufreq_driver_resolve_freq
appears to be used only from schedutil (why is it even then?), so we
could just pull it into cpufreq_schedutil.c and just plain skip the call
to cpufreq_frequency_table_target if the target freq has been indexed in
the EM table -- it should already be matching a real OPP.

Thoughts?
Quentin