Message-ID: <Zjn8RGSDfn3HrYYd@arm.com>
Date: Tue, 7 May 2024 12:02:44 +0200
From: Beata Michalska <beata.michalska@....com>
To: Viresh Kumar <viresh.kumar@...aro.org>
Cc: Vanshidhar Konda <vanshikonda@...amperecomputing.com>,
	linux-kernel@...r.kernel.org, linux-arm-kernel@...ts.infradead.org,
	ionela.voinescu@....com, sudeep.holla@....com, will@...nel.org,
	catalin.marinas@....com, vincent.guittot@...aro.org,
	sumitg@...dia.com, yang@...amperecomputing.com,
	lihuisong@...wei.com
Subject: Re: [PATCH v4 4/4] cpufreq: Use arch specific feedback for
 cpuinfo_cur_freq

On Tue, May 07, 2024 at 10:31:52AM +0200, Beata Michalska wrote:
> On Mon, Apr 29, 2024 at 02:55:15PM +0530, Viresh Kumar wrote:
> > On 26-04-24, 12:45, Beata Michalska wrote:
> > > It seems that we might need to revisit the discussion we've had around
> > > scaling_cur_freq and cpuinfo_cur_freq and the use of arch_freq_get_on_cpu.
> > > As Vanshi has raised, having both utilizing arch specific feedback for
> > > getting current frequency is bit problematic and might be confusing at best.
> > > As arch_freq_get_on_cpu is already used by show_scaling_cur_freq there are not
> > > many options we are left with, if we want to kee all archs aligned:
> > > we can either try to rework show_scaling_cur_freq and it's use of
> > > arch_freq_get_on_cpu, and move it to cpuinfo_cur_freq, which would align with
> > > relevant docs, though that will not work for x86, or we keep it only there and
> > > skip updating cpuinfo_cur_freq, going against the guidelines. Other options,
> > > purely theoretical, would involve making arch_freq_get_on_cpu aware of type of
> > > the info requested (hw vs sw) or adding yet another arch-specific implementation,
> > > and those are not really appealing alternatives to say at least.
> > > What's your opinion on this one ?
> > 
> > Hi Beata / Vanshidhar,
> > 
> > Lets forget for once what X86 and ARM may have done and think about it
> > once again. I also had a chat with Vincent today about this.
> > 
> > The documentation says it clearly, cpuinfo_cur_freq is the one
> > received from hardware and scaling_cur_freq is the one requested from
> > software.
> > 
> > Now, I know that X86 has made both of them quite similar and I
> > suggested to make them all aligned (and never received a reply on my
> > previous message).
> > 
> > There are few reasons why it may be worth keeping the definition (and
> > behavior) of the sysfs files as is, at least for ARM:
> > - First is that the documentation says so.
> > - There is no point providing the same information via both the
> >   interfaces, there are two interfaces here for a reason.
> > - There maybe tools around which depend on the documented behavior.
> > - From userspace, currently there is only one way to know the exact
> >   frequency that the cpufreq governors have requested from a platform,
> >   i.e. the value from scaling_cur_freq. If we make it similar to
> >   cpuinfo_cur_freq, then userspace will never know about the requested
> >   frequency and the eventual one and if they are same or different.
> > 
> > Lets keep the behavior as is and update only cpuinfo_cur_freq with
> > arch_freq_get_on_cpu().
> > 
> > Makes sense ?
> >
> First of all - apologies for late reply.
> 
> It all makes sense, though to clarify things up, for my own benefit, and to
> avoid any potential confusion ....
> 
> Adding arch_freq_get_on_cpu to cpuinfo_cur_freq does seem to be the right
> approach - no argue on this one. Doing that though means we need a way to
> skip calling arch_freq_get_on_cpu() from show_scaling_cur_freq(), to avoid
> having both providing the same information when that should not be the case.
> In the initial approach [1], that was handled by checking whether the cpufreq
> driver supports 'get' callback (and thus cpuinfo_cur_freq). In case it didn't,
> things remained unchanged for scaling_cur_freq. That does not seem to be a viable
> option though, as there are at least few drivers, that will support both:
> cpuinfo_cur_freq alongside scaling_cur_freq (+ APERF/MPERF) and for those,
> we would hit the initial problem of both relying on arch_freq_get_on_cpu.
> So I guess we need another way of avoiding calling arch_freq_get_on_cpu
> for show_scaling_cur_freq (and most probably avoid calling that one for
> cpuinfo_cur_freq). Quick idea on how to not bring arch specificity into
> cpufreq generic code would be to introduce a new flag for cpufreq drivers though
> that seems a bit stretched. Will ponder a bit about that but in the meantime
> suggestions are more than welcomed.
Alternatively we could add a parameter to arch_freq_get_on_cpu specifying type
of feedback required: hw vs sw. Then the arch specific implementation could
decide which to provide when. It will get slightly counter-intuitive, especially
for cases when sw feedback provides hw one, as it is the case for current
arch_freq_get_on_cpu() and scaling_cur_freq but at least the changes would be
minimal and it will contain handling the tricky bits inside arch specific
implementation - hiding those messy bits.

---
BR
Beata
> 
> Aside: I will most probably send the changes separately from this series to not
> mix things any further.
> 
> ---
> [1] https://lore.kernel.org/all/20230606155754.245998-1-beata.michalska@arm.com/
> ---
> BR
> Beata
> 
> 
> > -- 
> > viresh

Powered by blists - more mailing lists