Re: [PATCH] cpufreq: CPPC: Increase delay between perf counter reads

[Prashant Malani]

On Tue, 19 Aug 2025 at 02:29, Beata Michalska <beata.michalska@xxxxxxx> wrote:
>
> On Tue, Aug 05, 2025 at 06:00:09PM -0700, Prashant Malani wrote:
> > Thanks Yang,
> >
> > On Tue, 5 Aug 2025 at 17:26, Yang Shi <yang@xxxxxxxxxxxxxxxxxxxxxx> wrote:
> > > Thank you for cc'ing me the patch. I posted the similar patch ago and
> > > had some discussion on the mailing list. Then someone else from ARM
> > > pursued a different way to solve it. But I didn't follow very closely.
> > > If I remember correctly, a new sysfs interface, called cpuinfo_avg_freq
> > > was added. It should be the preferred way to get cpu frequency. Please
> > > see
> > > https://github.com/torvalds/linux/commit/fbb4a4759b541d09ebb8e391d5fa7f9a5a0cad61.
> > >
> > > Added Beata Michalska in the loop too, who is the author of the patch.
> > > Please feel free to correct me, if I'm wrong.
> >
> > Thanks for the additional context. Yeah, the issue is that :
> > - The new sysfs node is sampling period is too long (20ms) [1]
> > That makes it problematic for userspace use cases, so we need something
> > which takes less time.
> That actually specifies the duration, for which the most recently acquired
> sample is considered valid. Sampling is tick-based.

Thanks for the correction. I made an error in understanding the code.

>
> > - The central accuracy issue behind cpuinfo_cur_freq still needs to be handled.
> I'd really discourage increasing the sampling period for cppc.

The only true solution is to make the register reads (ref + delivered
combined) atomic.
We see that this solves the issue when conducting the same
measurements on firmware
(which is an RTOS, so no scheduler randomness).

Outside of that, I can't think of another mitigation beyond adding delay to make
the time deltas not matter so much.

Perhaps ARM should introduce a "snapshot" feature which takes the snapshot
of the AMU counters on a register write; IAC that's outside the scope
of this discussion.

Could you kindly explain why adding the udelay here is discouraged?

Best regards,

-- 
-Prashant