Re: [PATCH v7 3/4] arm64: Provide an AMU-based version of arch_freq_avg_get_on_cpu

[Sumit Gupta <sumitg@nvidia.com>]

On 11/10/24 21:59, Vanshidhar Konda wrote:
> External email: Use caution opening links or attachments
> 
> 
> On Thu, Oct 10, 2024 at 01:08:23PM GMT, Beata Michalska wrote:
>> On Thu, Oct 03, 2024 at 02:54:22PM -0700, Vanshidhar Konda wrote:
>>> On Thu, Oct 03, 2024 at 11:39:54PM GMT, Beata Michalska wrote:
>>> > On Thu, Sep 26, 2024 at 04:21:14PM -0700, Vanshidhar Konda wrote:
>>> > > On Thu, Sep 26, 2024 at 12:34:01PM GMT, Beata Michalska wrote:
>>> > > > On Tue, Sep 17, 2024 at 05:41:09PM +0530, Sumit Gupta wrote:
>>> > > > > Hi Beata,
>>> > > > Hi Sumit,
>>> > > > >
>>> > > > > Thank you for the patches.
>>> > > > Thank you for having a look at those.
>>> > > > >
>>> > > > > On 13/09/24 18:59, Beata Michalska wrote:
>>> > > > > > External email: Use caution opening links or attachments
>>> > > > > >
>>> > > > > >
>>> > > > > > With the Frequency Invariance Engine (FIE) being already 
>>> wired up with
>>> > > > > > sched tick and making use of relevant (core counter and 
>>> constant
>>> > > > > > counter) AMU counters, getting the average frequency for a 
>>> given CPU,
>>> > > > > > can be achieved by utilizing the frequency scale factor 
>>> which reflects
>>> > > > > > an average CPU frequency for the last tick period length.
>>> > > > > >
>>> > > > > > The solution is partially based on APERF/MPERF 
>>> implementation of
>>> > > > > > arch_freq_get_on_cpu.
>>> > > > > >
>>> > > > > > Suggested-by: Ionela Voinescu <ionela.voinescu@arm.com>
>>> > > > > > Signed-off-by: Beata Michalska <beata.michalska@arm.com>
>>> > > > > > ---
>>> > > > > >   arch/arm64/kernel/topology.c | 109 
>>> +++++++++++++++++++++++++++++++----
>>> > > > > >   1 file changed, 99 insertions(+), 10 deletions(-)
>>> > > > > >
>>>
>>> --- snip ----
>>>
>>> > > > >
>>> > > > >     ..
>>> > > > >   freq_comput:
>>> > > > >     scale = arch_scale_freq_capacity(cpu);
>>> > > > >     freq = scale * arch_scale_freq_ref(cpu);
>>> > > > >   ----
>>> > > > >
>>> > > > This boils down to the question what that function, and the 
>>> information it
>>> > > > provides, represent really. The 'unknown' here simply says the 
>>> CPU has been idle
>>> > > > for a while and as such the frequency data is a bit stale and 
>>> it does not
>>> > > > represent the average freq the CPU is actually running at 
>>> anymore, which is
>>> > > > the intention here really. Or, that the given CPU is a 
>>> non-housekeeping one.
>>> > > > Either way I believe this is a useful information, instead of 
>>> providing
>>> > > > stale data with no indication on whether the frequency is 
>>> really the 'current'
>>> > > > one or not.
>>> > > >
>>> > > > If that is somehow undesirable we can discuss this further, 
>>> though I'd rather
>>> > > > avoid exposing an interface where the feedback provided is open to
>>> > > > interpretation at all times.
>>> > >
>>> > > Would it make sense to identify that the frequency reporting is 
>>> unknown due to
>>> > > cpu being idle vs some other issue like being a non-housekeeping 
>>> CPU? Would
>>> > > returning a value of 0 make it easier for tools to represent that 
>>> the CPU is
>>> > > currently idle?
>>> > That is an option.
>>> > Another one would be to return an error for those cases. This would 
>>> make it
>>> > easier to distinguish between valid frequency &/| idle CPU vs 
>>> tickless CPU
>>> > (EINVAL vs ENOENT) ?
>>> >
>>>
>>> That seems like a good idea but I suspect it would be confusing to 
>>> the end user.
>>>
>>> If a user runs `cat /sys/devices/system/cpu/cpu2/cpuinfo_avg_freq` 
>>> they would
>>> get an error in some cases or get a number in some other iterations.
>>>
>> That is a fair point but I am not entirely convinced using '0' instead 
>> makes
>> things any more clearer as this is in no way a valid CPU frequency.
>> As long as we document the expected behaviour keeping the interface well
>> defined,  both options should be fine I guess.
>>
> 
> Another option could be to list out the reason as 'idle' or 'no entry' 
> instead of
> returning EINVAL or ENOENT. These wouldn't be valid values either but 
> cat on the
> sysfs node wouldn't return an error.
> 
> Thanks,
> Vanshidhar
> 

Ya, listing the clear reason sounds better.

Thank you,
Sumit Gupta

>> @Viresh: what is your opinion on that one ?
>>
>> ---
>> BR
>> Beata
>>> Thanks,

....

>>> > > > > > +               cpu = ref_cpu;
>>> > > > > > +               goto retry;
>>> > > > > > +       }
>>> > > > > > +       /*
>>> > > > > > +        * Reversed computation to the one used to determine
>>> > > > > > +        * the arch_freq_scale value
>>> > > > > > +        * (see amu_scale_freq_tick for details)
>>> > > > > > +        */
>>> > > > > > +       scale = arch_scale_freq_capacity(cpu);
>>> > > > > > +       freq = scale * arch_scale_freq_ref(cpu);
>>> > > > > > +       freq >>= SCHED_CAPACITY_SHIFT;
>>> > > > > > +       return freq;
>>> > > > > > +}
>>> > > > > > +
>>> > > > >
>>> > > > > >   static void amu_fie_setup(const struct cpumask *cpus)
>>> > > > > >   {
>>> > > > > >          int cpu;
>>> > > > > > --
>>> > > > > > 2.25.1
>>> > > > > >