RE: [PATCH] Remove some PMU events for FUJITSU-MONAKA
From: Koichi Okuno (Fujitsu)
Date: Thu Mar 27 2025 - 00:25:07 EST
Hi, James
Sorry for the late reply.
Also, the person in charge here has changed from Furudera to Okuno.
> Hi, James
> Thank you for your comment.
>
> > On 05/03/2025 6:40 am, Yoshihiro Furudera (Fujitsu) wrote:
> > > Hi, James
> > >
> > >>
> > >> Hi, James
> > >> Thank you for your comment.
> > >>
> > >>> On 27/02/2025 5:40 am, Yoshihiro Furudera wrote:
> > >>>> The following events are not counted properly:
> > >>>>
> > >>>> 0x0037 LL_CACHE_MISS_RD
> > >>>> 0x400B L3D_CACHE_LMISS_RD
> > >>>
> > >>> These two are discoverable so will still appear in
> > >>> /sys/bus/event_source/devices/armv8_pmuv3_0/events/ if the
> > >>> hardware says they exist. It might be better to change the json
> > >>> strings of these two to warn that they don't work if that's the
> > >>> case, otherwise Perf will still list them and you'll be worse off.
> > >>
> > >> I will leave these 2 events and
> > >> add a warning message to the description in the JSON file.
> > >> I will handle other events in the same way as these 2 events.
> > >
> > > I'm thinking of adding one of the following warnings to the
> > > description of events where inaccurate counts occur on
> FUJITSU-MONAKA.
> > > Is this okay?
> > >
> > > 1.Simple version
> > > "Note: This event does not count accurately."
> > >
> > > 2.Detailed version
> > > "Note: This event does not count accurately because it counts as a
> > > miss
> > regardless of whether the L3 prefetch is a hit or miss."
> > >
> > > I think "2.Detailed version" is better.
> > >
> > > example:
> > > {
> > > "EventCode": "0x0396",
> > > "EventName": "L2D_CACHE_REFILL_L3D_MISS",
> > > "BriefDescription": "This event counts operations that cause a
> > > miss of the
> > L3 cache. Note: This event does not count accurately because it counts
> > as a miss regardless of whether the L3 prefetch is a hit or miss."
> > > }
> > >
> > > Best Regards,
> > > Yoshihiro Furudera
> >
> > You could have both by using BriefDescription and PublicDescription.
> > The longer one being available with 'perf list -v'. I think that's what that
> feature is for.
>
> Thank you for your suggestion.
> After discussing this with the OS team,
> we would like to move forward with the policy of defining "2.Detailed version" in
> the Brief Description.
> Please let me make a final decision after discussing it with the hardware team.
>
> >
> > Either way you should probably also update the description before the
> > note, so add 'hit or miss' to the first sentence:
> >
> > "EventCode": "0x0396",
> > "EventName": "L2D_CACHE_REFILL_L3D_MISS",
> > "BriefDescription": "This event counts operations that cause a hit or
> > miss of the L3 cache. Note that this incorrectly counts both hits _and_
> misses."
>
> We will also discuss with the hardware team on this matter.
After discussing with the hardware team,
for all events where the count is incorrect,
we have decided to add the following text to
the end of the "BriefDescription":
"Note: This event does not count accurately."
I will repost the patch later.
>
> Best Regards,
> Yoshihiro Furudera
Best Regards,
Koichi Okuno