Re: [PATCH 2/3] perf: arm_cspmu: nvidia: update CNVLINK PMU events

[Will Deacon]

On Thu, Oct 24, 2024 at 02:10:55PM +0000, Besar Wicaksono wrote:
> > > > > diff --git a/drivers/perf/arm_cspmu/nvidia_cspmu.c
> > > > b/drivers/perf/arm_cspmu/nvidia_cspmu.c
> > > > > index ea2d44adfa7c..d1cd9975e71a 100644
> > > > > --- a/drivers/perf/arm_cspmu/nvidia_cspmu.c
> > > > > +++ b/drivers/perf/arm_cspmu/nvidia_cspmu.c
> > > > > @@ -112,6 +112,25 @@ static struct attribute *mcf_pmu_event_attrs[]
> > = {
> > > > >       NULL,
> > > > >  };
> > > > >
> > > > > +static struct attribute *mcf_cnvlink_pmu_event_attrs[] = {
> > > > > +     ARM_CSPMU_EVENT_ATTR(rd_bytes_cmem,                     0x0),
> > > > > +     ARM_CSPMU_EVENT_ATTR(rd_bytes_gmem,                     0x1),
> > > > > +     ARM_CSPMU_EVENT_ATTR(wr_bytes_cmem,                     0x2),
> > > > > +     ARM_CSPMU_EVENT_ATTR(wr_bytes_gmem,                     0x3),
> > > > > +     ARM_CSPMU_EVENT_ATTR(total_bytes_cmem,                  0x4),
> > > > > +     ARM_CSPMU_EVENT_ATTR(total_bytes_gmem,                  0x5),
> > > > > +     ARM_CSPMU_EVENT_ATTR(rd_req_cmem,                       0x6),
> > > > > +     ARM_CSPMU_EVENT_ATTR(rd_req_gmem,                       0x7),
> > > > > +     ARM_CSPMU_EVENT_ATTR(wr_req_cmem,                       0x8),
> > > > > +     ARM_CSPMU_EVENT_ATTR(wr_req_gmem,                       0x9),
> > > > > +     ARM_CSPMU_EVENT_ATTR(total_req_cmem,                    0xa),
> > > > > +     ARM_CSPMU_EVENT_ATTR(total_req_gmem,                    0xb),
> > > > > +     ARM_CSPMU_EVENT_ATTR(rd_cum_outs_cmem,                  0xc),
> > > > > +     ARM_CSPMU_EVENT_ATTR(rd_cum_outs_gmem,                  0xd),
> > > > > +     ARM_CSPMU_EVENT_ATTR(cycles,
> > > > ARM_CSPMU_EVT_CYCLES_DEFAULT),
> > > > > +     NULL,
> > > > > +};
> > > > > +
> > > > >  static struct attribute *generic_pmu_event_attrs[] = {
> > > > >       ARM_CSPMU_EVENT_ATTR(cycles,
> > > > ARM_CSPMU_EVT_CYCLES_DEFAULT),
> > > > >       NULL,
> > > > > @@ -234,7 +253,7 @@ static const struct nv_cspmu_match
> > > > nv_cspmu_match[] = {
> > > > >         .filter_default_val = NV_CNVL_FILTER_ID_MASK,
> > > > >         .name_pattern = "nvidia_cnvlink_pmu_%u",
> > > > >         .name_fmt = NAME_FMT_SOCKET,
> > > > > -       .event_attr = mcf_pmu_event_attrs,
> > > > > +       .event_attr = mcf_cnvlink_pmu_event_attrs,
> > > > >         .format_attr = cnvlink_pmu_format_attrs
> > > > >       },
> > > >
> > > > Hmm. Isn't this a user-visible change? For example, will scripts driving
> > > > 'perf' with the old event names continue to work after this patch?
> > > >
> > >
> > > Yes this is user visible. I am expecting user script to be updated accordingly.
> > > Would this be reasonable?
> > 
> > I don't think so, no. We don't tend to require userspace changes as a
> > result of upgrading the kernel.
> 
> Are you referring to userspace change just on the perf tool side?
> Cause this PMU doesn't have JSON scripts for alias/metric in the perf tool yet.

I'm not sure that matters, does it? If the mappings are exposed in sysfs,
then the tool will pick them up.

> Do you have suggestion of the proper approach?

I'd say leave the event names like they are and if you want to add aliases,
do that in userspace.

Will