RE: [PATCH 2/3] perf: arm_cspmu: nvidia: update CNVLINK PMU events
From: Besar Wicaksono
Date: Thu Oct 24 2024 - 10:21:51 EST
> -----Original Message-----
> From: Will Deacon <will@xxxxxxxxxx>
> Sent: Wednesday, October 23, 2024 11:27 AM
> To: Besar Wicaksono <bwicaksono@xxxxxxxxxx>
> Cc: suzuki.poulose@xxxxxxx; robin.murphy@xxxxxxx;
> catalin.marinas@xxxxxxx; mark.rutland@xxxxxxx; linux-arm-
> kernel@xxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; linux-
> tegra@xxxxxxxxxxxxxxx; Thierry Reding <treding@xxxxxxxxxx>; Jon Hunter
> <jonathanh@xxxxxxxxxx>; Vikram Sethi <vsethi@xxxxxxxxxx>; Rich Wiley
> <rwiley@xxxxxxxxxx>; Bob Knight <rknight@xxxxxxxxxx>
> Subject: Re: [PATCH 2/3] perf: arm_cspmu: nvidia: update CNVLINK PMU
> events
>
> External email: Use caution opening links or attachments
>
>
> On Tue, Oct 15, 2024 at 05:21:06PM +0000, Besar Wicaksono wrote:
> >
> >
> > > -----Original Message-----
> > > From: Will Deacon <will@xxxxxxxxxx>
> > > Sent: Monday, October 14, 2024 8:19 AM
> > > To: Besar Wicaksono <bwicaksono@xxxxxxxxxx>
> > > Cc: suzuki.poulose@xxxxxxx; robin.murphy@xxxxxxx;
> > > catalin.marinas@xxxxxxx; mark.rutland@xxxxxxx; linux-arm-
> > > kernel@xxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; linux-
> > > tegra@xxxxxxxxxxxxxxx; Thierry Reding <treding@xxxxxxxxxx>; Jon Hunter
> > > <jonathanh@xxxxxxxxxx>; Vikram Sethi <vsethi@xxxxxxxxxx>; Rich Wiley
> > > <rwiley@xxxxxxxxxx>; Bob Knight <rknight@xxxxxxxxxx>
> > > Subject: Re: [PATCH 2/3] perf: arm_cspmu: nvidia: update CNVLINK PMU
> > > events
> > >
> > > External email: Use caution opening links or attachments
> > >
> > >
> > > On Wed, Sep 18, 2024 at 09:58:45PM +0000, Besar Wicaksono wrote:
> > > > Rename loc* and rem* events in CNVLINK PMU to cmem* and gmem*
> > > events.
> > > >
> > > > Signed-off-by: Besar Wicaksono <bwicaksono@xxxxxxxxxx>
> > > > ---
> > > > drivers/perf/arm_cspmu/nvidia_cspmu.c | 21
> ++++++++++++++++++++-
> > > > 1 file changed, 20 insertions(+), 1 deletion(-)
> > > >
> > > > 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.
Do you have suggestion of the proper approach?
Thanks,
Besar