Re: [PATCH v5 2/2] perf vendor events hygon: Add Hygon Data Fabric PMU events and metrics

[Qi Liu]

Hi Sandipan,

Sorry for the delayed reply.

On 8/19/2026 1:11 PM, Sandipan Das wrote:
> On 18-08-2026 15:31, Qi Liu wrote:
> > Add perf JSON event definitions and metrics for the Hygon Data Fabric (DF)
> > uncore PMU.
> >
> > Signed-off-by: Qi Liu <liuqi@xxxxxxxx>
> > ---
> >   tools/perf/pmu-events/Build                   |   2 +-
> >   .../arch/x86/hygonm1/data-fabric.json         |  95 ++++++++++
> >   .../arch/x86/hygonm4/data-fabric.json         | 134 ++++++++++++++
> >   .../arch/x86/hygonm6/data-fabric.json         | 170 ++++++++++++++++++
> >   tools/perf/pmu-events/arch/x86/mapfile.csv    |   3 +
> >   tools/perf/pmu-events/jevents.py              |   2 +
> >   6 files changed, 405 insertions(+), 1 deletion(-)
> >   create mode 100644 tools/perf/pmu-events/arch/x86/hygonm1/data-fabric.json
> >   create mode 100644 tools/perf/pmu-events/arch/x86/hygonm4/data-fabric.json
> >   create mode 100644 tools/perf/pmu-events/arch/x86/hygonm6/data-fabric.json
> >
> > diff --git a/tools/perf/pmu-events/Build b/tools/perf/pmu-events/Build
> > index 372773b998e4..79a9b4cad202 100644
> > --- a/tools/perf/pmu-events/Build
> > +++ b/tools/perf/pmu-events/Build
> > @@ -95,7 +95,7 @@ endif
> >   
> >   ifeq ($(JEVENTS_ARCH),$(filter $(JEVENTS_ARCH),x86 all))
> >   # Generate Intel Json
> > -INTELS := $(shell ls -d pmu-events/arch/x86/*|grep -v amdzen|grep -v mapfile.csv)
> > +INTELS := $(shell ls -d pmu-events/arch/x86/*|grep -v amdzen|grep -v hygon|grep -v mapfile.csv)
> >   INTEL_METRICS = $(foreach x,$(INTELS),$(OUTPUT)$(x)/extra-metrics.json)
> >   INTEL_METRICGROUPS = $(foreach x,$(INTELS),$(OUTPUT)$(x)/extra-metricgroups.json)
> >   GEN_JSON += $(INTEL_METRICS) $(INTEL_METRICGROUPS)
> > diff --git a/tools/perf/pmu-events/arch/x86/hygonm1/data-fabric.json b/tools/perf/pmu-events/arch/x86/hygonm1/data-fabric.json
> > new file mode 100644
> > index 000000000000..8c6a8e9ba9a5
> > --- /dev/null
> > +++ b/tools/perf/pmu-events/arch/x86/hygonm1/data-fabric.json
> > @@ -0,0 +1,95 @@
> > +[
> > +  {
> > +    "EventName": "cpu_read_requests_cs_0",
> > +    "BriefDescription": "Number of CPU read requests to memory in 64-byte units for Coherent Subordinate 0 (CS0).",
>
> Coherent Station is the correct term, if you go by the terminology in AMD
> PPRs. Since CSes are the interfaces between the data fabric and the memory
> controllers (UMCs), it may be better to rephrase the descriptions as read
> or write requests to DRAM channel 0..N which are backed by UMC 0..N and
> CS 0..N.

Thanks,I will update "Coherent Subordinate" to "Coherent Station".

Yes, CSes provide the interface between the data fabric and UMCs. Since the 
UMC-to-DRAM-channel topology may vary across platforms, would it be reasonable 
to avoid referring to DRAM channels in the event descriptions?

I was thinking of wording the description as follows:
  "Number of CPU read requests associated with Coherent Station 0 (CS0) and the 
UMC."

> > +    "EventCode": "0x2",
> > +    "UMask": "0x07",
> > +    "CompId": "0x0",
> > +    "Unit": "HYGON_DF",
> > +    "ScaleUnit": "64Bytes"
>
> As the naming suggests, my understanding is that these events count the
> number of requests and each request carries a 64-byte payload. Perhaps
> the ScaleUnit needs to be factored into the bandwidth metrics instead of
> the individual events.

Thanks,I will remove the "64Bytes" ScaleUnit from the individual request
events in v6.

The bandwidth MetricExprs already apply the payload size: * 64 for the
CPU bandwidth and I/O upper-bound metrics, and * 32 for the I/O
lower-bound metric. I plan to retain those conversions in the metrics.

Does that match your suggestion?

Thanks,
Qi