Re: Some unc_cha_tor_* events appear to be "not supported"?

[Ian Rogers]

On Thu, Jul 25, 2024 at 9:39 AM Song Liu <song@xxxxxxxxxx> wrote:
>
> Hi Kan,
>
> Thanks for your quick response!
>
> On Thu, Jul 25, 2024 at 6:37 AM Liang, Kan <kan.liang@xxxxxxxxxxxxxxx> wrote:
> >
> >
> >
> > On 2024-07-24 6:39 p.m., Song Liu wrote:
> > > Hi folks,
> > >
> > > We noticed that some unc_cha_tor_occupancy and unc_cha_tor_inserts events
> > > appear to be "not supported" and always counting zero (see an example below).
> > > Is this a known issue? We are testing with 6.10 kernel and perf.
> > >
> > > Thanks,
> > > Song
> > >
> > >
> > > [root@kerneltest008.05.atn6 ~]# ./perf list | grep
> > > unc_cha_tor_occupancy.ia_hit_llcprefdrd
> > >   unc_cha_tor_occupancy.ia_hit_llcprefdrd
> > > [root@kerneltest008.05.atn6 ~]# ./perf stat -e
> > > unc_cha_tor_occupancy.ia_hit_llcprefdrd -a -- sleep 1
> > > WARNING: event 'N/A' not valid (bits 10,12-13,15 of config1 '4b433'
> >
> >
> > The value '4b433' is for the 'Filter1', according to the original event
> > list.
> >
> > https://github.com/intel/perfmon/blob/main/SKX/events/skylakex_uncore.json#L4634
> >
> >       "EventName": "UNC_CHA_TOR_OCCUPANCY.IA_MISS_LlcPrefDRD",
> >       "BriefDescription": "UNC_CHA_TOR_OCCUPANCY.IA_MISS_LlcPrefDRD",
> >       "PublicDescription": "UNC_CHA_TOR_OCCUPANCY.IA_MISS_LlcPrefDRD",
> >       "Counter": "0",
> >       "MSRValue": "0x00",
> >       "ELLC": "0",
> >       "Filter": "Filter1",
> >       "ExtSel": "0",
> >       "Deprecated": "0",
> >       "FILTER_VALUE": "0x4b433"
> >
> > There are two filters for CHA on SKX. Each filter is 32 bits wide.
> > So the Linux kernel driver uses config1 (64 bits wide) to represent both
> > of them. The low 32 bits are for filter0 and high 32 bits are for filter1.
> >
> > It should be an issue of the convert script, which set the filter1 value
> > to the low 32 bits.
> > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/perf/pmu-events/arch/x86/skylakex/uncore-cache.json#n4489
> >
> > Can you please try the below event?
> > uncore_cha/event=0x35,umask=0x11,config1=0x4b43300000000/
>
> This appears to work. I also tried events like:
>
> uncore_cha/UNC_CHA_TOR_OCCUPANCY.IA_MISS,config1=0x4043200000000/
> uncore_cha/UNC_CHA_TOR_INSERTS.IA_MISS,config1=0x4043200000000/
>
> Do you have an idea on how to fix the convert script? I haven't looked
> into these codes.

The converter script logic is here:
https://github.com/intel/perfmon/blob/main/scripts/create_perf_json.py#L395
but presumably this worked at some point or is needed on other models.
Checking the perf json I see:
```
$ grep -rl config1 tools/perf/pmu-events/arch/x86/
tools/perf/pmu-events/arch/x86/snowridgex/uncore-cache.json
tools/perf/pmu-events/arch/x86/skylakex/skx-metrics.json
tools/perf/pmu-events/arch/x86/skylakex/uncore-cache.json
tools/perf/pmu-events/arch/x86/cascadelakex/uncore-cache.json
tools/perf/pmu-events/arch/x86/cascadelakex/clx-metrics.json
```
Since CLX is near equal to SKX, and the metrics will presumably be
false hits, perhaps shifting this value will break snowridgex - I have
no such machine to test on.

Using config1 values rather than formats is odd to me. On a SkylakeX I see:
```
$ grep -r config1 /sys/devices/uncore_cha_0/format/
/sys/devices/uncore_cha_0/format/filter_rem:config1:32
/sys/devices/uncore_cha_0/format/filter_opc0:config1:41-50
/sys/devices/uncore_cha_0/format/filter_isoc:config1:63
/sys/devices/uncore_cha_0/format/filter_tid:config1:0-8
/sys/devices/uncore_cha_0/format/filter_loc:config1:33
/sys/devices/uncore_cha_0/format/filter_nc:config1:62
/sys/devices/uncore_cha_0/format/filter_opc1:config1:51-60
/sys/devices/uncore_cha_0/format/filter_all_op:config1:35
/sys/devices/uncore_cha_0/format/filter_not_nm:config1:37
/sys/devices/uncore_cha_0/format/filter_state:config1:17-26
/sys/devices/uncore_cha_0/format/filter_nm:config1:36
```

Adding Weilin and Perry as I'm unsure what the right fix should be. My
thought is the input Intel perfmon json should break apart the filter
value into opcodes, etc. and then in the converter script we add these
to perf events like filter_opc0=0xb4.

Thanks,
Ian

> Thanks,
> Song
>
> [...]