On Mon, 6 Nov 2023, Reinette Chatre wrote: > On 11/6/2023 9:03 AM, Reinette Chatre wrote: > > On 11/6/2023 1:53 AM, Ilpo Järvinen wrote: > >> On Fri, 3 Nov 2023, Reinette Chatre wrote: > >>> On 11/3/2023 3:39 AM, Ilpo Järvinen wrote: > >>>> On Thu, 2 Nov 2023, Reinette Chatre wrote: > >>>>> On 10/24/2023 2:26 AM, Ilpo Järvinen wrote: > >>>> > >>>>>> Add L2 CAT selftest. As measuring L2 misses is not easily available > >>>>>> with perf, use L3 accesses as a proxy for L2 CAT working or not. > >>>>> > >>>>> I understand the exact measurement is not available but I do notice some > >>>>> L2 related symbolic counters when I run "perf list". l2_rqsts.all_demand_miss > >>>>> looks promising. > >>>> > >>>> Okay, I was under impression that L2 misses are not available. Both based > >>>> on what you mentioned to me half an year ago and because of what flags I > >>>> found from the header. But I'll take another look into it. > >>> > >>> You are correct that when I did L2 testing a long time ago I used > >>> the model specific L2 miss counts. I was hoping that things have improved > >>> so that model specific counters are not needed, as you have tried here. > >>> I found the l2_rqsts symbol while looking for alternatives but I am not > >>> familiar enough with perf to know how these symbolic names are mapped. > >>> I was hoping that they could be a simple drop-in replacement to > >>> experiment with. > >> > >> According to perf_event_open() manpage, mapping those symbolic names > >> requires libpfm so this would add a library dependency? > > > > I do not see perf list using this library to determine the event and > > umask but I am in unfamiliar territory. I'll have to spend some more > > time here to determine options. > > tools/perf/pmu-events/README cleared it up for me. The architecture specific > tables are included in the perf binary. Potentially pmu-events.h could be > included or the test could just stick with the architectural events. > A quick look at the various cache.json files created the impression that > the events of interest may actually have the same event code and umask across > platforms. > I am not familiar with libpfm. This can surely be considered if it supports > this testing. Several selftests have library dependencies. man perf_event_open() says this: "If type is PERF_TYPE_RAW, then a custom "raw" config value is needed. Most CPUs support events that are not covered by the "generalized" events. These are implementation defined; see your CPU manual (for ex- ample the Intel Volume 3B documentation or the AMD BIOS and Kernel De- veloper Guide). The libpfm4 library can be used to translate from the name in the architectural manuals to the raw hex value perf_event_open() expects in this field." ...I've not come across libpfm myself either but to me it looks libpfm bridges between those architecture specific tables and perf_event_open(). That is, it could provide the binary value necessary in constructing the perf_event_attr struct. I think this is probably the function which maps string -> perf_event_attr: https://man7.org/linux/man-pages/man3/pfm_get_os_event_encoding.3.html -- i.
