Messages in this thread | | | From | Ian Rogers <> | Date | Sun, 17 Mar 2024 16:04:17 -0700 | Subject | Re: [PATCH v1] perf vendor events intel: Add umasks to PCU events. |
| |
On Fri, Mar 15, 2024 at 1:29 PM Arnaldo Carvalho de Melo <acme@kernel.org> wrote: > > On Wed, Feb 28, 2024 at 08:12:10AM -0800, Ian Rogers wrote: > > On Wed, Feb 28, 2024 at 6:55 AM Liang, Kan <kan.liang@linux.intel.com> wrote: > > > On 2024-02-27 9:12 p.m., Ian Rogers wrote: > > > > Kan, what's the right way to deal with this? > > > > > > There is no umask in the uncore spec. Please see P120 of the spec. > > > https://www.intel.com/content/www/us/en/develop/download/intel-xeon-processor-e5-2600-v2-product-family-uncore-performance-monitoring-reference.html > > > > > > It should be occ_sel. > > > The occ_sel = (the umask from the event list >> 6) & 0x3. > > > > > > I don't think we want to change either the kernel or the perf tool. > > > There is nothing to help either if we add the occ_sel in the original > > > event list. So it seems the only choice should be handling it in the > > > convertor tool. > > > > > > Is it possible to check the UNC_P_POWER_STATE_OCCUPANCY.CORES_C0 in IVT > > > and use "Filter": "occ_sel=0x1" to replace "UMask": "0x40"? > > > > > > It seems everything in the filter will directly be appended. Is my > > > understanding correct? > > > > Yep. From the manual: > > C0 = 1 > > C3 = 2 > > C6= 3 > > So I'll need to fix all 3 events during the conversion to perf json. > > The same problem exists for haswell and broadwell, so I'll do the fix > > on those architectures too. I was able to test on a jaketown but the > > kernel source makes me think the same fix is necessary there too. I'll > > send out a new patch with this. > > Please clarify if this remains reviewed and should be merged as-is.
v2. is already applied in perf-tools-next [1] so we don't need v1. Thanks, Ian
[1] https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/commit/?h=perf-tools-next&id=dd267d056fed323f1684fa52d2a864fc93ca3be0
> - Arnaldo >
| |