Re: [PATCH 0/2] perf vendor events arm64: Add A720/A520 events/metrics
From: James Clark
Date: Wed Feb 19 2025 - 10:26:30 EST
On 18/02/2025 10:33 pm, Ian Rogers wrote:
On Tue, Feb 18, 2025 at 2:19 PM Namhyung Kim <namhyung@xxxxxxxxxx> wrote:
On Tue, Feb 18, 2025 at 09:30:23AM +0000, James Clark wrote:
On 18/02/2025 12:41 am, Ian Rogers wrote:
On Fri, Feb 14, 2025 at 2:02 AM James Clark <james.clark@xxxxxxxxxx> wrote:
On 14/02/2025 5:49 am, Yangyu Chen wrote:
On 14 Feb 2025, at 09:12, Namhyung Kim <namhyung@xxxxxxxxxx> wrote:
Hello,
On Thu, Feb 13, 2025 at 11:11:01PM +0800, Yangyu Chen wrote:
This patchset adds the perf JSON files for the Cortex-A720 and Cortex-A520
processors. Some events have been tested on Raxda Orion 6 with Cix P1 SoC
(8xA720 + 4xA520) running mainline Kernel with ACPI mode.
I'm curious how the name of PMUs look like. It is cortex_a720 (or a520)?
The name of PMUs comes from Arm's documentation. I have included these
links in each patch.
I remember there's a logic to check the length of hex digits at the end.
Could you provide more details about this?
Ian, are you ok with this?
I think they wouldn't be merged because they're core PMUs, so should be
fine? Even though they would otherwise be merged because they're more
than 3 hex digits.
Do we know the PMU names? If they are cortex_a520 and cortex_a720 then
It will be "armv9_cortex_a720" from this line:
PMUV3_INIT_SIMPLE(armv9_cortex_a720)
I see, thanks!
this comment at least reads a little stale:
https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/util/pmus.c?h=perf-tools-next#n76
```
/*
* There is a '_{num}' suffix. For decimal suffixes any length
* will do, for hexadecimal ensure more than 2 hex digits so
* that S390's cpum_cf PMU doesn't match.
*/
```
James is right that core PMUs aren't put on the same list as uncore/other PMUs.
Ok, then I guess we're good.
I think you may be able to do things that look odd, like today the
"i915" PMU can be called just "i", I think the a520/a720 naming will
allow "armv9_cortex/cycles/" as an event name, then open it on two
PMUs if they are present.
I assumed that was the intended behavior. It seems fairly useful to be
able to open on ones with common prefixes.
We may only show one PMU in perf list as
that code I think assumes they're the same PMU as they only differ by
suffix:
https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/util/pmus.c?h=perf-tools-next#n384
Yeah that is the case. I didn't realise it when looking at the previous
fixes to keep the suffixes in perf stat output.
I can imagine aggregation possibly being broken, but I think that
works off the number of PMUs not the names of the PMUs, so it should
be okay. Probably the only thing broken that matter is perf list when
you have a BIG.little system with a520 and a720, this may be broken
with say a a53 and a72 today as both of those suffix lengths are >2,
but maybe they use the "armv8._pmuv3_0", "armv8._pmuv3_1", etc. naming
convention. I suspect the >2 here:
Also the case for a53 and a72 right now. Even "perf list --unit
armv8_cortex_a57" doesn't work because we deduplicate before filtering.
Adding -v fixes it though because that disables deduplication. Perhaps
we can change it to disable it with the --unit argument?
https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/util/pmus.c?h=perf-tools-next#n80
would still work and be correct if it were >4. If that changes then
this will also need to change:
https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/Documentation/ABI/testing/sysfs-bus-event_source-devices?h=perf-tools-next#n12
That could be an easy fix. If >4 is enough to still get rid of all the
uncore duplicates I can make the change?
Thanks,
Ian
Thanks,
Namhyung