Re: IBS perf test failures on 9950x3d

[Arnaldo Carvalho de Melo <acme@kernel.org>]

On Sun, May 18, 2025 at 11:22:24AM -0700, Namhyung Kim wrote:
> On Sat, May 17, 2025 at 05:33:19PM +0530, Ravi Bangoria wrote:
> > Hi Arnaldo,

> > > If it needs a specific kernel, then the test needs to cope with that,
> > > etc.

> > 1) Bunch of IBS kernel fixes went in recently (v6.15-rc1):

> >    https://lore.kernel.org/r/20250115054438.1021-1-ravi.bangoria@amd.com

> >    Since you are running on 6.14.5-300.fc42.x86_64, the failures are
> >    genuine and indicating that kernel should be fixed.

Telling that to the user and possibly skipping the test if viable using
uname to get the version and if less than v6.15-rc1 print "Skip (needs
v6.15-rc1 or newer)" may be an option.

Then if the bug somehow reappears, people running 'perf test' will flag
it.

> >    I'm wondering how would the test cope with the buggy kernel. Or
> >    probably 'perf test' a wrong place to put such selftests?
 
> Maybe... some kernel internal changes can be tested in the selftest.
 
> And 'perf test' should work with any kernel versions.  Is the test
> checking any invalid configs?  Probably then it should move to the
> selftest so that it can be tied to the fixed kernels and 'perf test'
> should only use valid configs.

Right, but then there are cases where something changed and then
'perf test' should adapt to accept new behaviour while keeping what made
sense for older kernels.

The fix Leo sent recently is one such case.

https://lore.kernel.org/all/20250519082755.1669187-1-leo.yan@arm.com/T/#u

And I think having it in 'perf test' as well may make the feature to be
tested more widely, both by those who run selftests as well as by people
trying just 'perf test'.

- Arnaldo