Re: [PATCH v1] perf test: Be more tolerant of metricgroup failures

[Arnaldo Carvalho de Melo]

On Mon, Apr 22, 2024 at 08:42:17AM -0700, Ian Rogers wrote:
> On Mon, Apr 22, 2024 at 4:51 AM Veronika Molnarova <vmolnaro@xxxxxxxxxx> wrote:
> >
> > Hi Ian,
> >
> > On Wed, Apr 3, 2024 at 6:48 PM Ian Rogers <irogers@xxxxxxxxxx> wrote:
> >>
> >> Previously "set -e" meant any non-zero exit code from perf stat would
> >> cause a test failure. As a non-zero exit happens when there aren't
> >> sufficient permissions, check for this case and make the exit code
> >> 2/skip for it.
> >>
> >> Signed-off-by: Ian Rogers <irogers@xxxxxxxxxx>
> >> ---
> >>  .../perf/tests/shell/stat_all_metricgroups.sh | 28 +++++++++++++++----
> >>  1 file changed, 22 insertions(+), 6 deletions(-)
> >>
> >> diff --git a/tools/perf/tests/shell/stat_all_metricgroups.sh b/tools/perf/tests/shell/stat_all_metricgroups.sh
> >> index 55ef9c9ded2d..d6db192b9f18 100755
> >> --- a/tools/perf/tests/shell/stat_all_metricgroups.sh
> >> +++ b/tools/perf/tests/shell/stat_all_metricgroups.sh
> >> @@ -1,9 +1,7 @@
> >> -#!/bin/sh
> >> +#!/bin/bash
> >>  # perf all metricgroups test
> >>  # SPDX-License-Identifier: GPL-2.0
> >>
> >> -set -e
> >> -
> >>  ParanoidAndNotRoot()
> >>  {
> >>    [ "$(id -u)" != 0 ] && [ "$(cat /proc/sys/kernel/perf_event_paranoid)" -gt $1 ]
> >> @@ -14,11 +12,29 @@ if ParanoidAndNotRoot 0
> >>  then
> >>    system_wide_flag=""
> >>  fi
> >> -
> >> +err=0
> >>  for m in $(perf list --raw-dump metricgroups)
> >>  do
> >>    echo "Testing $m"
> >> -  perf stat -M "$m" $system_wide_flag sleep 0.01
> >> +  result=$(perf stat -M "$m" $system_wide_flag sleep 0.01 2>&1)
> >> +  result_err=$?
> >> +  if [[ $result_err -gt 0 ]]
> >> +  then
> >> +    if [[ "$result" =~ \
> >> +          "Access to performance monitoring and observability operations is limited" ]]
> >> +    then
> >> +      echo "Permission failure"
> >> +      echo $result
> >> +      if [[ $err -eq 0 ]]
> >> +      then
> >> +        err=2 # Skip
> >> +      fi
> >> +    else
> >> +      echo "Metric group $m failed"
> >> +      echo $result
> >> +      err=1 # Fail
> >> +    fi
> >> +  fi
> >>  done
> >>
> >> -exit 0
> >> +exit $err
> >> --
> >> 2.44.0.478.gd926399ef9-goog
> >>
> >>
> >
> > The patch looks good and thanks for taking care of it.
> >
> > Just wanted to check what is the desired outcome for metric groups
> > with events that are invalid in per-thread mode causing the test to fail.
> >
> > ```
> > $ ./stat_all_metricgroups.sh
> > Testing smi
> > Metric group smi failed
> > Error: Invalid event (msr/smi/u) in per-thread mode, enable system wide with '-a'.
> > ```
> >
> > Wouldn't it be better if in these cases the test would result in skip instead of fail?
> 
> Hi Veronika,
> 
> I agree that fail isn't best here. I'm wondering:
> 
>  - why doesn't msr/smi/ support per-thread mode? Can't we save/restore
> the count on a context switch? The implementation is here:
> https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/arch/x86/events/msr.c?h=perf-tools-next#n234
> There's clearly something going on as pperf appears to have other restrictions:
> https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/arch/x86/events/msr.c?h=perf-tools-next#n115
> I'm wondering if aggregation is working right if these counters are
> more than per hyperthread (I'm guessing why the restrictions exist).
> 
>  - the tool error message is doing pretty good. In the test I guess we
> can spot the per-thread error and turn the fail to a skip. It's a
> shame to bucket things as skip, but it seems easier than listing
> metrics in the test or spotting particular events.
> 
> I can do a v2 to add this.

Ok, I'll wait for v2 then.

- Arnaldo