Re: [PATCH] perf tests: Add auto counter reload (ACR) sampling test

[Ian Rogers]

On Wed, Apr 15, 2026 at 8:06 PM Mi, Dapeng <dapeng1.mi@xxxxxxxxxxxxxxx> wrote:
>
>
> On 4/16/2026 1:13 AM, Ian Rogers wrote:
> > On Wed, Apr 15, 2026 at 2:19 AM Mi, Dapeng <dapeng1.mi@xxxxxxxxxxxxxxx> wrote:
> >>
> >> On 4/13/2026 9:30 AM, Ian Rogers wrote:
> >>> On Sun, Apr 12, 2026 at 6:14 PM Dapeng Mi <dapeng1.mi@xxxxxxxxxxxxxxx> wrote:
> >>>> Add auto counter reload sampling test to verify that the intended event
> >>>> records can be captured and the self-reloaded events won't generate any
> >>>> records.
> >>>>
> >>>> Signed-off-by: Dapeng Mi <dapeng1.mi@xxxxxxxxxxxxxxx>
> >>>> ---
> >>>>  tools/perf/tests/shell/record.sh | 42 ++++++++++++++++++++++++++++++++
> >>>>  1 file changed, 42 insertions(+)
> >>>>
> >>>> diff --git a/tools/perf/tests/shell/record.sh b/tools/perf/tests/shell/record.sh
> >>>> index 7cb81cf3444a..1068843282f5 100755
> >>>> --- a/tools/perf/tests/shell/record.sh
> >>>> +++ b/tools/perf/tests/shell/record.sh
> >>>> @@ -402,6 +402,47 @@ test_callgraph() {
> >>>>    echo "Callgraph test [Success]"
> >>>>  }
> >>>>
> >>>> +test_acr_sampling() {
> >>>> +  events="{instructions/period=20000,acr_mask=0x2/u,cycles/period=40000,acr_mask=0x3/u}"
> >>>> +  pebs_events="{instructions/period=20000,acr_mask=0x2/pu,cycles/period=40000,acr_mask=0x3/u}"
> >>>> +  echo "Auto counter reload sampling test"
> >>>> +  if ! perf record -o "${perfdata}" -e "${events}" ${testprog} 2> /dev/null
> >>>> +  then
> >>>> +    echo "Auto counter reload sampling [Skipped not supported]"
> >>>> +    return
> >>>> +  fi
> >>>> +  if ! perf script -i "${perfdata}" | grep -q "period=20000,acr_mask=0x2"
> >>>> +  then
> >>>> +    echo "Auto counter reload sampling [Failed missing instructions event]"
> >>>> +    err=1
> >>>> +    return
> >>>> +  fi
> >>>> +  if perf script -i "${perfdata}" | grep -q "period=40000,acr_mask=0x3"
> >>>> +  then
> >>>> +    echo "Auto counter reload sampling [Failed cycles event shouldn't be sampled]"
> >>>> +    err=1
> >>>> +    return
> >>>> +  fi
> >>>> +  if ! perf record -o "${perfdata}" -e "${pebs_events}" ${testprog} 2> /dev/null
> >>>> +  then
> >>>> +    echo "Auto counter reload PEBS sampling [Skipped not supported]"
> >>>> +    return
> >>>> +  fi
> >>>> +  if ! perf script -i "${perfdata}" | grep -q "period=20000,acr_mask=0x2"
> >>>> +  then
> >>>> +    echo "Auto counter reload PEBS sampling [Failed missing instructions event]"
> >>>> +    err=1
> >>>> +    return
> >>>> +  fi
> >>>> +  if perf script -i "${perfdata}" | grep -q "period=40000,acr_mask=0x3"
> >>>> +  then
> >>>> +    echo "Auto counter reload PEBS sampling [Failed cycles event shouldn't be sampled]"
> >>>> +    err=1
> >>>> +    return
> >>>> +  fi
> >>>> +  echo "Auto counter reload sampling [Success]"
> >>> Thanks Dapeng! Could we also test ratio-to-prev as well as just the
> >>> plain acr_mask?
> >> Ian, it could be hard to validate ratio-to-prev option in this test. On
> >> platforms which don't support ACR, perf would automatically fallback the
> >> ACR sampling to normal sampling, this would lead to false positives on
> >> checking the existence of records from cycles event. I didn't find a good
> >> way to figure out this fallback.
> >>
> >> Since we already have independent test case to validate the ratio_to_prev
> >> parsing. It may be unnecessary to add ratio-to-prev validation in this ACR
> >> test.
> > So the ratio_to_prev test covers parsing the ratio-to-prev option (it
> > only has `perf record` commands) and ensures it is acted upon, but it
> > doesn't check the option's functional behavior - what the output is.
> > This test makes sure the acr_mask option is functional and so would be
> > complemented by ratio-to-prev testing somethig like:
> > ```
> >   rtp_events="{instructions,cycles/period=40000,ratio-to-prev=0.5/u}"
> >
> >   # Test ratio-to-prev
> >   if ! perf record -o "${perfdata}" -e "${rtp_events}" ${testprog} 2> /dev/null
> >   then
> >     echo "Auto counter reload ratio-to-prev sampling [Failed record
> > ratio-to-prev]"
> >     err=1
> >     return
> >   fi
> >   if ! perf script -i "${perfdata}" | grep -q "instructions"
> >   then
> >     echo "Auto counter reload ratio-to-prev sampling [Failed missing
> > instructions event]"
> >     err=1
> >     return
> >   fi
> >   if perf script -i "${perfdata}" | grep -q "cycles"
> >   then
> >     echo "Auto counter reload ratio-to-prev sampling [Failed cycles
> > event shouldn't be sampled]"
> >     err=1
> >     return
> >   fi
> > ```
> > The first thing your test does is to make sure the acr_mask is
> > supported, so I'm not sure how the fall back can happen?
> > We could additionally check for
> > `/sys/bus/event_source/devices/cpu*/format/acr_mask`.
>
> Yes, that's initially what I intended to do, but it would cause false
> positives on some hybrid platform, like PTL which only supports ACR on E-core.
>
> ```
>
> $ grep . /sys/bus/event_source/devices/cpu_core/format/*
> /sys/bus/event_source/devices/cpu_core/format/cmask:config:24-31
> /sys/bus/event_source/devices/cpu_core/format/edge:config:18
> /sys/bus/event_source/devices/cpu_core/format/eq:config:36
> /sys/bus/event_source/devices/cpu_core/format/event:config:0-7
> /sys/bus/event_source/devices/cpu_core/format/frontend:config1:0-23
> /sys/bus/event_source/devices/cpu_core/format/inv:config:23
> /sys/bus/event_source/devices/cpu_core/format/ldlat:config1:0-15
> /sys/bus/event_source/devices/cpu_core/format/metrics_clear:config1:0
> /sys/bus/event_source/devices/cpu_core/format/offcore_rsp:config1:0-63
> /sys/bus/event_source/devices/cpu_core/format/pc:config:19
> /sys/bus/event_source/devices/cpu_core/format/umask:config:8-15,40-47
>
> $ grep . /sys/bus/event_source/devices/cpu_atom/format/*
> /sys/bus/event_source/devices/cpu_atom/format/acr_mask:config2:0-63
> /sys/bus/event_source/devices/cpu_atom/format/cmask:config:24-31
> /sys/bus/event_source/devices/cpu_atom/format/edge:config:18
> /sys/bus/event_source/devices/cpu_atom/format/eq:config:36
> /sys/bus/event_source/devices/cpu_atom/format/event:config:0-7
> /sys/bus/event_source/devices/cpu_atom/format/inv:config:23
> /sys/bus/event_source/devices/cpu_atom/format/ldlat:config1:0-15
> /sys/bus/event_source/devices/cpu_atom/format/offcore_rsp:config1:0-63
> /sys/bus/event_source/devices/cpu_atom/format/pc:config:19
> /sys/bus/event_source/devices/cpu_atom/format/snoop_rsp:config1:0-63
> /sys/bus/event_source/devices/cpu_atom/format/umask:config:8-15,40-47
>
> ```
>
> So strictly speaking, we need to check if it's a hybrid platform first and
> then check if the exact P-core/E-core supports ACR, then run the above ACR
> tests. It seems a little bit over-complicated and I'm not sure if it's
> accepted to add so much Intel specific check in the record.sh.
>
> How's your idea on this? Thanks.

Potentially auto counter reload and ratio to prev are generic
concepts, it just happens that we only have Intel e-core
implementations :-) Matching against cpu* already misses ARM core
PMUs. I'm always happy for as much testing as possible. The last code
coverage number I have for Linux v6.17 is 38.6% of lines under
tools/perf. Since then, I've added quite a bit of testing for commands
like `perf top`. Perhaps we now cover over 50% of the lines of code.

Thanks,
Ian

> >
> > Thanks,
> > Ian
> >
> >>> Thanks,
> >>> Ian
> >>>
> >>>> +}
> >>>> +
> >>>>  test_ratio_to_prev() {
> >>>>    echo "ratio-to-prev test"
> >>>>    if ! perf record -o /dev/null -e "{instructions, cycles/period=100000,ratio-to-prev=0.5/}" \
> >>>> @@ -457,6 +498,7 @@ test_leader_sampling
> >>>>  test_topdown_leader_sampling
> >>>>  test_precise_max
> >>>>  test_callgraph
> >>>> +test_acr_sampling
> >>>>  test_ratio_to_prev
> >>>>
> >>>>  # restore the default value
> >>>>
> >>>> base-commit: 4e03d6494f9504f8af46ba68a2a8b6877c196789
> >>>> --
> >>>> 2.34.1
> >>>>