Hi Kan,
On Thu, Apr 14, 2022 at 12:06 PM Liang, Kan <kan.liang@xxxxxxxxxxxxxxx> wrote:
>
>
>
> On 4/14/2022 12:09 PM, Ian Rogers wrote:
> > ```
> > $ perf stat -e '{BR_INST_RETIRED.NEAR_CALL,BR_INST_RETIRED.NEAR_TAKEN,BR_INST_RETIRED.NOT_TAKEN,cycles,cycles}:W'
> > -a sleep 1
> > Performance counter stats for 'system wide':
> >
> > <not counted> BR_INST_RETIRED.NEAR_CALL
> > (0.00%)
> > <not counted> BR_INST_RETIRED.NEAR_TAKEN
> > (0.00%)
> > <not counted> BR_INST_RETIRED.NOT_TAKEN
> > (0.00%)
> > <not counted> cycles
> > (0.00%)
> > <not counted> cycles
> > (0.00%)
> >
> > 1.005599088 seconds time elapsed
> >
> > Some events weren't counted. Try disabling the NMI watchdog:
> > echo 0 > /proc/sys/kernel/nmi_watchdog
> > perf stat ...
> > echo 1 > /proc/sys/kernel/nmi_watchdog
> > The events in group usually have to be from the same PMU. Try
> > reorganizing the group.
> > ```
> >
> > If we add two extra cycles or the original group is smaller then it is "fixed":
> > ```
> > $ perf stat -e '{BR_INST_RETIRED.NEAR_CALL,BR_INST_RETIRED.NEAR_TAKEN,BR_INST_RETIRED.NOT_TAKEN,cycles}:W'
> > -a sleep 1
> >
> > Performance counter stats for 'system wide':
> >
> > 20,378,789 BR_INST_RETIRED.NEAR_CALL
> > 168,420,963 BR_INST_RETIRED.NEAR_TAKEN
> > 96,330,608 BR_INST_RETIRED.NOT_TAKEN
> > 1,652,230,042 cycles
> >
> > 1.008757590 seconds time elapsed
> >
> > $ perf stat -e '{BR_INST_RETIRED.NEAR_CALL,BR_INST_RETIRED.NEAR_TAKEN,BR_INST_RETIRED.NOT_TAKEN,cycles,cycles,cycles}:W'
> > -a sleep 1
> >
> > Performance counter stats for 'system wide':
> >
> > 37,696,638 BR_INST_RETIRED.NEAR_CALL
> > (66.62%)
> > 298,535,151 BR_INST_RETIRED.NEAR_TAKEN
> > (66.63%)
> > 297,011,663 BR_INST_RETIRED.NOT_TAKEN
> > (66.63%)
> > 3,155,711,474 cycles
> > (66.65%)
> > 3,194,919,959 cycles
> > (66.74%)
> > 3,126,664,102 cycles
> > (66.72%)
> >
> > 1.006237962 seconds time elapsed
> > ```
> >
> > So the extra cycles is needed to fix weak groups when the nmi watchdog
> > is enabled and the group is an architecture dependent size.
>
> Yes, the size of the group depends on the architecture, but perf tool
> doesn't need to know the HW details. For this case, perf tool just sends
> the request with an extra cycles event in the group and lets kernel decide.
I prefer doing this in the kernel even if it'd be incomplete.
For the NMI watchdog, is it possible to check if it's enabled
at the moment, and set the fake_cpuc->idxmsk to prevent
scheduling events in validate_group()?
Thanks,
Namhyung
