How-To Perf and CoreFreq

[cyring]

• For some reasons Perf is back on work with its version 5.13.g62fb9874f5da
  No more perf top crash during its startup while CoreFreq is driving my AMD Ryzen 3950X
  I can now fine tune the hot spot in the Client which is obviously the FuseAll() function

• In screenshot above, various kernel modules have been blacklisted to leave the room to CoreFreq for its registrations.

cat /proc/cmdline

... nmi_watchdog=0 selinux=0 modprobe.blacklist=nouveau,joydev,mousedev,pcspkr,k10temp,acpi_cpufreq,rapl,intel_rapl_msr,intel_rapl_common,sp5100_tco,wmi,eeepc_wmi,asus_wmi idle=halt tsc=unstable acpi_enforce_resources=lax cpu0_hotplug audit=0 nowatchdog mitigations=off nokaslr sysrq_always_enabled

• Starting CoreFreq

## Build with CoreFreq's delay implementation (TSC based)
## Employ the family 17h dedicated APERF/MPERF counters

make DELAY_TSC=1 MSR_CORE_PERF_UCC=MSR_AMD_F17H_APERF MSR_CORE_PERF_URC=MSR_AMD_F17H_MPERF NO_UPPER=1 clean all

## Prepare kernel to switch to CoreFreq as the clock source
echo "corefreq" > /sys/devices/system/clocksource/clocksource0/current_clocksource

## Allows Driver to register to any kernel services
insmod corefreqk.ko Register_ClockSource=1 RDPMC_Enable=1 Register_Governor=1 Register_CPU_Freq=1 Register_CPU_Idle=1 Idle_Route=1 NMI_Disable=0

## Start the Daemon
corefreqd -d

## Start the Client in `Interrupts` view
corefreq-cli -t interrupts

• Starting Perf

perf top

kernel log is flooded with NMI messages which are also counted by CoreFreq

kernel: Uhhuh. NMI received for unknown reason 3d on CPU 20.
kernel: Do you have a strange power saving mode enabled?
kernel: Dazed and confused, but trying to continue

• Stopping CoreFreq
  You may face this message while exiting the Client

Press Enter and confirm to unregister the Driver from the kernel CPU-IDLE framework.

Ask me if you need more details