Skip to content

Latest commit

 

History

History
506 lines (403 loc) · 23.8 KB

README.md

File metadata and controls

506 lines (403 loc) · 23.8 KB

CoreFreq

Purpose

CoreFreq, a CPU monitoring software with BIOS like functionalities, is designed for the 64-bits Processors of architecture Intel Atom, Core2, Nehalem, SandyBridge and superiors; AMD Families from 0Fh ... up to 17h (Zen , Zen+ , Zen2), 18h (Hygon Dhyana), 19h (Zen3)

alt text

CoreFreq provides a framework to retrieve CPU data with a high degree of precision:

  • Core frequencies & ratios; SpeedStep (EIST), Turbo Boost, Hyper-Threading (HTT) and Base Clock
  • Performance counters including Time Stamp Counter (TSC), Unhalted Core Cycles (UCC), Unhalted Reference Cycles (URC)
  • Number of instructions per cycle or second, IPS, IPC, or CPI
  • CPU C-States C0 C1 C3 C6 C7 - C1E - Auto/UnDemotion of C1 C3
  • DTS Temperature and Tjunction Max, Thermal Monitoring TM1 TM2 state, Vcore
  • Topology map including Caches for boostrap & application CPU
  • Processor features, brand & architecture strings
  • In progress: Uncore, Memory Controller channels & geometry, DIMM timings,
    Stress tools, Power & Energy (RAPL, P-State, HWP, TDP), Overclocking, cpuidle & cpufreq driver, ClockSource, Mitigation Mechanisms

To reach this goal, CoreFreq implements a Linux Kernel module which employs the followings:

  • asm code to keep as near as possible the readings of the performance counters;
  • per-CPU, implements slab data memory and high-resolution timer;
  • compliant with suspend / resume and CPU Hot-Plug;
  • a shared memory to protect kernel from the user-space part of the software;
  • atomic synchronization of threads to avoid mutexes and deadlock.

Build & Run

Prerequisites

a- Intel only: For a better accuracy, disable the Kernel NMI Watchdog

Add the below parameter in the kernel boot loader { Grub, SysLinux } ...

nmi_watchdog=0

... and build with the fixed performance counters

make MSR_CORE_PERF_UC=MSR_CORE_PERF_FIXED_CTR1 MSR_CORE_PERF_URC=MSR_CORE_PERF_FIXED_CTR2

b- AMD and Intel: Some Virtualization

VMs don't provide access to the registers that the CoreFreq driver employs :

  • Fixed Performance Counters
  • Model Specific Registers
  • PCI Registers

However CoreFreq is making use of the virtualized performance counter :

  • HV_X64_MSR_VP_RUNTIME(0x40000010)

c- Rendering

The UI renders best with an ASCII console or a Xterm with VT100 support, ANSI colors; optionally transparency.
If bold and bright colors are not rendered then use the following terminal options:

Ubuntu Terminal

In the Preferences - Colors tab, select Show bold text in bright colors

alacritty terminal

Uncomment and set draw_bold_text_with_bright_colors: true in <config-file>

Dependencies

  • The Linux Kernel with a minimum version 3.3
  • The GNU C Library

Build

  1. Software needed:
  • GNU C Compiler with GNU extensions
  • GNU Make tool
  • Linux Kernel Header files to build modules
    • Mandatory : CONFIG_MODULES, CONFIG_SMP, CONFIG_X86_MSR
    • Optionally: CONFIG_HOTPLUG_CPU, CONFIG_CPU_IDLE, CONFIG_CPU_FREQ, CONFIG_PM_SLEEP, CONFIG_DMI, CONFIG_XEN, CONFIG_AMD_NB, CONFIG_HAVE_PERF_EVENTS, CONFIG_SCHED_MUQSS, CONFIG_SCHED_BMQ, CONFIG_SCHED_PDS
  1. Clone the source code into a working directory.
    git clone https://github.com/cyring/CoreFreq.git

  2. Build the programs.
    cd CoreFreq
    make

cc -Wall -pthread -c corefreqd.c -o corefreqd.o
cc -Wall -c corefreqm.c -o corefreqm.o
cc corefreqd.c corefreqm.c -o corefreqd -lpthread -lm -lrt
cc -Wall -c corefreq-cli.c -o corefreq-cli.o
cc -Wall -c corefreq-ui.c -o corefreq-ui.o
cc corefreq-cli.c corefreq-ui.c -o corefreq-cli -lm -lrt
make -C /lib/modules/x.y.z/build M=/workdir/CoreFreq modules
make[1]: Entering directory '/usr/lib/modules/x.y.z/build'
  CC [M]  /workdir/CoreFreq/corefreqk.o
  Building modules, stage 2.
  MODPOST 1 modules
  CC      /workdir/CoreFreq/corefreqk.mod.o
  LD [M]  /workdir/CoreFreq/corefreqk.ko
make[1]: Leaving directory '/usr/lib/modules/x.y.z/build'
  1. (Optionally) Sign the driver If module signature verification is enabled into Kernel, you will have to sign the corefreqk.ko driver.

Install

Manual

  1. Copying CoreFreq into the binaries directory
    make install

Distribution package

  1. Although CoreFreq is released in the ArchLinux AUR ; other sources of distribution may require to reload the systemd daemons:
    systemctl daemon-reload

Start

  1. When built from source code:
  • Load the kernel module, from current directory, as root.
    insmod corefreqk.ko
  • Start the daemon, as root.
    corefreqd
  • Start the client, as a user (in another terminal or console).
    corefreq-cli
  1. When manually installed or from a distribution package:
  • Load the kernel module, as root.
    modprobe corefreqk
  • Start the daemon, as root.
    systemctl start corefreqd
  • Start the client, as a user.
    corefreq-cli

Stop

  1. Press Ctrl+x or Ctrl+c to stop the client.

  2. Press Ctrl+c to stop the daemon (in foreground) or kill its background job.

  3. Unload the kernel module
    rmmod corefreqk.ko

Try

Download the CoreFreq Live CD from the Wiki
alt text

Screenshots

Linux kernel module

Use lsmod, dmesg or journalctl -k to check if the module is started:

CoreFreq: Processor [06_1A] Architecture [Nehalem/Bloomfield] CPU [8/8]

Daemon

CoreFreq Daemon.  Copyright (C) 2015-2021 CYRIL INGENIERIE

  Processor [Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz]
  Architecture [Nehalem/Bloomfield] 8/8 CPU Online.

Client

Without arguments, the corefreq-cli program displays Top Monitoring
alt text
Remark: Drawing will stall if the terminal width is lower than 80 columns, or its height is less than required.

  • With the option '-c', the client traces counters. alt text

  • Using option '-m' corefreq-cli shows the CPU topology alt text

  • With the option '-i' corefreq-cli traces the number of instructions per second / cycle

CPU     IPS            IPC            CPI
#00     0.000579/s     0.059728/c    16.742698/i
#01     0.000334/s     0.150569/c     6.641471/i
#02     0.000598/s     0.161326/c     6.198641/i
#03     0.000294/s     0.233535/c     4.282013/i
#04     0.000240/s     0.042931/c    23.293141/i
#05     0.000284/s     0.158661/c     6.302765/i
#06     0.000128/s     0.128031/c     7.810631/i
#07     0.000088/s     0.150406/c     6.648674/i
  • Use the option '-s' to show the Processor information (BSP)

alt text

ArchLinux

  • corefreq follows released tags
  • corefreq-git follows develop branch

Debian, Ubuntu

  • Installing the DKMS package will pull the Kernel development packages
    apt-get install dkms
  • Or, install selectively the development packages prerequisites.
    apt-get install libpthread-stubs0-dev

Red Hat, CentOS

  • Development packages prerequisites.
    yum install kernel-devel
    yum group install "Development Tools"

Q&A

  • Q: How many CPUs are supported by CoreFreq ?

    A: Up to 1024 CPUs can be built using the make CORE_COUNT option. 256 as a default.

  • Q: Turbo Technology is activated however CPUs don't reach those frequencies ?

  • Q: The CPU ratio does not go above its minimum value ?

  • Q: The UI shows erratic counters values !

    A: In the kernel boot command argument line, disable the NMI Watchdog
    nmi_watchdog=0

  • Q: The Processor does not enter the C-States ?

    A1: Check if at least one Idle driver is running.
    Accordingly to the Processor specs, provide a max_cstate value in the kernel argument as below.
    intel_idle.max_cstate=value

    A2: CoreFreq can also register itself as a cpuidle driver.
    This time, any idle driver will have to be blacklisted in the kernel command line; such as:
    modprobe.blacklist=intel_cstate idle=halt intel_idle.max_cstate=0
    Start the CoreFreq driver with the Register_CPU_Idle parameter:
    insmod corefreqk.ko Register_CPU_Idle=1

  • Q: The CoreFreq UI refreshes itself slowly, with a delay after the actual CPUs usage ?

    A: The sampling time to read the counters can be reduced or increased using a CoreFreq module argument:
    insmod corefreqk.ko SleepInterval=value
    where <value> is supplied in milliseconds between a minimum of 100 ms and a maximum of 4500 ms. 1000 ms is the default value.

  • Q: The base clock reports a wrong frequency value ?

    A: CoreFreq uses various algorithms to estimate the base clock.

    1. The delta of two TimeStamp counters during a defined interval
    2. The value provided in the Processor brand string divided by the maximum ratio (without Turbo)
    3. A static value advertised by the manufacturer specs.
    4. The MSR_FSB_FREQ bits provided with the Core, Core2 and Atom architectures.

    The CoreFreq module can be started as follow to ignore the first algorithm (frequency estimation):
    insmod corefreqk.ko AutoClock=0

    Remark: algorithms # 2, 3 and 4 will not return any under/over-clock frequency.

  • Q: The CPU temperature is wrong ?

    A: CoreFreq employs two MSR registers to calculate the temperature.
    MSR_IA32_TEMPERATURE_TARGET - MSR_IA32_THERM_STATUS [DTS]

    Remark: if the MSR_IA32_TEMPERATURE_TARGET is not provided by the Processor, a default value of 100 degree Celsius is considered as a target.

  • Q: The menu option "Memory Controller" does not open any window ?

    A: Although Uncore and IMC features are under development, they can be activated with the Experimental driver argument:
    insmod corefreqk.ko Experimental=1

  • Q: The Instructions and PMC0 counters are stuck to zero ?

  • Q: The Daemon crashes whenever its stress tools are executing !

    A: The PCE bit of control register CR4 allows RDPMC in ring 3
    echo "2" > /sys/devices/cpu/rdpmc
    or using systemd, create file /etc/tmpfiles.d/boot.conf and add line:
    w /sys/devices/cpu/rdpmc - - - - 2

    Next, load the driver with the RDPMC_Enable argument to override the CR4 register:
    insmod corefreqk.ko RDPMC_Enable=1

  • Q: How to solely control the P-States or the HWP Performance States ?

    A1: Without the Kernel cpufreq framework (aka CONFIG_CPU_FREQ), CoreFreq will take the full control over P-States.
    This allow the User to select a capped frequency from the UI, either per Core, either for the whole Processor.

    A2: With cpufreq built into Kernel, allow CoreFreq to register as a cpufreq driver.
    In the Kernel boot command line, two ways:

  1. disable cpufreq with the Kernel parameter
    cpufreq.off=1

  2. blacklist any P-state driver; such as:
    modprobe.blacklist=acpi_cpufreq,pcc_cpufreq intel_pstate=disable

  3. load the CoreFreq driver with its Register_CPU_Freq parameter:
    insmod corefreqk.ko Register_CPU_Freq=1

  • Q: The CPU freezes or the System crashes.

    A1: Changing the Max ratio frequency (aka P0 P-State) makes the Kernel TSC clock source unstable.

    1. Boot the Kernel with these command line parameters notsc nowatchdog
    2. Optionally, build the CoreFreq driver with its udelay() TSC implementation
      make DELAY_TSC=1
    3. Allow CoreFreq to register a new TSC clock source using driver arguments:
      insmod corefreqk.ko TurboBoost_Enable=0 Register_ClockSource=1
    4. Switch the current system clock source to corefreq
      echo "corefreq" > /sys/devices/system/clocksource/clocksource0/current_clocksource

    A2: [AMD][Zen] CCD temperatures:
    CoreFreq driver can be forced to use the Kernel function amd_smn_read()
    make LEGACY=2
    However amd_smn_read() serializes the SMU access through a mutex.
    CoreFreq CPU monitoring loops are executed in an interrupt context where any blocking call like Mutex will freeze the kernel.
    As a recommendation, don't use this option and make sure no other SMU driver is running.

    A3: This Processor is not or partially implemented in CoreFreq.
    Please open an issue in the CPU support Wiki page.

  • Q: No voltage is showing up with Nehalem or Westmere processors ?

    A: Build CoreFreq as below if one of those chips is present:
    make HWM_CHIPSET=W83627
    or
    make HWM_CHIPSET=IT8720

  • Q: [AMD][Zen] How to read the idle states ?

    A: As a workarround to the missing documentation of the hardware counters, CoreFreq implements virtual counters based on the TSC
    Those VPMC are estimated each time the Kernel is entering an idle state level.
    The prerequisities are:

    1. Boot the Kernel without its idle drivers and no TSC default clock source set
      modprobe.blacklist=acpi_cpufreq idle=halt tsc=unstable
    2. Build CoreFreq with its TSC implementation
      make DELAY_TSC=1
    3. Load and register the CoreFreq kernel module as the system handler
      insmod corefreqk.ko Register_ClockSource=1 Register_Governor=1 Register_CPU_Freq=1 Register_CPU_Idle=1 Idle_Route=1
    4. Define CoreFreq as the System clock source
      echo "corefreq" > /sys/devices/system/clocksource/clocksource0/current_clocksource
    5. Start the Daemon then the Client
      alt text
    • The registration is confirmed into the Settings window
    • The idle limit can be changed at any time in the Kernel window
      alt text
  • Q: How to enable transparency in the User Interface ?

    A: Transparency is a build option invoked by the compilation directive UI_TRANSPARENCY

    1. Build the project with UI_TRANSPARENCY enabled
      make UI_TRANSPARENCY=1
    2. Start the Client with one of its transparency compatible colors theme
      corefreq-cli -OE 2 -t
    3. Or switch to that theme from Menu > Theme, shortcut [E]
  • Q: What are the build options for CoreFreq ?

    A: Enter make help to display them:

o---------------------------------------------------------------o
|  make [all] [clean] [info] [help] [install] [module-install]  |
|                                                               |
|  CC=<COMPILER>                                                |
|    where <COMPILER> is cc, gcc       [clang partial support]  |
|                                                               |
|  WARNING=<ARG>                                                |
|    where default argument is -Wall                            |
|                                                               |
|  KERNELDIR=<PATH>                                             |
|    where <PATH> is the Kernel source directory                |
|                                                               |
|  CORE_COUNT=<N>                                               |
|    where <N> is 64, 128, 256, 512 or 1024 builtin CPU         |
|                                                               |
|  LEGACY=<L>                                                   |
|    where level <L> is 1 or 2                                  |
|    1: assembly level restriction such as CMPXCHG16            |
|    2:   kernel level restriction like amd_smn_read()          |
|                                                               |
|  UBENCH=<N>                                                   |
|    where <N> is 0 to disable or 1 to enable micro-benchmark   |
|                                                               |
|  TASK_ORDER=<N>                                               |
|    where <N> is the memory page unit of kernel allocation     |
|                                                               |
|  FEAT_DBG=<N>                                                 |
|    where <N> is 0 or N for FEATURE DEBUG level                |
|    3: XMM assembly in RING operations                         |
|                                                               |
|  DELAY_TSC=<N>                                                |
|    where <N> is 1 to build a TSC implementation of udelay()   |
|                                                               |
|  OPTIM_LVL=<N>                                                |
|    where <N> is 0, 1, 2 or 3 of the OPTIMIZATION level        |
|                                                               |
|  MAX_FREQ_HZ=<freq>                                           |
|    where <freq> is at least 4850000000 Hz                     |
|                                                               |
|  HWM_CHIPSET=<chipset>                                        |
|    where <chipset> is W83627 or IT8720 or COMPATIBLE          |
|                                                               |
|  Performance Counters:                                        |
|    -------------------------------------------------------    |
|   |     MSR_CORE_PERF_UCC     |     MSR_CORE_PERF_URC     |   |
|   |----------- REG -----------|----------- REG -----------|   |
|   | MSR_IA32_APERF            |  MSR_IA32_MPERF           |   |
|   | MSR_CORE_PERF_FIXED_CTR1  |  MSR_CORE_PERF_FIXED_CTR2 |   |
|   | MSR_PPERF                 |  MSR_PPERF                |   |
|   | MSR_AMD_F17H_APERF        |  MSR_AMD_F17H_MPERF       |   |
|    -------------------------------------------------------    |
|                                                               |
|  User Interface Layout:                                       |
|    NO_HEADER=<F>  NO_FOOTER=<F>  NO_UPPER=<F>  NO_LOWER=<F>   |
|      when <F> is 1: don't build and display this area part    |
|    UI_TRANSPARENCY=<F>                                        |
|      when <F> is 1: build with background transparency        |
|                                                               |
|  Example:                                                     |
|    make CC=gcc OPTIM_LVL=3 FEAT_DBG=1                         |
|         MSR_CORE_PERF_UCC=MSR_CORE_PERF_FIXED_CTR1            |
|         MSR_CORE_PERF_URC=MSR_CORE_PERF_FIXED_CTR2            |
|         HWM_CHIPSET=W83627 MAX_FREQ_HZ=5350000000             |
|         CORE_COUNT=1024 NO_FOOTER=1 NO_UPPER=1                |
|         clean all                                             |
o---------------------------------------------------------------o
  • Q: What are the parameters of the CoreFreq driver ?

    A: Use the modinfo command to list them:

$ modinfo corefreqk.ko
parm:           ArchID:Force an architecture (ID) (int)
parm:           AutoClock:Estimate Clock Frequency 0:Spec; 1:Once; 2:Auto (int)
parm:           SleepInterval:Timer interval (ms) (uint)
parm:           TickInterval:System requested interval (ms) (uint)
parm:           Experimental:Enable features under development (int)
parm:           CPU_Count:-1:Kernel(default); 0:Hardware; >0: User value (int)
parm:           Target_Ratio_Unlock:1:Target Ratio Unlock; 0:Lock (short)
parm:           Clock_Ratio_Unlock:1:MinRatio; 2:MaxRatio; 3:Both Unlock (short)
parm:           Turbo_Ratio_Unlock:1:Turbo Ratio Unlock; 0:Lock (short)
parm:           Uncore_Ratio_Unlock:1:Uncore Ratio Unlock; 0:Lock (short)
parm:           ServiceProcessor:Select a CPU to run services with (int)
parm:           RDPMC_Enable:Enable RDPMC bit in CR4 register (ushort)
parm:           NMI_Disable:Disable the NMI Handler (ushort)
parm:           Override_SubCstate:Override Sub C-States (array of ushort)
parm:           PkgCStateLimit:Package C-State Limit (short)
parm:           IOMWAIT_Enable:I/O MWAIT Redirection Enable (short)
parm:           CStateIORedir:Power Mgmt IO Redirection C-State (short)
parm:           Config_TDP_Level:Config TDP Control Level (short)
parm:           Custom_TDP_Offset:TDP Limit Offset (watt) (array of short)
parm:           Activate_TDP_Limit:Activate TDP Limiting (array of short)
parm:           Activate_TDP_Clamp:Activate TDP Clamping (array of short)
parm:           Custom_TDC_Offset:TDC Limit Offset (amp) (short)
parm:           Activate_TDC_Limit:Activate TDC Limiting (short)
parm:           L1_HW_PREFETCH_Disable:Disable L1 HW Prefetcher (short)
parm:           L1_HW_IP_PREFETCH_Disable:Disable L1 HW IP Prefetcher (short)
parm:           L2_HW_PREFETCH_Disable:Disable L2 HW Prefetcher (short)
parm:           L2_HW_CL_PREFETCH_Disable:Disable L2 HW CL Prefetcher (short)
parm:           SpeedStep_Enable:Enable SpeedStep (short)
parm:           C1E_Enable:Enable SpeedStep C1E (short)
parm:           TurboBoost_Enable:Enable Turbo Boost (short)
parm:           C3A_Enable:Enable C3 Auto Demotion (short)
parm:           C1A_Enable:Enable C3 Auto Demotion (short)
parm:           C3U_Enable:Enable C3 UnDemotion (short)
parm:           C1U_Enable:Enable C1 UnDemotion (short)
parm:           CC6_Enable:Enable Core C6 State (short)
parm:           PC6_Enable:Enable Package C6 State (short)
parm:           ODCM_Enable:Enable On-Demand Clock Modulation (short)
parm:           ODCM_DutyCycle:ODCM DutyCycle [0-7] | [0-14] (short)
parm:           PowerMGMT_Unlock:Unlock Power Management (short)
parm:           PowerPolicy:Power Policy Preference [0-15] (short)
parm:           Turbo_Activation_Ratio:Turbo Activation Ratio (short)
parm:           PState_FID:P-State Frequency Id (int)
parm:           PState_VID:P-State Voltage Id (int)
parm:           HWP_Enable:Hardware-Controlled Performance States (short)
parm:           HWP_EPP:Energy Performance Preference (short)
parm:           HDC_Enable:Hardware Duty Cycling (short)
parm:           EEO_Disable:Disable Energy Efficiency Optimization (short)
parm:           R2H_Disable:Disable Race to Halt (short)
parm:           Clear_Events:Clear Thermal and Power Events (uint)
parm:           ThermalScope:[0:None; 1:SMT; 2:Core; 3:Package] (int)
parm:           VoltageScope:[0:None; 1:SMT; 2:Core; 3:Package] (int)
parm:           PowerScope:[0:None; 1:SMT; 2:Core; 3:Package] (int)
parm:           Register_CPU_Idle:Register the Kernel cpuidle driver (short)
parm:           Register_CPU_Freq:Register the Kernel cpufreq driver (short)
parm:           Register_Governor:Register the Kernel governor (short)
parm:           Register_ClockSource:Register Clock Source driver (short)
parm:           Idle_Route:[0:Default; 1:I/O; 2:HALT; 3:MWAIT] (short)
parm:           Mech_IBRS:Mitigation Mechanism IBRS (short)
parm:           Mech_STIBP:Mitigation Mechanism STIBP (short)
parm:           Mech_SSBD:Mitigation Mechanism SSBD (short)
parm:           Mech_IBPB:Mitigation Mechanism IBPB (short)
parm:           Mech_L1D_FLUSH:Mitigation Mechanism Cache L1D Flush (short)
parm:           Mech_PSFD:Mitigation Mechanism PSFD (short)
parm:           WDT_Enable:Watchdog Hardware Timer (short)

Algorithm

alt text

About

CyrIng

Copyright (C) 2015-2021 CYRIL INGENIERIE