Different modes in intel_pstate / intel_pstate 的各种运行模式 | Breakertt Blog

An (potentially) easy-to-understand explanation on different modes in intel_pstate driver

一个（也许）简单易懂的对 intel_pstate 驱动的不同模式的一个解释

（中文）

Preliminary Knowledge - C-states and P-states

Processor P-states and C-states - Thomas-Krenn-Wiki-en

Checking if CPU Supports HWP

Run the command:

cat /proc/cpuinfo

If the flags section includes hwp, then HWP is supported. Generally, Skylake and later processors support HWP, but this may vary on servers.

Kernel Boot Options Behavior [3]

• Default:
  • If the CPU supports HWP: active mode with HWP.
  • If the CPU does not support HWP: passive mode.
• intel_pstate=no_hwp:

  Note that the intel_pstate=no_hwp setting causes the driver to start in the passive mode if it is not combined with intel_pstate=active.

  • Default: passive mode.
  • With intel_pstate=active in boot options: active mode with no HWP.

Actual Behavior of Different intel_pstate Operating Modes

Current operating mode: /sys/devices/system/cpu/intel_pstate/status

The scaling_driver used by the current operating mode: /sys/devices/system/cpu/cpu0/cpufreq/scaling_driver

• Active Mode with HWP:

  • Only needs to provide performance guidance hints to CPU.
  • Loads intel_pstate into scaling_driver as intel_pstate.
  • Governor: The OS only has the option to select between the provided performance/powersave governors.
  • Mechanism: The CPU adjusts P-states itself. A callback is registered to the CPUFreq core to take over P-states adjustments but doesn’t do anything as the CPU manages it itself.

• Active Mode with No HWP:

  • intel_pstate works like acpi-cpufreq, doing only performance level setting work.
  • Loads intel_pstate into scaling_driver as intel_pstate.
  • Governor: Only supports performance/powersave governors.
  • Mechanism: Uses hardware coordination feedback to adjust P-states. Registers a callback to the CPUFreq core, which contains algorithms for intel_pstate to adjust P-states.
    • Algorithm: Uses APERF and MPERF MSRs to calculate CPU utilization and select performance levels. [6]

• Passive Mode:

  • Loads intel_pstate into scaling_driver as intel_cpufreq.
  • Supports all governors.
  • Mechanism: Exposes all P-states (including Turbo Boost P-states) to the CPUFreq core without registering a callback, allowing Linux to use both Turbo Boost and its own P-states adjustment algorithms.
    • Algorithm: Uses IA32_PERF_CTL and IA32_PERF_STATUS. [6]

• Disabled:

  • intel_pstate is not loaded, and acpi-cpufreq is used as scaling_driver.
  • Mechanism: Similar to passive mode, but Turbo Boost cannot be used since the acpi generic driver cannot access all P-states.

    On the other hand, in the passive mode the driver behaves similarly to the generic acpi-cpufreq driver - it collaborates with the regular scaling governors. Although, it can use the full range of frequency steps. [1]

Best Practices

• Currently, for CPUs that do not support HWP, using passive mode with the performance governor is recommended if power consumption is not a concern. This is also what Intel advocates for as well [7].

---

前置知识 - C-states and P-states

Processor P-states and C-states - Thomas-Krenn-Wiki-en

检查 CPU 是否支持 hwp

cat /proc/cpuinfo

flags 里面有 hwp 相关就是有 hwp，一般来说 skylake 之后就会有，但是好像在服务器上不一样

内核 boot options 行为 [3]

• 默认
  • 如果 CPU 支持 hwp - active mode with HWP
  • 如果 CPU 不支持 hwp - passive
• intel_pstate=no_hwp

  [Note that the intel_pstate=no_hwp setting causes the driver to start in the passive mode if it is not combined with intel_pstate=active.]

  • 默认 - passive
  • boot options 里面还有 intel_pstate=active - active mode with no HWP

intel_pstate 不同运行模式的实际行为

当前使用的运行模式 - /sys/devices/system/cpu/intel_pstate/status

当前运行模式使用的 scaling_driver - /sys/devices/system/cpu/cpu0/cpufreq/scaling_drvier

• active mode with HWP
  • only needs to provide performance guidance hints to CPU
  • 会加载 intel_pstate 到 intel_pstate 的 scaling_driver
  • governor - OS 除了选择提供的 performance / powersave governor 以外没有别的操作空间
  • 原理 - CPU 自己会进行 P-states 的调整，会有一个回调注册到 CPUFreq core 来接管 P-states 调整但是实际上不干事情因为 CPU 自己会做
• active mode with no HWP
  • intel_pstate work as acpi-cpufreq doing only performance level setting work
  • 会加载 intel_pstate 到 intel_pstate 的 scaling_driver
  • governor - 只支持 performance / powersave
  • 原理 - 使用 Hardware coordination feedback 来进行 P-states 算法的调整，会注册一个回调到 CPUFreq core，这个回调里面会有一些算法来让 intel_pstate 调整 P-states
    • 算法 - use APERF and MPERF MSRs to calculate CPU utilization and do performance level selection by itself [6]
• passive
  • 会加载 intel_pstate 到 intel_cpufreq 的 scaling_driver
  • 支持所有 governer
  • 原理 - 通过把所有的 P-states （包含 Turbo Boost P-states）暴露给 CPUFreq core，同时不会注册回调到 CPUFreq core，这样 Linux 就可以同时使用 Turbo Boost 和自己的 P-states 调整算法
    • 算法 - 使用 IA32_PERF_CTL 和 IA32_PERF_STATUS [6]
• disable
  • 会不加载 intel_pstate，使用 acpi-cpufreq 作为 scaling_driver
  • 原理 - 别的都与 passive 相同，但是 Turbo boost 会无法使用，因为 acpi 通用驱动无法获得所有的 P-states

    On the other hand, in the passive mode the driver behaves similarly to the generic acpi-cpufreq driver - it collaborates with the regular scaling governors. Although, it can use the full range of frequency steps. [1]

最佳实践

• 目前看起来在不支持 HWP 的 CPU 上，不考虑功耗的情况下，用 passive + performance 就可以，这也是 Intel 在推的 [7].

References

[1] https://wiki.gentoo.org/wiki/Power_management/Processor

[2] https://wiki.archlinux.org/title/CPU_frequency_scaling

[3] https://www.kernel.org/doc/html/v6.9/admin-guide/pm/intel_pstate.html

[4] Using Processor Performance P-States with Linux on Intel-based ThinkSystem Servers

lp1946.pdf

[5] https://web.archive.org/web/20220926200749/https://makiras.org/archives/344?amp=1

[6] Intel Power Management & MSR_SAFE

04-2020-01-29-prace-ee-DC-RS.pdf

[7] https://www.phoronix.com/news/P-State-Passive-Def-For-No-HWP