tl;dr; not worth the effort but an interesting exercise
Overview of updating microcode on Linux
- Figure out your CPUID and current microcode version
- Check for a newer version and download it
- Integrate it in your boot process
This example will be useful for other AMD CPUs as well. Some parts should be useful for Intel CPUs, but there is generally more information about them elsewhere anyway.
Getting the CPUID
I couldn't figure out a nice way to figure out your CPUID. Initially I was confused by the output of `dmidecode` for the "Processor" section. Where "ID" read as "10 0F 87 00 FF FB 8B 17". It turned out though that it has nothing to do with the actual CPUID that is needed to obtain and upload microcode.
What I did was to open http://instlatx64.atw.hu/ which was linked from MC Extractor README and find my CPU by model name "Ryzen 5 3600". So I saw the proper CPUID. MC Extractor is used to extract and analyze microcode from BIOS images and other sources so it is an interesting project to explore further.
In this case the CPUID is "00870F10". After the fact, I could figure out that this ID can also be obtained by `cpuid -r` looking at the `eax` register at `0x00000001` and `0x80000001`. But I can't say that this is universal. You can check more details about the CPUID instruction itself.
Getting current microcode version
This must be much more straightforward. Current microcode you can see in the BIOS settings, `dmidecode`, `/proc/cpuinfo`, `cpuid` (on windows hwinfo and CPUz showed it too). It was 8701030 for me (motherboard has AGESA 1.2.0.B, while the zenbleed fix should be in 1.2.0.C, see AMD bulletin).
Also with `dmesg`
```
$ sudo dmesg | grep -i microcode
[ 0.431255] Zenbleed: please update your microcode for the most optimal fix
[ 1.127054] microcode: Current revision: 0x08701030
```
Finding a newer version
You can look at the excellent CPUMicrocodes repo. In the `AMD` folder you can search for files matching your CPUID. I found and donloaded this one cpu00870F10_ver08701033_2023-10-06_E71C3D44.bin
It is not in a format understandable by Linux kernel though. So it has to be packaged appropriately.
Packaging as Kernel microcode
First you need to clone and compile amd-ucodegen
```
$ git clone https://github.com/AndyLavr/amd-ucodegen.git
$ cd amd-ucodegen
$ make
$ ./amd-ucodegen -o ~/packaged-08701033.bin cpu00870F10_ver08701033_2023-10-06_E71C3D44.bin
```
Now you need to check your CPU model. for me this is 23 decimal and 17 hex. So I have to either integrate this file into the existing `/usr/lib/firmware/amd-ucode/microcode_amd_fam17h.bin` or just overwrite it. For completeness, lets see how to integrate into it.
```
$ git clone https://github.com/AMDESE/amd_ucode_info.git
$ sudo mv /usr/lib/firmware/amd-ucode/microcode_amd_fam17h.bin /usr/lib/firmware/amd-ucode/microcode_amd_fam17h.bin_bak
$ sudo python amd_ucode_info/amd_ucode_info.py -m /usr/lib/firmware/amd-ucode/microcode_amd_fam17h.bin /usr/lib/firmware/amd-ucode/microcode_amd_fam17h.bin_bak packaged-08701033.bin
```
Verify operation with
```
$ python amd_ucode_info/amd_ucode_info.py /usr/lib/firmware/amd-ucode/microcode_amd_fam17h.bin
Microcode patches in /usr/lib/firmware/amd-ucode/microcode_amd_fam17h.bin:
Family=0x17 Model=0x08 Stepping=0x02: Patch=0x0800820d Length=3200 bytes
Family=0x17 Model=0x31 Stepping=0x00: Patch=0x0830107b Length=3200 bytes
Family=0x17 Model=0xa0 Stepping=0x00: Patch=0x08a00008 Length=3200 bytes
Family=0x17 Model=0x01 Stepping=0x02: Patch=0x0800126e Length=3200 bytes
Family=0x17 Model=0x71 Stepping=0x00: Patch=0x08701033 Length=3200 bytes
```
Integrate it in your boot process
Different linux distributions use different approaches to install early microcode. You can check archwiki for information. On Fedora 39 I had to just call `dracut -f` and the magic was done.
btw an alternative to all this is to integrate the original raw microcode into your mainboard BIOS image. There are some howtos about it. But I think it is more dangerous than loading on boot.
Verifying the result
```
$ sudo dmesg | grep -i microcode
[ 1.126852] microcode: Current revision: 0x08701033
[ 1.126854] microcode: Updated early from: 0x08701030
```
You can see the difference with the previous call. No mention of Zenbleed anymore.
Performance effect
The performance effect is hard to measure and is prone to statistical error. But it appears that single-core performance is slightly better while multi-core performance is slightly worse. The penalty in multi-thread is slightly worse than the positive effect in single-thread performance.
See my scientific Geekbench results:
Obviously things depend on your workload. Biggest margin has
- Photo filter ~ 6%
- Text and PDF processing ~ 4%
HTH
No comments:
Post a Comment