Beyond Linux® From Scratch (System V Edition) - Version r11.2-1338

Chapter 3. After LFS Configuration Issues

About Firmware

On some recent PCs it can be necessary, or desirable, to load firmware to make them work at their best. There is a directory, /lib/firmware, where the kernel or kernel drivers look for firmware images.

Currently, most firmware can be found at a git repository: https://git.kernel.org/cgit/linux/kernel/git/firmware/linux-firmware.git/tree/. For convenience, the LFS Project has created a mirror, updated daily, where these firmware files can be accessed via wget or a web browser at https://anduin.linuxfromscratch.org/BLFS/linux-firmware/.

To get the firmware, either point a browser to one of the above repositories and then download the item(s) which you need, or install git-2.39.2 and clone that repository.

For some other firmware, particularly for Intel microcode and certain wifi devices, the needed firmware is not available in the above repository. Some of this will be addressed below, but a search of the Internet for needed firmware is sometimes necessary.

Firmware files are conventionally referred to as blobs because you cannot determine what they will do. Note that firmware is distributed under various different licenses which do not permit disassembly or reverse-engineering.

Firmware for PCs falls into four categories:

[Note]

Note

Although not needed to load a firmware blob, the following tools may be useful for determining, obtaining, or preparing the needed firmware in order to load it into the system: cpio-2.13, git-2.39.2, pciutils-3.9.0, and Wget-1.21.3

User Notes: https://wiki.linuxfromscratch.org/blfs/wiki/aboutfirmware

Microcode updates for CPUs

In general, microcode can be loaded by the BIOS or UEFI, and it might be updated by upgrading to a newer version of those. On linux, you can also load the microcode from the kernel if you are using an AMD family 10h or later processor (first introduced late 2007), or an Intel processor from 1998 and later (Pentium4, Core, etc), if updated microcode has been released. These updates only last until the machine is powered off, so they need to be applied on every boot.

Intel provide updates of their microcode for Skylake and later processors as new vulnerabilities come to light, and have in the past provided updates for processors from SandyBridge onwards, although those are no-longer supported for new fixes. New versions of AMD firmware are rare and usually only apply to a few models, although motherboard manufacturers get AGESA (AMD Generic Encapsulated Software Architecture) updates to change BIOS values, e.g. to support more memory variants, new vulnerability fixes or newer CPUs.

There were two ways of loading the microcode, described as 'early' and 'late'. Early loading happens before userspace has been started, late loading happens after userspace has started. However, late loading is known to be problematic and not supported anymore (see the kernel commit x86/microcode: Taint and warn on late loading.) Indeed, early loading is needed to work around one particular erratum in early Intel Haswell processors which had TSX enabled. (See Intel Disables TSX Instructions: Erratum Found in Haswell, Haswell-E/EP, Broadwell-Y.) Without this update glibc can do the wrong thing in uncommon situations.

In previous versions of this book, late loading of microcode to see if it gets applied was recommended, followed by using an initrd to force early loading. But now that the contents of the Intel microcode tarball is documented, and AMD microcode can be read by a Python script to determine which machines it covers, there is no real reason to use late loading.

It might be still possible to manually force late loading of microcode. But it may cause kernel malfunction and you should take the risk yourself. You will need to reconfigure your kernel for either method. The instructions here will show you how to create an initrd for early loading. It is also possible to build the same microcode bin file into the kernel, which allows early loading but requires the kernel to be recompiled to update the microcode.

To confirm what processor(s) you have (if more than one, they will be identical) look in /proc/cpuinfo. Determine the decimal values of the cpu family, model and stepping by running the following command (it will also report the current microcode version): 

head -n7 /proc/cpuinfo

Convert the cpu family, model and stepping to pairs of hexadecimal digits, and remember the value of the “microcode” field. You can now check if there is any microcode available.

If you are creating an initrd to update firmware for different machines, as a distro would do, go down to 'Early loading of microcode' and cat all the Intel blobs to GenuineIntel.bin or cat all the AMD blobs to AuthenticAMD.bin. This creates a larger initrd - for all Intel machines in the 20200609 update the size was 3.0 MB compared to typically 24 KB for one machine.

Intel Microcode for the CPU

The first step is to get the most recent version of the Intel microcode. This must be done by navigating to https://github.com/intel/Intel-Linux-Processor-Microcode-Data-Files/releases/ and downloading the latest file there. As of this writing the most secure version of the microcode is microcode-20230214. Extract this file in the normal way, the microcode is in the intel-ucode directory, containing various blobs with names in the form XX-YY-ZZ. There are also various other files, and a releasenote.

In the past, intel did not provide any details of which blobs had changed versions, but now the releasenote details this. You can compare the microcode version in /proc/cpuinfo with the version for your CPU model in the releasenote to know if there is an update.

The recent firmware for older processors is provided to deal with vulnerabilities which have now been made public, and for some of these such as Microarchitectural Data Sampling (MDS) you might wish to increase the protection by disabling hyperthreading, or alternatively to disable the kernel's default mitigation because of its impact on compile times. Please read the online documentation at https://www.kernel.org/doc/html/latest/admin-guide/hw-vuln/index.html.

For an Icelake mobile (described as Intel(R) Core(TM) i7-1065G7 CPU) the relevant values are cpu family 6, model 126, stepping 5 so in this case the required identification is 06-7e-05. The releasenote says the latest microcode for it is versioned 0xb8. If the value of the “microcode” field in /proc/cpuinfo is 0xb8 or greater, it indicates the microcode update is already applied by the BIOS. Otherwise, configure the kernel to support loading Intel microcode, and then proceed to the section called “Early loading of microcode”: 

General Setup --->
  [*] Initial RAM filesystem and RAM disk (initramfs/initrd) support [CONFIG_BLK_DEV_INITRD]
Processor type and features  --->
  [*] CPU microcode loading support  [CONFIG_MICROCODE]
  [*]      Intel microcode loading support [CONFIG_MICROCODE_INTEL]

AMD Microcode for the CPU

Begin by downloading a container of firmware for your CPU family from https://anduin.linuxfromscratch.org/BLFS/linux-firmware/amd-ucode/. The family is always specified in hex. Families 10h to 14h (16 to 20) are in microcode_amd.bin. Families 15h, 16h, 17h (Zen, Zen+, Zen2) and 19h (Zen3) have their own containers, but very few machines are likely to get updated microcode. Instead, AMD provide an updated AGESA to the motherboard makers, who may provide an updated BIOS using this. There is a Python3 script at https://github.com/AMDESE/amd_ucode_info/blob/master/amd_ucode_info.py. Download that script and run it against the bin file to check which processors have updates.

For the very old Athlon(tm) II X2 in these examples the values were cpu family 16, model 5, stepping 3 giving an identification of Family=0x10 Model=0x05 Stepping=0x03. One line of the amd_ucode_info.py script output describes the microcode version for it: 

Family=0x10 Model=0x05 Stepping=0x03: Patch=0x010000c8 Length=960 bytes

If the value of the “microcode” field in /proc/cpuinfo is 0x10000c8 or greater, it indicates the BIOS has already applied the microcode update. Otherwise, configure the kernel to support loading AMD microcode, and then proceed to the section called “Early loading of microcode”: 

General Setup --->
  [*] Initial RAM filesystem and RAM disk (initramfs/initrd) support [CONFIG_BLK_DEV_INITRD]
Processor type and features  --->
  [*] CPU microcode loading support  [CONFIG_MICROCODE]
  [*]      AMD microcode loading support [CONFIG_MICROCODE_AMD]

Early loading of microcode

If you have established that updated microcode is available for your system, it is time to prepare it for early loading. This requires an additional package, cpio-2.13 and the creation of an initrd which will need to be added to grub.cfg.

It does not matter where you prepare the initrd, and once it is working you can apply the same initrd to later LFS systems or newer kernels on this same machine, at least until any newer microcode is released. Use the following commands: 

mkdir -p initrd/kernel/x86/microcode
cd initrd

For an AMD machine, use the following command (replace <MYCONTAINER> with the name of the container for your CPU's family): 

cp -v ../<MYCONTAINER> kernel/x86/microcode/AuthenticAMD.bin

Or for an Intel machine copy the appropriate blob using this command: 

cp -v ../intel-ucode/<XX-YY-ZZ> kernel/x86/microcode/GenuineIntel.bin

Now prepare the initrd: 

find . | cpio -o -H newc > /boot/microcode.img

You now need to add a new entry to /boot/grub/grub.cfg and here you should add a new line after the linux line within the stanza. If /boot is a separate mountpoint: 

initrd /microcode.img

or this if it is not: 

initrd /boot/microcode.img

If you are already booting with an initrd (see the section called “About initramfs”), you should run mkinitramfs again after putting the appropriate blob or container into /lib/firmware. More precisely, put an intel blob in a /lib/firmware/intel-ucode directory or an AMD container in a /lib/firmware/amd-ucode directory before running mkinitramfs. Alternatively, you can have both initrd on the same line, such as initrd /microcode.img /other-initrd.img (adapt that as above if /boot is not a separate mountpoint).

You can now reboot with the added initrd, and then use the following command to check that the early load worked: 

dmesg | grep -e 'microcode' -e 'Linux version' -e 'Command line'

If you updated to address vulnerabilities, you can look at the output of the lscpu command to see what is now reported.

The places and times where early loading happens are very different in AMD and Intel machines. First, an example of an Intel (Icelake mobile) with early loading: 

[    0.000000] microcode: microcode updated early to revision 0xb8, date = 2022-08-31
[    0.000000] Linux version 6.1.11 (xry111@stargazer) (gcc (GCC) 12.2.0, GNU ld (GNU Binutils) 2.40) #2 SMP PREEMPT_DYNAMIC Tue Feb 14 23:23:31 CST 2023
[    0.000000] Command line: BOOT_IMAGE=/vmlinuz-6.1.11-lfs-11.3-rc1 root=PARTUUID=<CLASSIFIED> ro
[    0.452924] microcode: sig=0x706e5, pf=0x80, revision=0xb8
[    0.453197] microcode: Microcode Update Driver: v2.2.

A historic AMD example: 

[    0.000000] Linux version 4.15.3 (ken@testserver) (gcc version 7.3.0 (GCC))
               #2 SMP Sun Feb 18 02:32:03 GMT 2018
[    0.000000] Command line: BOOT_IMAGE=/vmlinuz-4.15.3-sda5 root=/dev/sda5 ro
[    0.307619] microcode: microcode updated early to new patch_level=0x010000c8
[    0.307678] microcode: CPU0: patch_level=0x010000c8
[    0.307723] microcode: CPU1: patch_level=0x010000c8
[    0.307795] microcode: Microcode Update Driver: v2.2.

Firmware for Video Cards

Firmware for ATI video chips (R600 and later)

These instructions do NOT apply to old radeons before the R600 family. For those, the firmware is in the kernel's /lib/firmware/ directory. Nor do they apply if you intend to avoid a graphical setup such as Xorg and are content to use the default 80x25 display rather than a framebuffer.

Early radeon devices only needed a single 2K blob of firmware. Recent devices need several different blobs, and some of them are much bigger. The total size of the radeon firmware directory is over 500K — on a large modern system you can probably spare the space, but it is still redundant to install all the unused files each time you build a system.

A better approach is to install pciutils-3.9.0 and then use lspci to identify which VGA controller is installed.

With that information, check the RadeonFeature page of the Xorg wiki for Decoder ring for engineering vs marketing names to identify the family (you may need to know this for the Xorg driver in BLFS — Southern Islands and Sea Islands use the radeonsi driver) and the specific model.

Now that you know which controller you are using, consult the Radeon page of the Gentoo wiki which has a table listing the required firmware blobs for the various chipsets. Note that Southern Islands and Sea Islands chips use different firmware for kernel 3.17 and later compared to earlier kernels. Identify and download the required blobs then install them: 

mkdir -pv /lib/firmware/radeon
cp -v <YOUR_BLOBS> /lib/firmware/radeon

There are actually two ways of installing this firmware. BLFS, in the 'Kernel Configuration for additional firmware' section part of the Xorg ATI Driver-19.1.0 section gives an example of compiling the firmware into the kernel - that is slightly faster to load, but uses more kernel memory. Here we will use the alternative method of making the radeon driver a module. In your kernel config set the following: 

Device Drivers --->
  Graphics support --->
      Direct Rendering Manager --->
        [*] Direct Rendering Manager (XFree86 ... support)  [CONFIG_DRM]
      [M] ATI Radeon                                        [CONFIG_DRM_RADEON]

Loading several large blobs from /lib/firmware takes a noticeable time, during which the screen will be blank. If you do not enable the penguin framebuffer logo, or change the console size by using a bigger font, that probably does not matter. If desired, you can slightly reduce the time if you follow the alternate method of specifying 'y' for CONFIG_DRM_RADEON covered in BLFS at the link above — you must specify each needed radeon blob if you do that.

Firmware for AMD/ATI amdgpu video chips

All video controllers using the amdgpu kernel driver require firmware, whether you will be using the xorg amdgpu driver, the xserver's modesetting driver, or just kernel modesetting to get a console framebuffer larger than 80x25.

Install pciutils-3.9.0 and use that to check the model name (look for 'VGA compatible controller:'). If you have an APU (Accelerated Processing Unit, i.e. CPU and video on the same chip) that will probably tell you the name. If you have a separate amdgpu video card you will need to search to determine which name it uses (e.g. a card described as Advanced Micro Devices, Inc. [AMD/ATI] Baffin [Radeon RX 550 640SP / RX 560/560X] needs Polaris11 firmware. There is a table of "Family, Chipset name, Product name and Firmware" at the end of the Kernel sections in AMDGPU page of the Gentoo wiki.

Once you have identified the firmware name, install all the relevant files for it. For example, the Baffin card mentioned above has 21 different polaris11* files, APUs such as renoir and picasso have at least 12 files and might gain more in future updates (e.g. the raven APU now has a 13th file, raven_ta.bin). 

mkdir -pv /lib/firmware/amdgpu
cp -v <YOUR_BLOBS> /lib/firmware/amdgpu

If disk space is not a problem, you could install all the current amdgpu firmware files and not worry about exactly which chipset is installed.

Building the kernel amdgpu driver as a module is recommended. In your kernel .config set at least the following options and review the other AMDGPU options according to your target hardware, for example "ACP (Audio Co-Processor) Configuration": 

Device Drivers --->
  Graphics support --->
      Direct Rendering Manager --->
        [*] Direct Rendering Manager (XFree86 ... support)  [CONFIG_DRM]
        [M] AMD GPU                                         [CONFIG_DRM_AMDGPU]
        Display Engine Configuration --->
          [*] AMD DC - Enable new display engine (NEW)      [CONFIG_DRM_AMD_DC]

As written above at the end of the section on 'Firmware for ATI video chips', loading large blobs from /lib/firmware can take a noticeable time during which the screen will be blank. On a slow machine you might wish to refer to the 'Kernel Configuration for additional firmware' part of Xorg AMDGPU Driver-23.0.0 and compile all the required modules into the kernel to reduce this time, at the cost of using more kernel memory.

Firmware for Nvidia video chips

Nvidia has released basic signed firmware for recent graphics chips, but significantly after the chips and its own binary drivers were first available. For other chips it has been necessary to extract the firmware from the binary driver.

For more exact information about which chips need extracted firmware, see https://nouveau.freedesktop.org/wiki/VideoAcceleration/#firmware.

First, the kernel Nvidia driver must be activated: 

Device Drivers --->
  Graphics support --->
      Direct Rendering Manager --->
        <*> Direct Rendering Manager (XFree86 ... support)  [CONFIG_DRM]
      <*/M> Nouveau (NVIDIA) cards                          [CONFIG_DRM_NOUVEAU]

If the necessary firmware is available in the nvidia/ directory of linux-firmware, copy it to /lib/firmware/nouveau.

If the firmware has not been made available in linux-firmware, for the old chips mentioned in the nouveau wiki link above ensure you have installed Python-2.7.18 and run the following commands: 

wget https://raw.github.com/imirkin/re-vp2/master/extract_firmware.py
wget https://us.download.nvidia.com/XFree86/Linux-x86/325.15/NVIDIA-Linux-x86-325.15.run
sh NVIDIA-Linux-x86-325.15.run --extract-only
python2 extract_firmware.py
mkdir -p /lib/firmware/nouveau
cp -d nv* vuc-* /lib/firmware/nouveau/

Firmware for Network Interfaces

The kernel likes to load firmware for some network drivers, particularly those from Realtek (the /lib/linux-firmware/rtl_nic/) directory, but they generally appear to work without it. Therefore, you can boot the kernel, check dmesg for messages about this missing firmware, and if necessary download the firmware and put it in the specified directory in /lib/firmware so that it will be found on subsequent boots. Note that with current kernels this works whether or not the driver is compiled in or built as a module, there is no need to build this firmware into the kernel. Here is an example where the R8169 driver has been compiled in but the firmware was not made available. Once the firmware had been provided, there was no mention of it on later boots. 

dmesg | grep firmware | grep r8169
[    7.018028] r8169 0000:01:00.0: Direct firmware load for rtl_nic/rtl8168g-2.fw failed with error -2
[    7.018036] r8169 0000:01:00.0 eth0: unable to load firmware patch rtl_nic/rtl8168g-2.fw (-2)

Firmware for Other Devices

Identifying the correct firmware will typically require you to install pciutils-3.9.0, and then use lspci to identify the device. You should then search online to check which module it uses, which firmware, and where to obtain the firmware — not all of it is in linux-firmware.

If possible, you should begin by using a wired connection when you first boot your LFS system. To use a wireless connection you will need to use a network tools such as Wireless Tools-29 and wpa_supplicant-2.10.

Different countries have different regulations on the radio spectrum usage of wireless devices. You can install a firmware to make the wireless devices obey local spectrum regulations, so you won't be inquired by local authority or find your wireless NIC jamming the frequencies of other devices (for example, remote controllers). The regulatory database firmware can be downloaded from https://kernel.org/pub/software/network/wireless-regdb/. To install it, simply extract regulatory.db and regulatory.db.p7s from the tarball into /lib/firmware. The access point would send a country code to your wireless NIC, and wpa_supplicant-2.10 would tell the kernel to load the regulation of this country from regulatory.db, and enforce it.

Firmware may also be needed for other devices such as some SCSI controllers, bluetooth adaptors, or TV recorders. The same principles apply.