[workaround] manual install of AMDGPU 20.10

May 16, 2020
271
16
38
51
Antwerp, Belgium
commandline.be
When manually installing the unified AMDGPU 20.10 drivers, do this.

Download drivers from : https://www.amd.com/en/support/kb/release-notes/rn-amdgpu-unified-linux-20-10

change id=debian to id=ubuntu in /etc/os-release and save-quit the file.

only then proceed to install the driver.​
cp amdgpu-pro-20.10-1048554-ubuntu-18.04.tar.xz /usr/src/​
tar -xf amdgpu-pro-20.10-1048554-ubuntu-18.04.tar.xz​
cd amdgpu-pro-20.10-1048554-ubuntu-18.04​
./amdgpu-pro-install --no-32 --opencl=pal​

change back id=ubuntu to id=debian in /etc/os-release
reboot

source:https://github.com/RadeonOpenCompute/ROCK-Kernel-Driver/issues/77

This actually works, a really long time since i ever even consider to attempt that.

However: you may find the VM configured for PCI-E passthrough is not booting.

Try: qm start vmid
it may display a line containing

Failed to mmap 0000:..:.... BAR .. Performance may be slow​
see also the below command for more clues (if replaced the real ID with 0000:..:.... BAR)

dmesg | grep BAR​
[ 0.200987] pci 0000:..:.... BAR 0: assigned to efifb
[ 366.105096] vfio-pci 0000:..:.... BAR 0: can't reserve [mem 0xe0000000-0xefffffff 64bit pref]
[ 381.184838] vfio-pci 0000:..:.... BAR 0: can't reserve [mem 0xe0000000-0xefffffff 64bit pref]
[ 642.538394] vfio-pci 0000:..:.... BAR 0: can't reserve [mem 0xe0000000-0xefffffff 64bit pref]

my assumption is altering cmdline in /etc/default/grub to contain efifb=off should fix this
source: https://passthroughpo.st/explaining-csm-efifboff-setting-boot-gpu-manually/

and yes

dmesg | grep BAR
[ 0.201136] pci 0000:..:.... BAR 0: assigned to efifb

Do understand the VM must typically have been installed with UEFI/OVMF bios instead of Seagate BIOS. It is not easy (or possible without great efforts) to switch after either was installed.
 
Last edited: