PCI(e) pass through fixed in 7.4?

[Ozarc]

Hi, currently on Proxmox 7.2 and want to update to latest version which is 7.4 at this time. but want to make sure the following has been fixed before I do, as last time its caused havoc on my VMs, resulting in them not booting with my Nvidia card:

PCI(e) pass through related:

• Systems passing through a GPU may be affected from the switch to the SYS_FB (system frame buffer) KConfig build options using the simplefb module as driver in the new default 5.15 based kernel. The sys-fb allows taking over the FB from the firmware/earlier boot stages. Note that Proxmox VE uses the legacy simplefb driver over the modern simpledrm one due to regressions and issues we encountered on testing with the latter. Most of those issues are already fixed in newer kernels and Proxmox VE may try to switch to the modern, DRM based FB driver once it moves to 5.17, or newer, as its default kernel. If your systems is configured to pass through the (i)GPU, and you had to avoid the host kernel claiming the device, you may now need to also add video=simplefbff to the kernel boot command line.
• Setups using vendor-reset for PCIe pass through need to adapt to changes of the new default 5.15 based kernel, see For details see this issue. They must run the command echo 'device_specific' > /sys/bus/pci/devices/<PCI-ID>/reset_method before the VM is started. This can be automated by using a systemd service or using a on-boot cron script. Alternatively one can also use a VM hook script with the pre-start hook.

7.4 roadmap says "Latest 5.15 Kernel as stable default" and doesn't mention anything about it in its known issues section, so i assume it has but want to be 100% this time.

Thanks

 

[leesteken]

You still need this work-around (initcall_blacklist=sysfb_init) when doing passthrough of the boot GPU. (Unless you use the latest optional kernel and load amdgpu for the VGA-function of the AMD GPU instead of blacklisting or early binding to vfio-pci.)
If you install vendor-reset, you do need to enable it for each GPU that you want to use it on (with a hookscript for example). This is not a bug but a feature of recent Linux kernels (5.15+).