PVE 5.2; kernel and vfio passthrough errors: Mapping multiple BARs

[Republicus]

Update: I have posted flooded log messages in second post which may be easier to identify and resolve then the remainder of this original post.

Prior to upgrading to PVE 5.2 my guest VMs were booting without issue using OVMF.

Some issues I am noticing:

1) Windows 10 guests, OVMF, boot to windows loading splash screen and hang.
2) Windows 10 guests, OVMF, with GPU passthrough crashes after driver is initialized.
3) Windows 7 guests, OVMF, with GPU passthrough loads in background while video on GPU is stuck on Windows loading splash screen (verified with connection through RDP that system does load).
4) Linux guests, OVMF, with GPU passthrough have graphical glitches and tears through GPU.
5) All guests, SeaBios, load drivers fine with no graphical glitches.

I was unable to get Windows 10 to boot at all past the loading screen (with or without GPU passthrough) until I applied the patch to pve-edk2-firmware as referenced by Nicholas Sherlock on his blog "Proxmox 5.1’s version of the OVMF firmware contains two commits (2ac1730 and 147fd35) that are intended to mark the pagetables as read-only during startup [...] patched OVMF to revert the effect of these two commits[.]"

With the above patch applied Windows 10 guests will boot without GPU passthrough, but crash after driver initialization with GPU passthrough.

The patch is intended to make macOS systems bootable. So additionally, note that macOS boots fine after patched with OVMF. No graphical glitches or crashing as seen on both Windows and Linux. This may be because when using Clover bootloader - the dsdt/ssdt is patched on-the-fly and is no longer using the native ovmf dsdt/ssdt tables.

Everything worked fine on PVE 5.1 (no-subscription repo). My card does support UEFI but it appears something is going

I can't be sure if this bug report addresses any of the same issues, or if the released fix is implemented in proxmox - but it sounds partly similiar to my woes.

https://bugs.launchpad.net/ubuntu/+source/edk2/+bug/1725560

The bug description mentiones IDE devices, which I am not using. But the bug summary is:

- OVMF UEFI firmware causes Windows 10 to not boot after upgrade
+ OVMF UEFI firmware causes Windows 10 w/ GPU passthrough to not boot
+ after upgrade

Has anyone else noticed any issues with OVMF/UEFI guests after upgrading to PVE 5.2?

 

[Republicus]

Actually I was premature in reporting success with SeaBIOS.

My logs are filling up quickly and reporting:

Jun 12 13:52:26 pve kernel: [  615.796277] caller vfio_pci_vga_rw+0x16d/0x240 [vfio_pci] mapping multiple BARs
Jun 12 13:52:26 pve kernel: [  615.796393] resource sanity check: requesting [mem 0x000a0000-0x000bffff], which spans more than PCI Bus 0000:20 [mem 0x000a0000-0x000b0000 window]

I havn't tried yet adding to grub:

"video=efifb:off"

as suggested for "BAR 3: can't reserve [mem] error" as that'ts not the particular message I am seeing - neither that, nor is my GPU in the primary pcie slot.

Any thoughts on these messages?

 

[hal900]

Actually I was premature in reporting success with SeaBIOS.

My logs are filling up quickly and reporting:

Jun 12 13:52:26 pve kernel: [  615.796277] caller vfio_pci_vga_rw+0x16d/0x240 [vfio_pci] mapping multiple BARs
Jun 12 13:52:26 pve kernel: [  615.796393] resource sanity check: requesting [mem 0x000a0000-0x000bffff], which spans more than PCI Bus 0000:20 [mem 0x000a0000-0x000b0000 window]

I havn't tried yet adding to grub:

"video=efifb:off"

as suggested for "BAR 3: can't reserve [mem] error" as that'ts not the particular message I am seeing - neither that, nor is my GPU in the primary pcie slot.

Any thoughts on these messages?

I was getting the same error trying to pass through the Intel onboard GPU on an i5-4570 to a Windows 10 guest (using SeaBIOS, OVMF never worked). I went back to Proxmox 5.1 and it's working without issues.