Looping Invalidation Time-out Error and possible fix

[BoyStot]

Thanks for the work on this patch its really helpful.

I installed this on Proxmox 7.2 including a hookscript for adding the device_specific reset method and it fires whenever a VM uses my device, a Radeon Pro WX5100, and it is running as it should.

[ 892.924004] vfio-pci 0000:07:00.0: AMD_POLARIS10: version 1.1
[ 892.924013] vfio-pci 0000:07:00.0: AMD_POLARIS10: performing pre-reset
[ 892.943929] vfio-pci 0000:07:00.0: AMD_POLARIS10: performing reset
[ 892.943942] vfio-pci 0000:07:00.0: AMD_POLARIS10: CLOCK_CNTL: 0x0, PC: 0x2a44
[ 892.943945] vfio-pci 0000:07:00.0: AMD_POLARIS10: performing post-reset
[ 892.983937] vfio-pci 0000:07:00.0: AMD_POLARIS10: reset result = 0

However when using vendor-reset as the only reset method, if I stop my VM and restart it I would still get a bug where there were a continuous stream of DMAR errors which lock up the node and force me to do a cold boot.

[ 893.004144] DMAR: VT-d detected Invalidation Completion Error: SID 0
[ 893.004146] DMAR: QI HEAD: UNKNOWN qw0 = 0x0, qw1 = 0x0
[ 893.004148] DMAR: QI PRIOR: Invalidation Wait qw0 = 0x200000025, qw1 = 0x185f6729c4
[ 893.004150] DMAR: Invalidation Completion Error (ICE) cleared
[ 893.004266] DMAR: VT-d detected Invalidation Time-out Error: SID 0
[ 893.004268] DMAR: QI HEAD: UNKNOWN qw0 = 0x0, qw1 = 0x0
[ 893.004269] DMAR: QI PRIOR: Invalidation Wait qw0 = 0x200000025, qw1 = 0x185f6729cc
[ 893.004272] DMAR: Invalidation Time-out Error (ITE) cleared

A lot of searching and testing different things and I found this pci reset script that when added to the post-stop hook in Proxmox allows the VM to stop and start without crashing the host.

I still get 3-4 of the DMAR errors on VM start, however they stop right away and the VM launches and I can use the GPU without issue.

Only thing that remains is that Proxmox hook scripts don't run when the VM guest does a reboot rather than shutdown, but vendor-reset does still reset the device in this case and I see the AMD_POLARIS10 messages.

Is there something that this script is doing that could be included in vendor-reset so that guest reboots could be made to work also?

Thanks!